Machine learning applications in detection and diagnosis of urology cancers: a systematic literature review

Abstract

Deep learning integration in cancer diagnosis enhances accuracy and diagnosis speed which helps clinical decision-making and improves health outcomes. Despite all these benefits in cancer diagnosis, the present AI models in urology cancer diagnosis have not been sufficiently reviewed systematically. This paper reviews the artificial intelligence approaches used in cancer diagnosis, prediction, and treatment of urology cancer. AI models and their applications in urology subspecialties are evaluated and discussed. The Scopus, Microsoft Academic and PubMed/MEDLINE databases were searched in November 2022 using the terms “artificial intelligence”, “neural network”, “machine learning,” or “deep learning” combined with the phrase “urology cancers”. The search was limited to publications published within the previous 20 years to identify cutting-edge deep-learning applications published in English. Irrelevant review articles and publications were eliminated. The included research involves two kinds of research analysis: quantitative and qualitative. 48 articles were included in this survey. 25 studies proposed several approaches for prostate cancers, while 15 were for bladder cancers. 8 studies discussed renal cell carcinoma and kidney cancer. The models presented to detect urology cancers have achieved high detection accuracy (77–95%). Deep learning approaches that use convolutional neural networks have achieved the highest accuracy among other techniques. Although it is still progressing, the development of AI models for urology cancer detection, prediction, and therapy has shown significant promise. Additional research is required to employ more extensive, higher-quality, and more recent datasets to the clinical performance of the proposed AI models in urology cancer applications.

Full text

REVIEW
Machine learning applications in detection and diagnosis of urology
cancers: a systematic literature review
M. Lubbad
1
•D. Karaboga
1
•A. Basturk
1
•B. Akay
1
•U. Nalbantoglu
1
•I. Pacal
2
Received: 28 March 2023 / Accepted: 7 December 2023
The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024
Abstract
Deep learning integration in cancer diagnosis enhances accuracy and diagnosis speed which helps clinical decision-making
and improves health outcomes. Despite all these benefits in cancer diagnosis, the present AI models in urology cancer
diagnosis have not been sufficiently reviewed systematically. This paper reviews the artificial intelligence approaches used in
cancer diagnosis, prediction, and treatment of urology cancer. AI models and their applications in urology subspecialties are
evaluated and discussed. The Scopus, Microsoft Academic and PubMed/MEDLINE databases were searched in November
2022 using the terms ‘‘artificial intelligence’’, ‘‘neural network’’, ‘‘machine learning,’’ or ‘‘deep learning’’ combined with the
phrase ‘‘urology cancers’’. The search was limited to publications published within the previous 20 years to identify cutting-
edge deep-learning applications published in English. Irrelevant review articles and publications were eliminated. The
included research involves two kinds of research analysis: quantitative and qualitative. 48 articles were included in this
survey. 25 studies proposed several approaches for prostate cancers, while 15 were for bladder cancers. 8 studies discussed
renal cell carcinoma and kidney cancer. The models presented to detect urology cancers have achieved high detection
accuracy (77–95%). Deep learning approaches that use convolutional neural networks have achieved the highest accuracy
among other techniques. Although it is still progressing, the development of AI models for urology cancer detection,
prediction, and therapy has shown significant promise. Additional research is required to employ more extensive, higher-
quality, and more recent datasets to the clinical performance of the proposed AI models in urology cancer applications.
Keywords Bladder cancer Renal-cell carcinoma Prostate cancer Machine learning Deep learning Medical image
analysis Urology cancers Artificial intelligence
1 Introduction
Cancer is a complex disease characterized by the uncon-
trolled proliferation and division of cells, influenced by
genetic and environmental factors. According to the
National Cancer Institute (NIH), cancer has had a profound
impact, affecting nearly 150,580 individuals and causing
50,540 deaths per month in the USA in 2020 [1]. Conse-
quently, cancer has emerged as one of the leading causes of
mortality worldwide. Early detection and diagnosis of
cancer are crucial in improving long-term survival rates.
Medical imaging plays a pivotal role in the early detection,
follow-up, and post-treatment monitoring of cancer
patients [2]. However, manually examining a large volume
of medical images can be time-consuming and prone to
human error. To address this challenge, computer-aided
diagnosis (CAD) systems have been developed since the
early 1900s to assist physicians in interpreting medical
M. Lubbad and D. Karaboga have contributed equally to this
work.
&M. Lubbad
engmlubbad@gmail.com
D. Karaboga
karaboga@erciyes.edu.tr
A. Basturk
ab@erciyes.edu.tr
B. Akay
bahriye@erciyes.edu.tr
U. Nalbantoglu
nalbantoglu@erciyes.edu.tr
I. Pacal
ishak.pacal@igdir.edu.tr
1
Department of Computer Engineering, Engineering Faculty,
Erciyes University, Kayseri, Turkey
2
Department of Computer Engineering, Engineering Faculty,
Igdir University, Igdir, Turkey
123
Neural Computing and Applications
https://doi.org/10.1007/s00521-023-09375-2(0123456789().,-volV)(0123456789().,-volV)

images [3]. In recent years, the scientific community has
turned its attention to Artificial Intelligence (AI) tech-
niques, which have shown promising results in detecting
and classifying various severe diseases. AI aims to repli-
cate intelligent behaviors humans exhibit in machines and
computer systems. Among the different AI techniques,
machine learning has gained significant traction. Machine
learning involves automatically extracting valuable fea-
tures using deep architectures that combine linear and
nonlinear processing [4].
The application of machine learning algorithms in
urology cancer detection offers several advantages. These
algorithms have improved accuracy compared to tradi-
tional methods, enabling high precision in identifying
urology cancers. Additionally, they excel at early detec-
tion, leading to timely interventions and improved treat-
ment outcomes. Automated analysis capabilities reduce the
burden on healthcare professionals and increase efficiency
by automating the interpretation of medical images and
data. Integration with clinical workflows allows seamless
adoption in healthcare settings, while the analysis of
patient-specific data contributes to personalized medicine,
optimizing treatment plans and enhancing patient care.
Deep learning, a prominent subset of machine learning,
seeks to mimic the human brain’s organization, observa-
tion, and decision-making mechanisms based on data
inputs. Deep learning employs image data to construct
hierarchical features. A key advantage of deep learning is
its ability to extract high-level features directly from raw
images. GPU cards and parallel programming architectures
have further expedited the application of deep learning
techniques in various domains. The ever-advancing field of
deep learning in medical diagnostics has seen remarkable
studies that advance the use of AI in the interpretation of
medical imagery for urology cancer detection. Recent work
in this area has leveraged deep learning to refine the clas-
sification of prostate cancer by extracting sophisticated
patterns from histopathologic images [5]. Another study
has made significant strides by introducing an interactive
and explainable deep learning model that aids in the
interpretation of MRI scans for prostate cancer, enhancing
the transparency and interpretability of AI in the medical
diagnostics [6]. These developments represent a significant
leap forward in the utilization of AI for urology cancer
diagnostics, underscoring a trend towards more precise,
nuanced, and transparent AI applications. This aligns with
the ongoing shift towards precision medicine and the
delivery of patient-centric care. This study aims to conduct
a systematic review of research focusing on detecting and
identifying urology cancerous tissues using AI and deep
learning techniques. The primary research questions
addressed in this study are: What are the prevailing deep
learning techniques researchers employ to detect different
types of urology cancers, and how are these techniques
integrated into the diagnosis process? While the number of
review studies on this topic is limited [7], our study sys-
tematically reviews recent deep-learning models and
techniques proposed in the literature to diagnose urology
cancers. Concludes the review by proposing directions for
future research.
To achieve this objective, we conducted a comprehen-
sive search in November 2022, utilizing the PubMed
Medline, Scopus, and Microsoft Academic databases and
manual analysis. The search terms used included ‘‘artificial
intelligence,’’ ‘‘deep learning,’’ ‘‘machine learning,’’ or
‘‘artificial neural network,’’ combined with the phrase
‘‘urology cancers.’’ This article represents the latest sys-
tematic review of scientific papers utilizing deep learning
techniques to detect different urology cancers. The inclu-
sion of relevant articles was based on the PRISMA
methodology.
A total of 381 studies were identified through the data-
base search. After a thorough evaluation, we selected 48
papers for inclusion in this study. These papers were sub-
sequently categorized into 25 studies on prostate cancers,
15 studies on bladder cancers, and 8 studies on renal cell
carcinoma (kidney cancer). This research encompasses the
most up-to-date studies available at the time of article
submission. The remaining sections of this paper are
structured as follows: Sect. 2provides a concise overview
of the primary deep-learning techniques employed in
diagnosing urology cancers. Section 3outlines the
methodology adopted for this systematic review. In Sect. 4,
we present deep learning applications in detecting urology
cancers, categorizing them into four main groups based on
cancer type: renal cell carcinoma, bladder cancer, prostate
cancer, and other types of urology cancers. Each study is
summarized in detail, accompanied by a table highlighting
the deep learning methods utilized, the datasets employed,
and the corresponding results. Section 5discusses the
findings obtained and elucidates the existing challenges.
Finally, Sect. 6
2 Machine & deep learning applications
in health
Deep learning has been defined as an artificial neural net-
work approach since the 1940s. The ‘‘perceptron’’ is a
simple form used as the first fundamental component in
deep learning [6]. Artificial neural networks (ANNs) have
advanced rapidly in the recent decade, and ‘‘deep learning’’
has begun to be used. This study may classify deep learning
models into unsupervised, supervised, and deep hybrid
networks. Convolutional Neural Networks (CNNs) are a
kind of supervised learning that is now the most often used
Neural Computing and Applications
123

architecture in medical image processing [7,8]. The term
‘‘hybrid deep networks’’ refers to systems developed by
combining one or more deep learning architectures to
provide more outstanding performance. Recent studies
show the promising performance of CNNs [7] in cancer
detection and diagnosis. Deep belief neural networks and
variations, CNNs, and their Autoencoders, are the most
utilized deep learning approaches for producing exact
findings and outputs in urological malignancies. Conven-
tional machine learning approaches are often designed for
specific challenges and focus on predetermined architec-
tural features. Although deep learning algorithms do not
require an accurate definition of features, they rely on data
and combine dimensionally complex features to compre-
hend to produce a result. Because of deep learning meth-
ods’ excellent efficiency and effectiveness, many
traditional machine learning methods, such as decision
trees, support vector machines (SVMs), Random Forest
(RF), and Gauss mixed models, have been displaced by
deep learning approaches. As they are manual, these old
techniques for urological cancer diagnosis are neither long-
lasting, scalable, nor time-consuming. Due to significant
false-positive rates, these techniques are ineffective in
rapid urology cancer detection for urological pictures.
2.1 Machine learning: convolutional neural
networks (CNNs)
CNNs; are one of the most common multi-layer perceptron
networks, whose basic design evolved from Hubel’s study
in the cortical area of a cat [9]. Visual patterns are usually
recognized without intermediaries from unprocessed pic-
ture elements by CNNs [10]. The primary suggested model
for CNN is to detect handwritten digits (LeNet-5). A CNN
conducts a convolution process rather than a matrix oper-
ation in image layers [11]. CNN, the most common
architecture in deep learning architecture, is the most uti-
lized in research that employs ML algorithms in image
processing for medical analysis [12]. The most common
explanation for these input pictures is often analyZed using
CNNs, modifying them while preserving spatial connec-
tions. Because the links and interactions of malignant tissue
with normal tissue are frequently interpreted through spa-
tial relationships, spatial relations are essential for medical
picture analysis.
Deep learning may adapt well to images; CNN is one of
the proposed deep learning techniques for image process-
ing issues. Rather than being directly coupled, CNN layers
are structured in multiple blocks. Like the visual area, a
data flow removes the traditional methods problems
between these blocks. It can also automatically identify
data features, eliminating manual feature extraction chal-
lenges. Because CNN shares characteristics, the number of
variables is lower than it should be. However, the model
allows for quicker learning and avoids overfitting issues.
Therefore, CNNs perform and adapt well in medical image
processing because they can deal with big data and train on
many parameters (millions). Several comparable benefits
encourage the implementation of CNNs in various appli-
cations [8].
2.2 Deep learning: CNNs models
CNN has become the most common deep learning tech-
nique, providing the highest throughput among other AI
algorithms. Several studies have reported that deep learn-
ing algorithms improved diagnostic accuracy and enhanced
efficiency, reducing reading times without compromising
the detection and classification accuracy [9]. Researchers
have collaborated by proposing new approaches to increase
the efficiency of CNNs. These CNN algorithms are often
utilized in the study of urology cancer and the processing
of medical images. Figure 1shows an example of CNN
architecture. Different typical and recent Deep CNN
architectures, which are now utilized as building blocks in
various classification, segmentation, and detection designs,
will be briefly discussed in this section. The earliest CNN
model is the LeNet architecture, described in research by
Yann LeCun [23].
The scale of this building, which is fundamentally flat
and essential, is relatively small. This structure was used to
classify handwritten pictures.
CNN launched when the AlexNet architecture design
[25] won the ImageNet competition in 2012, and compa-
rable to the LeNet design, it is a more robust structure with
more feature filters. The AlexNet architecture success has
altered the path of image processing research significantly.
The ZFNet design [27] improves the AlexNet design’s
architecture, leading to winning the ImageNet competition
in 2013. Furthermore, with a 7.3 percent error rate, VGG
architecture [28] achieved significant success, first pre-
sented in the 2014 ImageNet Large Scale Visual Recog-
nition Challenge (ILSVRC) competition. The most popular
forms used as a backbone in CNN structures for object
classification and detection of VGG design are VGG16 and
VGG19. Figure 2depicts the architecture of VGG.
Furthermore, several additional designs were intro-
duced, including DenseNet [32], EfficientNet [33], and
MobileNets [31]. VGG, Inception, and ResNet are the most
often used architectures for cancer analysis and image
classification. In addition, while CNN structures were first
employed for classification, they were widely adopted
because of their effectiveness in object identification and
segmentation applications.
Many practical CNN-based algorithms were published
in this field since object identification is one of the most
Neural Computing and Applications
123

prominent bases of deep learning. Because of its success,
faster R-CNN [34], one of these two-phase topologies in
object identification, has been raised. In contrast to these
models, YOLO [37], RetinaNet [36], SSD [35] and their
variations were presented, which are single-stage tech-
niques that are quicker than two-step approaches. Struc-
tures such as CenterNet [39], and CornerNet [38], anchor-
free methods, have also been developed. Efficient and
Scalable Object Detection (EfficientDet), one of the most
recent methods, achieves 55.1% COCO AP; with these
object detection algorithms, the most recent object detec-
tion technology is employed as efficient algorithms that are
often used in cancer analysis. One of the most popular
CNN models for object segmentation is the Fully Convo-
lutional Networks (FCNs) [40] architecture, a very suc-
cessful and widely used basic architecture suggested for
semantic segmentation. It comprises only pooling, convo-
lution, and layers. The most significant distinction between
CNNs and FCNs is that FCNs use a deconvolutional and
pooling layer instead of the fully connected layer [41].
Based on the amplification in the deconvolution layer, FCN
design can be described by FCN-8 s, FCN-16 s, or FCN-
32 s. SegNet [43], UNet [42], RefineNet [45], and MaskR-
CNN [44] are some of the most powerful and widely used
architectures suggested for semantic segmentation. These
algorithms, such as gland and polyp segmentation, are the
most popular and successful segmentation algorithms uti-
lized in urology cancer analysis.
2.3 Recurrent neural networks
In evaluating and dealing with sequential data, recurrent
neural networks (RNNs) are utilized more frequently than
other approaches and perform better. RNNs have been used
in text mining, such as machine translation, text prediction,
and speech recognition, due to their capability to process
and create text [46,47]. The parameters of the RNN model
at different time intervals are shared based on the nature of
the data. The output of one layer in flat RNN is generally
appended to the following entry and sent back to that layer,
causing memory capacity problems. These difficulties
make network training challenging and temporal depen-
dencies difficult to represent. As a solution to this problem,
LSTMs were created, and the issue with RNNs was han-
dled by adding memory cells and different gates. In tem-
poral data, short or long data can be learned. RNNs are
hardly utilized in processing medical images and are usu-
ally employed as part of hybrid models that include CNNs
or other learning techniques. In most cases, these hybrid
models are used in segmentation procedures. Figure 3b
depicts an example of an RNN.
Fig. 1 A CNN architecture
Fig. 2 depicts the architecture of VGG16 [10]
Neural Computing and Applications
123

2.4 Autoencoders (AEs)
Another unsupervised deep learning approach is the
Autoencoders, which learns features from unlabeled input
data. Autoencoders (AEs) are composed of input, output,
and hidden layers. Even though the topology is equivalent
to feedforward neural networks, the purpose of the hidden
layer is to produce a distinct set of inputs. Figure 3depicts
an example of an AEs chart (c). AEs take input data,
extract codes from it, and then use these codes to produce
output data. The underlying principle is that the input data
be viable as the input data. As a result, they contain a cost
function that penalizes the model when there is a difference
between inputs and outputs. To be labeled, AEs do not
require training data. Encodings are often lower in size to
reduce complexity and computational cost. They also
provide output like the input training data. Because of these
benefits, it is commonly employed after CNN in processing
medical data when labeled data is limited. The most pop-
ular forms of AEs are Denoising Autoencoders, Stacked
Autoencoders, Variational Autoencoders, and Sparse
Autoencoders [48,49].
2.5 Deep belief networks (DBNs) and restricted
Boltzmann machines (RBMs)
Geoffrey Hinton et al. characterized Restricted Boltzmann
Machines (RBM) as a fast algorithm in 2006 [52], although
it was initially presented in 1986 [51]. RBMs are graphs
with two parts of hidden and visible balanced connections.
RBM is a Boltzmann Machine version of a random pro-
ductive neural network capable of learning the distribution
probability on the input set. RBMs are often used in size
reduction, classification, and feature learning. Figure 3d, e)
shows typical RBM and DBN charts. Hinton et al. intro-
duced deep belief networks, which use an unsupervised
learning method to construct one layer at a time [53]. The
fundamental concept is to train each layer of the network
independently using an RBM network that models the
preceding layer output. Using this technique to combine
Fig. 3 Deep learning architectures aCNN, bRNNs, cDBNs, dAEs, eRBMs [50]
Neural Computing and Applications
123

several simpler RBM models is an efficient way to learn a
new model. A DBN may understand the probabilistic
replication of trained inputs. Then, the DBN layers extract
the features. Following the learning process, a DBN may
be learned more about working in a supervised method.
3 Methodology
For conducting the research, the PubMed Medline, Scopus,
and Microsoft Academic databases were used up to
November 2022 using the following terms: ‘‘artificial
intelligence’’, ‘‘deep learning’’, ‘‘machine learning’’, ‘‘ar-
tificial neural network’’, ‘‘AI’’, ‘‘ML’’ or ‘‘DL’’ combined
with the term ‘‘urology cancer’’. The search was bound to
studies initially published in English within the last
20 years to detect recent deep-learning applications.
Review articles and irrelevant publications on the subject
were excluded. The following chart (Fig. 4) describes our
article methodology (selection process) in compliance with
the PRISMA criteria. There are about 48 articles that are
adequately discussed. As shown in Fig. 4, the first step in
the process is identification, which identifies related
research, then removes the duplicated study. The second
step is screening, which includes the most related analysis
after removing irrelevant and duplicated research. The
screened records are assessed for eligibility, and only
entitled research is selected. The included study involves
Studies identified through
database searching
(n =381)
ScreeningIncluded Eligibility Identification
Studies after irrelevant
excluded (n = 176)
Full-text articles
excluded, with reasons
(n = 49)
Studies included in
quantitative
synthesis (meta-
analysis)
(n = 48)
Full-text articles
assessed for eligibility
(n = 76)
Studies included in
qualitative synthesis
(n =51)
Studies after duplicates removed
(n = 80)
Studies screened
(n = 125)
Fig. 4 PRISMA: The flow
diagram displays the movement
of data across the various stages
of a systematic review. It
depicts the number of records
identified, those included and
those omitted, as well as the
reasons for exclusions.
Depending on the kind of
review (new or updated) and the
sources utilized to locate
research, several templates are
available
Neural Computing and Applications
123

two kinds of research analysis: quantitative and qualitative.
Figure 5, illustrates the classification of the included
studies. Twenty-five studies proposed several approaches
for prostate cancers, 15 studies for bladder cancers, and
eight discussed renal cell carcinoma, which is kidney
cancer.
3.1 Key terms and search methodology
The method was conducted using 3 databases and the
search keywords listed below: ‘Deep Learning’ OR
‘Machine Learning’ OR ‘Neural Network’ OR ‘Artificial
Intelligence’ OR ‘AI’ OR.
ML’ OR ‘DL’ AND ‘urology cancer’. The duration of
the search extended from the beginning of the databases
until November 2022 and included research papers pub-
lished in English only.
3.2 Databases and information sources
The methodology of the search was performed using sev-
eral to find relevant studies in 3 different databases. The
databases were Scopus, PubMed, and Microsoft Academic.
Table 1shows the databases and information sources used
in this review.
4 Deep learning applications in urology
cancers
4.1 Overview
Cancer is a disease that weakens the body and damages its
immune system. Immune systems have limited ability to
fight against cancer, and it appears to be in high ratios in
elders due to weak immunity of their bodies. The body of
elders puts them in danger of experiencing threatening
diseases. Cancer is a severe illness that, if diagnosed early,
doctors and patients can take proper management, reducing
morbidity and mortality rates [11]. Globally, the number of
elders diagnosed with cancer increases day by day. Other
research over the past decades shows that the percentage of
people living with cancer aged 50 years or above has
increased to more than 17%. Nevertheless, this is alarming
that this ratio may exceed 40% after two decades. Data
shows that mortality and morbidity are high in elderly
patients compared to younger counterparts. The response to
cancer treatment is also different in younger and elderly
patients because their bodies have additional capabilities,
and their immune systems are weak compared to younger
patients. Older adults also suffer from many other diseases
affecting their immune system, and recovery healthcare
providers are a significant part of people’s lives and
essential to successful cancer treatment. They prescribe the
medication setup and monitor cancer and patient condition
through proper testing. People select their healthcare pro-
vider according to their needs; requirements, and the best
cancer treatment is chosen according to people’s needs.
Some resources determine the healthcare provided
Fig. 5 A distribution of studies
classified by urology system
diseases and cancers
Table 1 Summary of search results from five databases
Search method Total # of articles found
Scopus 224
PubMed 125
Microsoft academic 32
Subtotal 381
Neural Computing and Applications
123

according to the provision, such as cancer testing sites and
care services locators from government and state cancer/
aids hotlines from Health Resources and Services Admin-
istration (HRSA). Therefore, get guidance from different
drug campaigns and social security administrations
according to the patient and their conditions [12].
It is vital to highlight that transferring information from
narrative Electronic health records (EHRs) into databases
is practically impossible, costly, and time-consuming.
High-volume databases are also only effective in covering
some patient characteristics. However, a viable solution is
associated with DNN-based machine learning or DNN
applied to natural language process (NLP) to extract
information from narrative Electronic health records
(EHRs). In the study, NLP techniques were utilized by
Leyh-Bannurah et al. (2018) for training language models
on industry-standard by transfer learning for EHRs. This
new pipeline offers new capabilities for efficient and pre-
cise data management for clinical research from the nar-
rative documentation [13]. Because of the multiple issues
associated with environmental pollution, cancer has been
one of the first killers of the human population. As it has
been examined, its growth is increasing daily, and some
studies have used visualization and the bibliometric
approach to conduct deep mining. This article has dis-
cussed the different information technologies that have
provided technical methods and opportunities to overcome
cancer problems [14].
4.2 Datasets
The most crucial component of deep learning approaches is
precisely acquiring the appropriate data. A dataset with
adequate training and test data is required for deep learning
architectures. Sometimes, a dataset may need more data to
train deep learning models. In such instances, data aug-
mentation or transfer learning is generally utilized by
performing fundamental visual changes that do not alter the
semantic-level image label. Horizontal rotation, mirroring,
zooming, and cropping are a few examples of operations.
4.2.1 Private versus public
During the specification of this topic, the number of public
datasets is bounded to other medical issues; therefore,
private datasets are employed in most studied approaches.
Using a private dataset in the training or testing phase
compares the proposed method’s accuracy and perfor-
mance to other techniques. Public datasets were utilized in
59 of the research assessed in this study, whereas private
datasets were used in 76 investigations. The prostate cancer
grade assessment (PANDA) challenge and Digital atlas for
cystoscopy are the only open-access resource for urology
cancer diagnosis data sets. We gathered and presented the
datasets we utilized in the evaluated papers since datasets
are the most significant aspect of deep learning methods.
We anticipate that future researchers will find it easier to
use and comprehend by giving all the available data in
urology cancer analysis in a single table. Table 2lists the
datasets utilized by the academic studies in this study.
4.2.2 Imaging modality
Urology cancers are diagnosed using a variety of imaging
modalities based on their type. Computer tomography
(CT), radiographic or histopathological images are the
primary imaging modality used to diagnose urology can-
cers. Endoscopic imaging methods are a new modality and
more extensively utilized than histopathological and radi-
ological imaging for diagnosing bladder cancer because
they allow for faster diagnosis, which is a form of endo-
scopic technology. Endoscopic is the image produced using
the cystoscopy camera.
MRI, ultrasonography, and histopathology are the prin-
cipal imaging modalities for identifying prostate and penile
cancer. Table 1illustrates datasets and imaging modalities
used in academic research to study urology cancer types.
The major studies focus on prostate and bladder cancers, as
shown in Fig. 6a distribution of studies grouped by urology
cancers.
4.3 Renal cell carcinoma (RCC)
It is a kidney cancer that is the most common type in
adults. RCC has three main types: Papillary (pRCC),
Chromophobe (ChRCC), and Clear cell renal cell carci-
noma (ccRCC). ccRCC is the most common subtype of
RCC, and its diagnosis has been improving unceasingly in
developed countries lately. This improvement is mainly
recognized in detecting imaging techniques, such as cross-
sectional imaging, enabling the clinical T1 stage detection
with a 7-cm smaller tumor and having better incidental
renal masses identification, which is malignancy suspi-
cious. Park et al. (2019) [9] aimed to classify molecular
biomarkers related to aggressive clinical T1 stage ccRCCs
of 7 cm and utilized it in developing a risk prediction
instrument to define a treatment plan.
Baghdadi et al. (2020) [11] were able to improve the
feasibility of using AI with image processing to distinguish
oncocytoma from the chromophobe type of renal cell car-
cinoma (ChRCC) on computed tomography imaging.
Another researcher, Tanaka1. [12], used several deep
learning approaches to identify if a minor (B4 cm) solid
renal mass is malicious or not on multiphase different
improved CT, deciding that CNN was able to distinguish a
small (B4 cm) solid renal masses in dynamic CT images,
Neural Computing and Applications
123

Table 2 Summarization of datasets used in the discussed studies
Author
(Year)
Dataset source Availability Imaging modality Description Advantages Disadvantages
Singhal
et al.
(2022)
[15]
Muljibhai Patel Urological Hospital
(MPUH)
Private Core Needle
Biopsy (CNB)
580 CNB slides from 110 individuals
The dataset was split into training (155) and
testing sets (425)
Large sample size
specific to urology
cancers
Limited external
validation due to
private dataset
Wouter
et al.
(2020)
[16]
Prostate cANcer graDe assessment
(PANDA) challenge
Open
access
Core Needle
Biopsy (CNB)
3586 biopsies were used from Radboud
University Medical Center for training and
1201 for testing
Promotes
reproducibility with
open access
Potential lack of
diversity from a
single center
Lin F, Ma
C. et al.
(2020)
[17]
The Cancer Genome Atlas-Kidney Renal
Clear Cell Carcinoma (TCGA-KIRC)
database
Open
access
CT A training dataset was from diagnosed
patients admitted to one hospital between
October 2009 and November 2019
410 patients, (370) training, (20) internal test
dataset, (20) from the TCGA-KIRC
database three-phase renal CT images with
nephrographic phase (NP), pre-contrast
phase (PCP), and corticomedullary phase
(CMP) were included, too, and internally
and externally tested the DL model
Comprehensive data
from multi-phase
imaging
Population bias
possible due to
dataset origin
Baghdadi
et al.
(2020)
[11]
Patients with pathological diagnoses of
oncocytoma and ChRCC was included. CT
scans of 212 renal masses were reviewed,
and images that included the kidney were
selected based on the existence of imaging
data available through the Roswell Park
Comprehensive Cancer
CenterPicture Archiving and
Communication System. The CT scans
were originated from several different
radiology facilities and no standard
protocol was used for the image
acquisition. The early (arterial/cortical or
venous/nephrogenic) and non-contrast
phases were exported in abdominal view,
with the earliest contrast phase is used in
cases with multiple early phases available.
The Roswell Park Comprehensive Cancer
Center
Private CT, Radiographic
images
CT scans of 212 renal masses were
reviewed. A data query was performed on
the internal Picture Archiving and
Communication System database to
retrieve and export de-identified CT
images comprising matrices with 512 *
512 pixels in the axial planes
The CNN was trained and validated to
identify the kidney tumor areas in images
from 192 patients, and 20 patients were
used to test the network and extract
measures
Non-standardized image
acquisition protocols.
Diverse imaging
modalities. Diverse
imaging modalities
Diverse imaging
modalities
Non-standardized
imaging protocols
Non-standardized
imaging protocols
Neural Computing and Applications
123

Table 2 (continued)
Author
(Year)
Dataset source Availability Imaging modality Description Advantages Disadvantages
Park et al.
(2019)
[18]
Department of Urology and Urological
Science Institute, Korea
Private Radiographic,
histopathological
One thousand sixty-nine nephrectomies
were performed for ccRCC of B7
between 2008 and 2014, and 177 cases
were evaluatedThe dataset was randomly
partitioned into two separate training and
validation groups. The prediction models
were built using the training group, which
accounted for 70% of the data (123
individuals, including 28 with active
ccRCC). The validation group (54
individuals, including 12 with active
ccRCC) was utilized to evaluate the
model’s efficacy in predicting active
ccRCC
Mix of radiographic and
histopathological data
May not be
generalizable due
to private dataset
Eminaga
et al.
(2018)
[19]
Digital atlas for cystoscopy Open
access
Endoscopic There were 479 patient cases with 44
urologic findings. The image’s color was
linearly standardized before being
equalized with contrast-limited adaptive
histogram equalization They ended up
with 18,681 photos after rotating them in
10-degree increments and flipping them
horizontally or vertically. A training set
(60%), a validation set (10%), and a test
set (30%) were randomly generated from
the study data set
Large augmented
dataset
Representativeness
may be
questioned
Ikeda et al.
(2020)
[20]
University of Tsukuba Hospital between
February 2017 and July 2018
Private Endoscopic A total of 2102 cystoscopic pictures were
utilized to construct a dataset with a 2:8
ratio of test and training images,
containing 1671 images of healthy cells
and 431 images of tumor tissues
The endoscopic bladder image was acquired
using a TIFF file with a resolution of
1350 91080 pixels and white light
Images had turbidity in the urine, and out-of-
focus images were eliminated
Healthy versus tumor
tissue comparison
Lack of diversity
due to single
institution dataset
Neural Computing and Applications
123

Table 2 (continued)
Author
(Year)
Dataset source Availability Imaging modality Description Advantages Disadvantages
Shkolyar
et al.
(2019)
[21]
Digital atlas for cystoscopy Open
access
Endoscopic One hundred subjects (141 videos) had an
office-based cystoscopy, and the
transurethral excision of bladder tumors
was gathered and annotated
Video frames (n= 611) with histologically
proven papillary bladder cancer were
chosen for algorithm development, and the
tumor was delineated
Based on 95 subjects, a training set was
created (417 cancer and 2,335 regular
frames). Five participants were used to
create a test set (211 cancer, 1,002 regular
frames)
Dynamic video data Focus limited to
one cancer type
Ikeda et al.
(2021)
[22]
Collected at the University of Tsukuba
Hospital
Private endoscopic 2102 cystoscopic images, consisting of 1671
images of normal tissue and 431 images of
tumor lesions
High-resolution imaging Dataset privacy
could hinder
validation
Wu et al.
(2019)
[23]
Collected with IRB approval Private CT CT scans of 123 patients with 129 total
cancers
(pre and posttreatment) (bladder cancer)
The training set consisted of 77 lesions. The
validation set consisted of 10 lesions; The
test set was composed of 42 lesions
Inclusion of pre and
post-treatment scans
Small sample size
limiting
robustness
Tekeuchi
et al.
(2018)
[24]
Collected with IRB approval Private CT 123 subjects of CT scans with 157 MIBC
(Pre and post- chemotherapy) (bladder
cancer)
[mean age: 63 years, range: 43–84 years and
23 females [mean age: 63 years, range:
37–82 years])
Focus on a specific
cancer subtype
Limited cohort size
and demographic
diversity
Takeuchi
et al.
(2018)
[25]
Collected with IRB approval Private MRI,
histopathological
A total of 334 patients underwent 3-Tesla
multiparametric magnetic resonance
imaging (mpMRI) before ultrasound-
guided -MRI targeted transrectal 12-core
prostate biopsy between 2013 and 2017
232 patients were used as training cases, and
102 patients were for the test to examine
the probability of prostate cancer
existence,
Utilization of advanced
imaging techniques
Lack of external
validation due to
privacy
Neural Computing and Applications
123

mainly in the corticomedullary image model. Oberai
(2020) [29] built a workflow that applies convolutional
neural networks (CNNs) to detect lipid-poor, solid, and
different increasing renal masses using multiphase con-
trast-enhanced CT (CECT) images and to evaluate the
resulting network performance. He developed a CNN-
based classifier to diagnose solid malicious renal masses
based on multiphase CECT images.
The CNN-based classifier is established to be trained
and accurately discriminate malignant renal lesions. Lin F.
(2020) examines the performance of various methodologies
of deep learning models on discriminating high- from low-
grade clear cell renal cell carcinoma (ccRCC). He shows
that the CT-based deep Learning model can classify
ccRCC with simple IC in routine clinical practice [17]. The
studies for imaging diagnosis techniques of renal cell
cancer using various deep learning techniques are dis-
cussed in Table 3.
4.4 Bladder cancer
Bladder cancer is one of the most apparent cancers in the
USA. Radical cystectomy gives the best domestic control
for patients with localized muscle-invasive or recurring
non-muscle offensive bladder cancer. However, there is
proper local cancer control, and minimally 50% of patients
have been undergoing cystectomy, which develops the
metastases -within two of the cystectomy years. The early
evaluation of therapeutic effectiveness and prediction of
treatment failure will be constructive for clinicians.
Moreover, in this study, the investigation has been done on
the likelihood that the radionics—dependent predictive
models have been used [30].
As stated above, bladder cancer has universally one of
the most apparent malignancies, with an assessed 430,000
recent diagnoses yearly. As well as the formal diagnosis of
individuals with bladder cancer based on white light cys-
toscopy, almost 2 million cystoscopies have been per-
formed yearly in the USA and Europe. The latest
progression within deep learning-based image processing
might address some of the restrictions of cystoscopy and
TURBT. Some video structures holding histologically
inveterate papillary urothelial carcinoma have been chosen
and manually glossed [31]. Bladder cancer detection using
cystoscopy imaging diagnosis is shown in Fig. 7
Bladder cancer has accounted for 5% of all recent
cancers in the United States of America [14]. In this regard,
the standard treatment method has been used for bladder
cancer, including the radical cystectomy of the bladder.
Moreover, half of the patients undergoing cystectomy had
been taken to have only aggressive cancer at the time of
surgery. It can also be said that the presence of neoplasms
has been blown out to the perivascular tissue, which has
Table 2 (continued)
Author
(Year)
Dataset source Availability Imaging modality Description Advantages Disadvantages
Wong et al.
(2019)
[26]
Institute of Urology, Keck School of
Medicine, University of Southern
California, Los Angeles, United States
Private Histopathological 338 RARP cases, 19 different training
variables
Continence outcomes from a separate set of
consecutive historical RARPs (January
2015 to August 2016)
Detailed collection of
variables
Dataset limited to
postoperative
cases
Bonekamp
et al.
(2018)
[27]
From the Tokyo Medical and Dental
University Ethics Review Committee
(approval number M2017-214)
Private MRI,
Histopathological
316 men with MRI-transrectal US fusion
biopsy
The data were divided into a training data set
(n= 301), used to develop the CAD
algorithm, and two evaluation data sets
(n= 34)
Combination of MRI
and US data
Reproducibility
limited by dataset
privacy
Ma et al.
(2017)
[28]
NA NA CT CT image volumes from 92 patients were
divided into five disjoint subsets
Each of the five subsets is taken as the
testing dataset, randomly selecting one
subset as the validation set for parameter
setting, and the three remaining subsets are
the training set
The use of a cross-
validation setup
Details on data
availability are
lacking
Neural Computing and Applications
123

gone unnoticed at the time of treatment [32]. The bladder’s
urothelial carcinoma is one of the most dangerous urinary
cancers. The white light cystoscopy forms the main stone
for diagnosing this disease. The suspicious scratches of the
bladder have been endoscopically resected for histopatho-
logic investigation.
Furthermore, in the case of the UCB, there are some
cellular structures and nuclear entrances [31]. Risk strati-
fication has been significant in evaluating the optimal
treatment for some patients with bladder cancer. This type
of cancer has vastly divided into two risk sets depending on
the invasion of the bladder’s muscular walls. For some
patients diagnosed with MICB, therapy is radical cystec-
tomy and pelvic lymph node dissection (LN).
Conversely, efforts have been to further risk stratify
MIBC based on contrary or irregular histology and clinical
staging. The study aims to mature a model based on
Artificial Intelligence [33]. There are some computerized
decision support systems for muscle offensive bladder
cancer treatment response assessment using image infor-
mation. This information has been obtained from the
computed tomography investigations, which had been
developed within the laboratory. On the other side, the
CDSS-T tool approximates the likelihood that the patient
has mainly responded to neoadjuvant chemotherapy; fur-
thermore, for all the pre-treatment and post-treatment CT
scans of the pelvis that has been attained along some of the
healthcare. The stage of pathology cancer has been utilized
in the standard [34].
Urothelial cell carcinoma is one of the most apparent
kinds of bladder cancer. It is one of the most significant
challenges in oncology and urology because of their
affluence to recur and progress. The WHO presents the
grading system dividing the histological band of UCC into
three grades. Moreover, it can also be said that all the
exported images had been annotated by utilizing the free-
hand explanation tool, which has been developed within
the house. In addition, all the sectors holding UCC had
been demarcated [35]. As aforementioned, radical cystec-
tomy has been taken as the gold standard for treating
bladder cancer; in this regard, half of the patients develop
metastases.
At the same time, radical cystectomy enhances the
survival of patients with domestically progressed bladder
cancer. In some clinical trials, the downstaging with drugs
before having surgery had been shown to have essential
survival benefits. Moreover, some of the patients, along
with progressed disease, would benefit from neoadjuvant
chemotherapy and some of its disadvantages. DL-CNN
models are trained and tested in their paper for the treat-
ment [36]. In reaction to the neoadjuvant chemotherapy for
cancer that starts to spread in an individual’s bladder,
several physicians interpreted the CT examinations after
giving the chemotherapy. In this article, the CAD system
has been observed to treat bladder cancer. Different
examples have been shown for the pre-treatment and the
bladder’s post-treatment cancer in the form of specific
pairs; this has been clearly shown through the help of
pictures in this article. The use of CAD and some of the
radiometric features can quickly improve the CT’s accu-
racy in identifying the complete response [37]. The spatial
and temporal development has numerous sites within the
clinical properties of bladder cancer and some of the fre-
quencies at various intravesical sites when the diagnosis is
made. In addition, some of the standards presented within
the study for the treatment of non-muscular bladder cancer
Fig. 6 A distribution of studies
grouped by urology system
cancers
Neural Computing and Applications
123

have an endoscopic transurethral resection of the tumor in
the bladder. In this regard, cystoscopy is one of the most
effective treatments for bladder cancer. The treatment
through this method has been assessed through cystoscopy
images within the convolutional neural network [38].
The studies for imaging diagnosis techniques of bladder
cancer using various deep learning techniques are dis-
cussed in Table 4.
4.5 Prostate cancer
Prostate cancer is commonly found in the USA, and males
are the ones that have been diagnosed with this on a large
scale. It has been discussed again that MRI is an increas-
ingly used technique to evaluate prostate cancer. A con-
volutional neural network (CNN) is the predominant
method of machine learning to recognize medical images.
This study described the deep learning approach, the subset
of AI for automatic localization and prostrate segmentation
[39]. A deep learning algorithm has been explained in
Table 3 Summarization of recent works employing deep learning for the renal cell carcinoma diagnosis
Author
(Year)
Study goal/research question Dataset Method used Primary findings/
used tool
Characteristics Limitations
Lin et al.
(2020)
[17]
To define the effects of several
deep learning model
performance methodologies
for detecting high- from
low-grade
410 patients,
(370) training,
(20) internal
test dataset,
(20) from the
TCGA-KIRC
database
ResidualNetwork
(ResNet)
Internal test
accuracy was
73.7%, external
test accuracy
was 77.9%,
Internal test
negative
predictive value
was 69.2%, and
external test
negative
predictive value
was 87.1%. The
internal test
positive
predictive value
was 84.6%,
external test
negative
predictive value
was 73.5%
CT-based DL
model can be
conveniently
applied for
grading RCC
with image
cropping (IC)
in clinical
practice
The small sample
size does not
compare
performance
with other
models;
2-dimensional
image analysis
was used, not
considering the
tumor size
Baghdadi
et al.
(2020)
[11]
To analyze the feasibility of
using image processing and
AI to distinguish
oncocytoma from the
chromophobe subtype of
renal cell carcinoma
(ChRCC) on computed
tomography imaging
Review CT scans
of 212 renal
masses, to test
the network
segmentation
and extract
radiographic
measures for
tumor subtype
differentiation
Convolutional
Neural
Networks
(CNNs)
95% accuracy,
100%
sensitivity, and
89% specificity
It can be used in
enhancing the
accuracy of
diagnosis in
CD117( ?)
renal tumor
biopsies
to avoid the
need for
diagnostic
resection
Limited numbers
of images
available for
training the
CNN for this
rare kidney
tumor
segmentation,
lack of multi-
institutional
validation
Park et al.
(2019)
[18]
To identify molecular
biomarkers related to the
aggressive clinical T1 stage
ccRCCs of B7 cm,
Among 1069
nephrectomies
performed for
ccRCC of B7
between 2008
and 2014, 177
cases were
evaluated
DNN, DCN,
DBN, and RNN
Six parameters
were used, and
the accuracy
reached 0.760
and 0.759 in
DNN and
logistic
regression
models,
respectively
Support
stratifying
patients with
clinical T1
stage ccRCC
Can be
substantial,
based only on
postoperative
formalin-fixed
paraffin-
embedded
tissue (FFPE)
samples
Neural Computing and Applications
123

equivalent or the most remarkable performance to the
pathology experts in predicting muscle-invasive bladder
cancer and the subsets from the histopathological slides.
Here, it can also be said that four of the pathologists given,
along with 186 tiles, chose to act randomly and achieved a
general correctness rate of almost 40%. N of the individ-
uals matches the exactness rate of the algorithm in subtype
categorization. The initial trials and loss function alterna-
tion in these and some other genes have been linked with
some aggressive diseases [40]. The Gleason grading has
been one of the most powerful predictors of the prognosis
for several patients with prostate cancer since the 1960s. It
has been discussed in the article that its application is
challenging and requires proper protocol and procedure.
Different patients were made under observation for the
study, and then the results came out to be much more
positive. Throughout the manuscripts, it has been seen that
different kinds of metrics were being used for the annotator
variability [41]. The convolutional neural network is the
predominant machine learning method for recognizing
medical images; this study has described the deep learning
approach, a subset of AI, for automatic localization and
prostrate segmentation [39].
The objective of this article was to stimulate medical
adoption, assess performance, and provide consistency of
machine learning for detecting prostate cancer in various
patients using effective methods. U-Net is the method that
has been used, and it showed almost the same kind of
performance as the PI-RADS. An automated model has
been proposed in the article for prostate cancer. P-DNN has
been used for the MR prostate segments where different
low-level images have been lined up. Other techniques and
technologies have been used to detect and treat prostate
cancer, increasing daily on a large scale [42]. Men sus-
pected of possessing significant prostate cancer (sPC) tend
to undergo the necessary MRI for the prostate. However,
deep learning’s potential for providing diagnostic support
for the interpretation of humans needs further evaluation.
Schelb et al. (2019) aimed to compare the clinical assess-
ment’s performance to a deep learning system enhanced for
segmentation, which is trained using T2 diffusion and
weighted MRI to segment and detect suspicious lesions for
PC. Three hundred twelve men were evaluated in their
study, and conclusions were derived from the set. U-Net is
trained with T2 diffusion, and weighted MRI seems to
achieve similar performance and capability to Data System
assessment and Prostate Imaging Reporting [43].
Prostate cancer’s early detection raises the likelihood of
a patient’s survival. Reda et al. (2017) created a non-in-
vasive automated system for CAD, or computer-aided
diagnosis of the segments of prostate cancer on DW-MRI
or diffusion-weighted magnetic resonance images that are
acquired at different b0 values, which seems to estimate its
ADC or apparent diffusion coefficient and its classifies
their descriptors with a deep learning network that is
appropriately trained. It should be noted that for a more
effective and robust evolution, the attributes are fused with
a general probabilistic shape. The experiments in their
study obtained 92.3% accuracy, 100% specificity, and
83.3% sensitivity, which indicates that the presented sys-
tem can be considered reliable [44]. Notably, localization
and early identification of prostate tumors generally pose a
significant challenge and issue to medical communities.
Several imaging techniques, including PET, have shown
some success. However, no accurate and robust solution
has yet been developed. Rubinstein et al. (2019) used an
unsupervised learning approach to detect prostate cancer
foci in different dynamic PET images. The proposed
method seems to extract three feature classes from the
imaging data: in-depth features, kinetic biological, and
statistical features, which are learned using a convolutional
autoencoder that is deep stacked. The proposed method
generates satisfying outcomes for large cancer foci [45].
Several methodological changes to diagnose prostate can-
cer, including MRI, prostate ultrasonography, and digital
rectal examination, have seemingly evolved over the years.
Fig. 7 Representation of bladder cancer
Neural Computing and Applications
123

Hong K (2020) focuses on AI-based deep learning clinical
applications for PI-RADS classifications for assisting
multiparametric MRI. Image classification AI system based
on deep learning is proposed. This system seems to assign a
PI-RAD score to a lesion segmented and detected by a
radiologist. It is vital to note that the optimal utilization of
AI is still in clinical practice. A compelling study with
Table 4 Summarization of recent works employing deep learning for image processing to diagnose bladder cancer
Author
(Year)
Study goal/research
question
Dataset Method used Primary
findings/used
tool
Characteristics Limitations
Eminaga
et al.
(2018)
[19]
To develop
computer-aided
diagnosis tools to
extract features
using deep learning
Digital atlas for
cystoscopy,
consists of 479
images, covers
44 cystoscopy
findings
NN models
ResNet50,6
VGG-16,7,
VGG-19,7
Xception9 and
InceptionV3,8.
To classify the
cystoscopic
images
The highest F1
score has been
achieved by
the Xception-
based model
(99.52%) and
ResNet50
(99.48%)
Using cystoscopic
images, deep-
learning models can
classify carcinoma
in situ from cystitis
Not enough datasets,
as only cystoscopic
images are used to
diagnose
Ikeda
et al.
(2020)
[20]
Image classification,
binary classes, the
healthy and tumoral
urothelium
One thousand six
hundred
seventy-one
images of
healthy cells
and 431 images
of tumor tissues
DCNN is used to
detect abnormal
images of the
lesion area. The
Inception-v3
model is used to
extract features
83.7%
specificity and
93.0%
sensitivity
Used a pre-trained
DCNN model and
the supervised
learning for feature
extraction and
classification
Relatively few
training samples
were used
Shkolyar
et al.
(2019)
[21]
Proposed a deep
learning algorithm
to detect augmented
cystoscopic bladder
cancer, using image
segmentation of
benign and cancer
It validated 54
patients, 95
patients for
training and five
for testing.
Training with
417 cancer and
2,335 normal
‘‘TUMNet’’,
‘‘CystoNet’’,
image analysis
platforms. It
was based on
convolutional
neural networks
Specificity was
99% per
tumor.
Sensitivity was
90%. Per-
frame
sensitivity was
88%,
The proposed
approach was
accurately detecting
papillary bladder
cancers. With high
specificity and
sensitivity. CystoNet
may improve the
diagnostic yield of
cystoscopy and the
efficacy of TURBT
The training set was
small. Precision is
suitable but not
sufficient to
develop a support
system. Bladder
tumors were
determined
histopathologically
Ikeda
et al.
(2020)
[22]
Deep learning is used
to diagnose bladder
cancer
2102 cystoscopic
images,
consisting of
1671 images of
normal tissue
and 431 images
of tumor lesions
A CNN model
with ImageNet
to learn
cystoscopy
images was
performed 1637
times in
supervised
learning
Specificity of
94.0%, and
sensitivity of
89.7%,
Enhance the accuracy
of bladder cancer
diagnosis and
treatment
The proposed
technique should
be verified in
clinical use
Wu et al.
(2019)
[23]
Comparing different
DL-CNN models to
predict response to
treatment in bladder
cancer
CT scans of 123
patients with
129 total
Cancers (pre
and
posttreatment)
Multiple DL-
CNN models
with structure
modification
and layer
freezing
Specificity of
80%, Accuracy
of 70%, and
Sensitivity of
60%
––
Tekeuchi
et al.
(2018)
[24]
Examining whether
CDSS-T based on
responded
neoadjuvant
chemotherapy
enhances patient
identification
123 subjects with
157 of CT scans
MIBC (Pre and
post-
chemotherapy)
Multilayer
Artificial Neural
Network (ANN)
Using CDSS-T:
0.80, 0.74
AUCs Without
CDSS-T: 0.77
AUC
––
Neural Computing and Applications
123

several patients with urologists and radiologists is needed
[46].
Deep learning algorithms have gained enormous success
in cancer image segmentation. However, knowing the
variances between deep learning and human understanding
is authoritative. Both MRI results of cancer-located cells
are compared with pathologist’s identified cancerous
positions. A sample of 307 images helped determine the
location of cancer cells which was later developed to
identify the location of cancer with the help of 3D recon-
struction of pathological images. Deep learning focused on
Lymphocyte aggregation and dilated prostatic ducts found
during the diagnosis of prostate cancer. It would not be
wrong to say that deep learning algorithms help achieve
more significant and specific targets with a cancer pres-
ence. Deep learning tends to detect even if the cancer is not
visible, where the results of tumor segmentation are shown
in Fig. 8[47]. The studies for imaging diagnosis techniques
of prostate cancer using various deep learning techniques
are discussed in Table 5.
4.6 Other types of urology cancer
Another type of cancer that can be diagnosed in the urology
system, penile cancer, in males, is relatively considered to
be a rare neoplasm. It is related to all efficient psychosocial
issues, which are entirely linked with the diagnosis and the
treatment; this is cancer in which under-treatment or even
over-treatment is directly associated with physical and
psychological problems. MRI has been discussed in this
article for the diagnosis, which has been further recom-
mended in the guidelines of EAU. In addition to that, MRI
also helps to add up all the relevant information. In short,
MRI is the diagnostic imaging discussed for penile cancer
and is further poised to be among the stop-shops for the
different stages of penile cancer [48].
A systematic literature review was conducted on
machine learning applications for detecting and diagnosing
urology cancers. The following papers provide valuable
insights into the advancements made in this area, ordered
from past to present: in the study by Cha et al. (2017) [30],
various methods, including lesion segmentation, ROI cre-
ation, DL-CNN, EF-SL, and RF-ROI were employed. The
study conducted ROC performance analysis and calculated
the area under the curve (AUC) to assess the performance
of these methods. Marit et al. (2019) [31] analyzed bladder
cancer using TURBT or flexible clinical cystoscopy. The
deep learning algorithm, CystoNet, proved effective with
clinician experience. CystoNet has the potential to aid in
training. Bono et al. (2020) [40] utilized the CRPC method
and highlighted its high AUC values across subgroups and
for each image. The study merged the overall outcome into
Cohorts A and B, primarily benefiting men. Cha et al.
(2016) [32] retrospectively analyzed 62 cases from
abdominal imaging datasets, demonstrating examples of
DL-CNN segmented bladder cancer. Ilaria et al. (2020)
Fig. 8 Results of tumour segmentation
Neural Computing and Applications
123

[35] achieved automated urothelium classification using a
segmentation and categorization networks. The results
were obtained through the annotation of a dataset com-
prising 32 billion pixels. Cha et al. (2019) [37] investigated
the use of chemotherapy treatment with MVAC in 300
individuals with bladder cancer. The results indicated that
half of the population with bladder cancer had a pathologic
stage while undergoing chemotherapy. Wu et al. (2019)
[36] employed DL-CNN models for bladder cancer treat-
ment, adjusting the weights randomly. The study demon-
strated that automatic adjustments in DL-CNN structure
and weights effectively removed certain structures.
Lin et al. (2020) [17]. The study focused on deep
learning model validation to improve the reliability and
robustness of the models. The authors proposed a novel
validation method that performs better than traditional
methods. Tanaka et al. (2020) [12] The study aimed to
identify risk factors in urology cancers and developed a
risk prediction tool that proved valuable in predicting and
understanding the risk factors associated with the pro-
gression of the disease. Oberai et al. (2020) [29]. The study
focused on enhancing CT learning for improved cancer
diagnosis. The authors incorporated learning algorithms to
analyze CT scans, leading to more accurate and predictive
diagnostic outcomes. Park et al. (2019) [18]. The study
investigated nuclear grade prediction for improved cancer
assessment. The prediction of atomic grade provided
valuable insights for evaluating cancer aggressiveness and
determining appropriate treatment strategies. Rodrigo et al.
(2018) [4]. The study focused on renal tumor prediction to
aid in understanding potential treatment options. Analyzing
renal tumor applications contributed to successful tumor
management and improved patient outcomes. Jun et al.
(2019) [47]. The study employed 3D reconstruction of MRI
images for improved prostate cancer diagnosis. 3D recon-
struction provided better results and enabled the identifi-
cation of cancer clues not visible to the human eye.
Junichiro et al. (2018) [27]. This study aimed to develop
AI-machine learning-based diagnosis methods. The results
significantly improved diagnostic capabilities through AI
and machine learning. Baghdadi et al. (2020) [11]. The
study focused on the automatic discretion of renal cell
classification. The proposed method successfully achieved
renal cell discretion, contributing to improved diagnosis
Table 5 Summarization of recent works used deep learning for image processing to diagnose prostate cancer
Author
(Year)
Study goal/
research
question
Dataset Method used Primary findings/used tool Characteristics Limitations
Takeuchi
et al.
(2018)
[25]
Used deep
learning with
multilayer
ANN to
Predict
prostate
cancer
Collected from
334 patients of
transrectal
biopsy
(mpMRI)
Multilayer
ANN
22 different selected variables – –
Wong et al.
(2019)
[26]
Detect the
recovery of
urinary
continence
after RARP
338 RARP
patients
3ML
algorithms
ML algorithms beat the
statistical regression model in
predicting biochemical
recurrence following RARP
(AUC 0.865). RF tree, K-NN,
and LR’s accuracy were
0.953, 0.976, and 0.976,
respectively
Surgeons effective
APMs achieved
higher continency
rates at 3 and
6 months after
RARP
–
Bonekamp
et al.
(2018)
[27]
Lesion
segmentation
and radionics
analysis
Patients of 316
using (MRI
transrectal
fusion biopsy)
ML-based
radionics
models
ADC (AUC global = 0.84;
AUC zone-specific B0.87)
RML (AUC global = 0.88,
p= 0.176; AUC zone-
specific B0.89, pC0.493)
– ML versus
ADC
comparison
showed no
significant
different
result
Ma et al.
(2017)
[28]
Use deep
learning to
segment the
prostate on
CT images
CT image
volumes from
92 patients
were divided
into five
disjoint subsets
CNN-based
prostate
segmentation
and multi-
atlas label
fusion
The Dice similarity coefficient
(DSC) for the whole volume is
86.80%, with a minimum of
75.26% for the base, and
79.91% for the apex
The automatic
method performs
well and can be
applied to diagnose
various types of
cancers
A small
number of
samples were
used in the
study
Neural Computing and Applications
123

and treatment of renal cell cancer. Yang et al. (2020) [49]
Segmentation through neural networking was employed in
this study. Using neural networking techniques improved
the precision and accuracy of tumor segmentation,
enhancing treatment outcomes. Guanyu et al. (2020) [50].
The study introduced an enhanced neural networking
approach for cancer treatment. Harmo et al. (2020) [33],
the study used a deep learning model to spatially resolve
prediction maps for evaluating the research topic. The
authors achieved positive effects in most of the 307
patients included in the study. Wang et al. (2019) [14]
explored various information technologies in cancer treat-
ment, such as cloud computing, IoT, and decision support
system technology. These technologies collectively con-
tribute to the improvement of cancer treatment outcomes.
Park et al. (2018) [18]. The study investigated the use of
machine learning algorithms for predicting prostate cancer
aggressiveness. The authors achieved high accuracy in
predicting prostate cancer aggressiveness, which can help
clinicians make more informed treatment decisions.
Cha et al. (2019) [30],the study administered
chemotherapy treatment to 300 individuals with bladder
cancer. The results revealed that half of the population with
bladder cancer had a pathologic stage while undergoing
chemotherapy. Atsushi et al. (2020) [38]. The study
included approximately 2102 images diagnosed with
bladder cancer, consisting of normal and tumor tissue
samples. The dataset evaluation showed that 50% of the
images had positive effects, and the other 50% had adverse
effects. Wu et al. (2019) [36] DL-CNN models were uti-
lized for treatment in this study. The weights of the DL-
CNN were randomly adjusted, and the structure of DL-
CNN was effectively used. The results demonstrated the
automatic removal of specific structures due to the imple-
mented changes. Hadjiiski et al. (2020) [34]. This study
focused on using a CAD system to predict bladder cancer
treatment. The results indicated that properly utilizing a
CAD system could enhance the accuracy level of CT scans.
Satheesh et al. (2020) [48] conducted a study focusing on
the MRI method for the treatment of penile cancer. The
study highlighted the widespread use of MRI in the man-
agement of penile cancer cases. The MRI method was
found to be effective in providing accurate treatment for
penile cancer. This emphasizes the importance of utilizing
MRI as a valuable tool in the diagnosis and treatment
planning for penile cancer patients. Satheesh et al. (2020)
[48]. The study discussed and utilized MRI methods for
treating penile cancer on a large scale. MRI proved to be a
valuable tool for accurate treatment of penile cancer.
Prathamesh et al. (2020) [51]. This study employed the
DNN method to distribute immune cells within the tumor.
The DNN method showed promise in enhancing the
understanding and treatment of cancer. Reda et al. (2017)
[44] introduced a computer-aided diagnosis (CAD) system
that utilized diffusion-weighted magnetic resonance
imaging (DW-MRI) and a Deep Learning Network. The
presented method demonstrated high accuracy, specificity,
and sensitivity, establishing it as a reliable tool.Ma et al.
(2017) [28] employed deep learning techniques, including
Convolutional Neural Networks (CNNs) and Multi-Atlas
Segmentation, for prostate segmentation in computed
tomography (CT) scans. The method achieved an 86.80%
Dice similarity coefficient, indicating effective automatic
segmentation.
Leyh-Bannurah et al. (2018) [13] integrated Deep
Neural Network (DNN) and Natural Language Processing
(NLP) techniques to develop a scalable solution for effi-
cient and precise data management in Electronic Health
Records (EHRs). The proposed approach demonstrated
generalizability to other EHRs, offering improved data
handling capabilities. Schelb et al. (2019) [43] applied the
U-Net method for detecting and treating prostate cancer.
Comparing it to the Prostate Imaging Reporting and Data
System assessment, U-Net exhibited a sensitivity of 99%,
showing its potential as an effective tool. Lindgren et al.
(2019) [52] employed Convolutional Neural Networking
(CNN) to achieve highly precise automated segmentation
of bone volumes and skeletal structures. This method
demonstrated accurate results and improved skeletal seg-
mentation. Ruud et al. (2019) [53] explored the use of deep
learning techniques as a cost-effective alternative to mag-
netic resonance imaging (MRI) for prostate segmentation
in transrectal ultrasound (TRUS) images. The study
showed that deep learning methods could achieve fast and
accurate prostate segmentation. Arif et al. (2020) [54]
introduced a deep learning computer-aided detection
(CAD) method for the segmentation and identification of
clinically significant prostate cancer (csPC) and the con-
firmation of low-risk cancer (LRC) in patients undergoing
surveillance. Their presented deep learning method
demonstrated effectiveness and promising results, high-
lighting its potential in improving the accuracy and effi-
ciency of prostate cancer diagnosis and surveillance. This
approach holds promise for enhancing patient care and
management in the field of prostate cancer research. Jae
et al. (2018) [55] proposed the use of a Deep Belief Net-
work combined with Dempster-Shafer theory (DBN-DS)
for predicting patient data in the context of understanding
diseases. Their method, DBN-DS, outperformed other
approaches, enabling the prediction of more patient data
and providing additional possibilities for comprehending
the disease. The area under the curve (AUC) of DBN-DS
was 0.777, compared to 0.620 for the Partin tables, indi-
cating its superior predictive performance.
Lucas et al. (2019) [56] employed a Convolutional
Neural Network (CNN) for the comparison of non-typical
Neural Computing and Applications
123

and malignant areas in prostate pathology. Their approach
achieved an accuracy of 92% in distinguishing between
these areas, with a sensitivity of 90% and a specificity of
93%. Furthermore, the classification between Gleason
pattern (GP) C4 and GP B3 exhibited an accuracy of
90%, with sensitivity and specificity values of 77% and
94%, respectively. These findings highlight the effective-
ness of CNN-based methods in accurately identifying and
distinguishing different areas of prostate pathology. Eirini
et al. (2019) [57] utilized Positron Emission Tomography/
Computed Tomography (PET/CT) for automated and
manual volume segmentations in urology cancers. The
study reported a segmentation result of 0.78 for automated
volume segmentation and 0.79 for manual volume seg-
mentation. Automated PET/CT utilized total lesion uptake
and abnormal voxels to generate these results. Addition-
ally, promising outcomes related to overall survival were
observed when values were obtained manually. Ohad et al.
(2019) [58] focused on Gleason grading using the Aperio
ScanScope CS scanner. Their model achieved an accuracy
of 91.5% for coarse segregation of images, with a sensi-
tivity level of 0.93 and specificities of 0.90. The study also
reported promising results regarding overall survival when
values were obtained manually. Yoichiro et al. (2019) [59]
utilized deep learning algorithms to acquire explainable
features from histopathology images. Their approach pro-
vided an automated and reliable means of feature acqui-
sition, potentially aiding in the analysis of cancerous
tissues.
Yan et al. (2019) [42] employed the Deep Neural Net-
work P-DNN for effective feature extraction in the seg-
mentation of prostate MR images. P-DNN demonstrated
successful feature extraction, showcasing its utility in
prostate image analysis. Atsushi et al. (2020) [38] suc-
cessfully classified prostate cancer samples using the
Gleason and tissue microarray approaches. This method
proved to be a reliable and effective means of prostate
cancer classification. Wang et al. (2019) [14] explored the
application of various information technologies, such as
cloud computing, IoT, and decision support systems, in
cancer treatment. Their study highlighted the usefulness of
these technologies in improving cancer treatment out-
comes. Rubinstein et al. (2019) [45] presented an unsu-
pervised learning approach for detecting large foci cancer
in PET scan imaging. Their algorithm showed promising
results, particularly in cases where tomographic devices
Table 6 Summarization of review articles in urology cancers recognition and classification using AI
Authors
(Year)
Searched keywords Methodology\type Coverage
period
Coverage quantity Database
source
Rodrigo
SI. et al.
[67]
‘‘Renal cell carcinoma’’,
‘‘urolithiasis’’, ‘‘bladder cancer’’,
‘‘prostate cancer’’ combined with
‘‘machine learning’’, ‘‘deep
learning’’ and ‘‘artificial neural
network’’
Selective search,
Literature review
May 2019 43 studies were included PubMed
MEDLINE
database
Milap et.
al. [68]
‘‘Urology,’’ ‘‘artificial intelligence,’’
‘‘machine learning,’’ ‘‘deep
learning,’’ ‘‘artificial neural
networks,’’ ‘‘computer vision,’’
and ‘‘natural language
processing’’
Selective search, iTRUE
study
NA 47 articles were selected NA
Misgana
et. al.
[69]
‘‘Cystoscopy’’, ‘‘deep learning’’,
‘‘machine learning’’ and
‘‘convolutional neural network’’
Literature search,
comprehensive review
October
2019
There was one conference, one
abstract, and two articles
addressing the use of AI
algorithms in cystoscopic
image recognition
PubMed,
MEDLINE
Enrico et.
al. [70]
NA A non-systematic review
of the literature
NA NA MEDLINE,
PubMed, the
Cochrane
Database,
and Embase
Lubbad et.
al.
(2023)—
This
research
The term ‘‘artificial intelligence’’,
‘‘neural network’’, ‘‘machine
learning,’’ or ‘‘deep learning’’
combined with the phrase
‘‘urology cancers’’
Literature research using
different databases.
Then using PRISMA
methodology,
systematic survey
Until
November
2022
Near 381 articles were
identified. Forty-eight articles
were included in the
qualitative synthesis process
PubMed
MEDLINE
database
Neural Computing and Applications
123

may not clearly visualize the foci. Alexander et al. (2020)
[39] employed multiparametric magnetic resonance imag-
ing (mpMRI) and Deep Learning Convolutional Neural
Networks (CNNs) for the automatic identification and
segmentation of the prostate organ. This deep learning
CNN approach proved to be efficient and quick. Amogh
et al. (2020) [60] focused on using deep learning models to
improve the accuracy of lung cancer detection and classi-
fication. Their proposed method achieved high accuracy in
distinguishing between benign and malignant nodules.
Schelb et al. (2020) [61] employed deep learning, in con-
junction with the U-Net method, achieving performance
comparable to the Prostate Imaging Reporting and Data
System assessment. This approach shows promise in
improving prostate cancer detection and diagnosis. Yang
et al. (2020) [49] developed a deep learning-based model
for lung nodule classification using CT images. Utilizing a
hybrid network architecture, their model accurately dif-
ferentiated between benign and malignant nodules.
These studies collectively demonstrate the advance-
ments made in machine learning applications for detecting
and diagnosing urology cancers and lung nodules. Inte-
grating deep learning techniques, such as CNNs and DNNs,
has shown promise in improving accuracy, efficiency, and
automation in cancer detection and classification, con-
tributing to enhanced clinical decision-making and early
diagnosis.
5 Discussion
We examined the most recent studies using deep learning
algorithms to detect and diagnose urological cancer. We
gathered all the works and grouped them into four key
categories based on the common types of urology cancer:
bladder cancer, prostate cancer, renal cell carcinoma, and
other types of cancers in the urology system. Most of the
recent research focused on diagnosing prostate cancer (25
articles), and many researchers analyZed bladder cancers
(15 studies); other researchers detected renal cell carci-
noma (8 papers). The last category includes different types
of cancer in the urology system, which get less attention
than the first three main types due to their widespread
worldwide.
A summary table for each category is presented with
distinct aspects. To create a more comprehensive com-
parison, we put the works in tables. These tables contain
several columns that provide datasets, detecting methods,
and other information. From the presented tables, it is
noticed that deep learning techniques, such as convolu-
tional neural networks (CNN) and Multilayer Artificial
neural networks (ML-ANN), are the most used methods by
researchers to diagnose urology cancers because they are
exceptionally good at finding patterns in vast amounts of
data, deriving relationships in complex information, results
in more accurate disease diagnosis. Dataset is crucial in
diagnosing cancers using deep learning techniques.
Researchers use several types of images to diagnose urol-
ogy cancers. CT, Radiographic, endoscopic, and
histopathological are the most recent imaging modalities
researchers used in diagnosing urology cancers. CT is the
most used imaging modality in detecting renal cell carci-
noma. Endoscopy is the image produced from cystoscopy,
and it is recently widely used for diagnosing bladder can-
cer. MRI and histopathological images are the most used in
recent research for diagnosing prostate cancer. In terms of
the high accuracy rates achieved in different papers, which
exceeded 90% on average, we can say deep learning
approaches are promising in cancer diagnosis, it also has
some challenges. One of the critical challenges is acquiring
a suitable dataset since most of the datasets used by
researchers are private, and very few public efficient
datasets are available for researchers. The deep learning
architectures are unlikely to deliver effective results on
insufficient datasets because they are prone to overfitting
after a certain point. With extensive, high-quality training
sample datasets, data-hungry deep learning algorithms can
yield considerable performance improvements.
While deploying automated systems in urological cancer
diagnostics requires rigorous certification and regulation
clearance, clinical trials have been conducted to satisfy
those needs. AI diagnosis of prostate cancer biopsies [62]
and multimodal data [63] are pioneering attempts. The
community perspective is leaning towards the power of
machine learning-based systems beyond being hype [64].
Real-life challenges with successful outcomes, such as
PANDA challenge, encourage this view [65]. Some
industry-scale software that are technology-ready to be
deployed in the field, such as Dr. Answer, led by the
National IT Industry Promotion Agency of Korea, have
been developed and are in experimental use [66]. However,
as of today, machine learning applications in urology
cancer diagnostics is in the infantry period, and a routine
translation to clinical use will probably need further
attempts.
6 Conclusion and future work
While conducting this research, it is apparent that the topic
is promising in the research area, despite having a few
numbers of related research articles. Most of the research
found is newly published from 2018 to 2022. However,
when we expanded the research scope to include every
research in machine learning applications in urology can-
cer, we found some papers from 2005. Still, they needed to
Neural Computing and Applications
123

be deeply involved in the subject’s core. This research has
been compared to the most recent published review papers,
as shown in Table 6; we have found that this study is the
most recent review based on the PubMed Medline, Scopus,
and Microsoft academic databases, using the primary
resource of articles up to November 2022. A robust
methodology (PRISMA) was performed to achieve the
most relevant articles. Unfortunately, most of the datasets
were private; they are rare due to the particularity of this
research. Researchers face a significant challenge in col-
lecting, refining and preparing a substantial-high-quality
dataset. Table 5summarizes AI and deep learning appli-
cations used in Urology cancers. Most of the reviewed
papers in this research proposed cystoscopy as a standard
procedure used in diagnosing Bladder cancer, and it might
be significant to employ it in diagnosing other types of
cancer in the urology system. Therefore, this topic is
chosen for future research, increasing the urologist’s effi-
ciency in diagnosing urology cancers and boosting accu-
racy to a high rate.
•An accurate preliminary result prediction would assist
urologists in improving patient selection, selecting
optimal treatment choices, and personalizing patient
therapy.
•Besides, deep learning and computer-aided detection
leverage visual attributes to assist radiologists in
detecting malignancies.
•Utilizing AI models for urology cancer has not been
extensively researched to determine their clinical
performance.
In this paper, we thoroughly review the applications of
machine learning algorithms in detecting and diagnosing
urology cancers. To provide a clear and concise analysis of
the advantages and disadvantages of machine learning
algorithms in urology cancer detection and diagnosis, we
can address the advantages as follows:
•Improved accuracy: machine learning algorithms have
demonstrated the ability to achieve high accuracy rates
in detecting urology cancers, outperforming traditional
methods.
•Early detection: machine learning algorithms enable
the detection of urology cancers at early stages, leading
to improved prognosis and treatment outcomes.
•Automated analysis: these algorithms automate the
analysis of medical images and data, reducing the
burden on healthcare professionals and potentially
increasing efficiency.
•Integration with clinical workflow: machine learning
algorithms can be integrated into existing clinical
systems, facilitating seamless adoption in healthcare
settings.
•Personalized medicine: by analyzing patient-specific
data, machine learning algorithms have the potential to
contribute to customized treatment plans and optimized
patient care.
On the other hand, the Machine learning algorithms
have some limitations due to the following:
•Need for large datasets: machine learning algorithms
require large datasets for training, which may pose
challenges in urology cancer detection due to limited
availability or privacy concerns.
•Interpretability: some machine learning algorithms,
such as deep learning models, are considered black
boxes, making it difficult to interpret and explain their
decision-making process, which can be a concern in
clinical settings.
•Generalizability: the performance of machine learning
algorithms heavily relies on the quality and represen-
tativeness of the training data, limiting their generaliz-
ability to diverse patient populations or different
healthcare institutions.
•Validation and clinical integration: before wide-
spread adoption, machine learning algorithms for
urology cancer detection and diagnosis need rigorous
validation studies and consideration of practical aspects
such as regulatory compliance and integration with
existing clinical workflows.
By addressing these advantages and disadvantages, our
paper comprehensively evaluates machine learning algo-
rithms in urology cancer detection and diagnosis, offering
valuable insights for researchers, clinicians, and healthcare
policymakers. Further research is required to validate the
performance of AI models in urology cancer detection in
the early stages. Therefore, as future work, a plan has been
made to develop a new AI model that uses deep learning
techniques to identify and recognize tumors in real time.
Multiple stages of real-time experiments will be done to
validate the clinical performance of the proposed model.
Author’s contribution ML wrote the main manuscript text, DK, AB,
and BA prepared Tables 1,2and 3, and UN and IP prepared Figs. 2
and 3. All authors reviewed the manuscript.
Funding Not Applicable.
Data availability Data is available on request from the authors.
Declarations
Conflict of interest The authors state they have no known conflicting
commercial interests or personal relationships that could have influ-
enced this article’s work.
Neural Computing and Applications
123

Human and animal rights The authors did not perform any experi-
ments with animals for conducting this research.
Informed consent There is not any patient data used in this research.
References
1. NIH (2021) Home American cancer society–cancer facts statis-
tics. [Online] Available: https://cancerstatisticscenter.cancer.org/
?_ga=2.191090925.69577. 2 Jan 2837.1589031645-
278983135.1589031645#!/
2. Fass L (2008) Imaging and cancer: a review. Mol Oncol
2:115–152
3. Doi K (2007) Computer-aided diagnosis in medical imaging:
historical review, current status and future potential. Comput Med
Imaging Gr 31:198–211
4. Suarez-Ibarrola R, Basulto-Martinez M, Heinze A, Gratzke C,
Miernik A (2018) Radiomics applications in renal tumor assess-
ment: a comprehensive review of the literature. Cancers (Basel)
12(6):3–17
5. Erak E, Oliveira LD, Mendes AA, Dairo O, Ertunc O, Kulac I,
Baena-Del Valle JA, Jones T, Hicks JL, Glavaris S, Guner G,
Vidal ID, Markowski M, d. l. Calle (2023) Predicting prostate
cancer molecular subtype with deep learning on histopathologic
images. Mod Pathol 36(10):0893–3952
6. Hamm CA, Baumga
¨rtner GL, Biessmann F, Beetz NL, Harten-
stein A, Savic LJ, Frobo
¨se K, Dra
¨ger F, Schallenberg S, Rudolph
M, Baur ADJ, Hamm B, Haas M, Hofbauer S (2023) Interactive
explainable deep learning model informs prostate cancer diag-
nosis at MRI. Radiol Soc North Am 307(4):e222276
7. N. W. T. C. M. L. B. Xu (2015) Empirical evaluation of rectified
activations in convolutional network. arXiv preprint. 1505.00853
8. Pacal I, Karaboga D, Basturk A, Akay B, Nalbantoglu U (2020)
A comprehensive review of deep learning in colon cancer.
Comput Biol Med 126(104003):0010–4825
9. Kim J-E, Nam N-E, Shim J-S, Jung Y-H, Cho B-H, Hwang J
(2020) Transfer learning via deep neural networks for implant
fixture system classification using periapical radiographs. J Clin
Med 9(4):1117
10. Yulian Z, Wang L, Mingxia L, Chunjun Q, Ambereen Y, Aytekin
O, Dinggang S (2017) MRI-based prostate cancer detection with
high-level representation and hierarchical classification. Am
Assoc Phys Med 44(3):1028–1039
11. Baghdadi A, Aldhaam NA, Elsayed AS, Hussein AA, Cavuoto
LA, Kauffman E, Guru KA (2020) Automated differentiation of
benign renal oncocytoma and chromophobe renal cell carcinoma
on computed tomography using deep learning. BJU int
125(4):553–560
12. Tanaka T, Huang Y, Marukawa Y, Tsuboi Y, Masaoka Y, Kojima
K, Iguchi T, Hiraki T, Gobara H, Yanai H, Nasu Y, Kanazawa S
(2020) Differentiation of small (B4 cm) renal masses on mul-
tiphase contrast-enhanced CT by deep learning. Am J Roentgenol
214(3):3–11
13. Leyh-Bannurah S-R, Tian Z, Karakiewicz PI, Wolffgang U,
Sauter G, Fisch M, Pehrke D, Huland H, Graefen M, Buda
¨us L
(2018) Deep learning for natural language processing in urology:
state-of-the-art automated extraction of detailed pathologic
prostate cancer data from narratively written electronic health
records. JCO Clin Cancer Inform 2:1–9
14. Wang X, Guo J, Gu D, Yang Y, Yang X, Zhu K (2019) Tracking
knowledge evolution, hotspots and future directions of emerging
technologies in cancers research: a bibliometrics review. J Cancer
10(12):2643
15. Singhal N, Soni S, Bonthu S, Chattopadhyay N, Samanta P, Joshi
U, Jojera A, Chharchhodawala T, Agarwal A, Desai M, Ganpule
A (2022) A deep learning system for prostate cancer diagnosis
and grading in whole slide images of core needle biopsies. Sci
Rep 12:3383
16. Bulten W, Litjens G, Pinckaers H, Stro
¨m P, Eklund M, Kartasalo
K Demkin M, Dane S (2020) The PANDA challenge: prostate
cANcer graDe assessment using the gleason grading system. In:
23rd international conference on medical image computing and
computer assisted intervention (MICCAI 2020)
17. Lin F, Ma C, Xu J, Lei Y, Li Q, Lan Y, Sun M, Long W, Cui E
(2020) A CT-based deep learning model for predicting the
nuclear grade of clear cell renal cell carcinoma. Eur J Radiol
129:1–15
18. Park JS, Lee HJ, Cho NH, Kim J, Jang WS, Heo JE, Ham WS
(2019) ’’Risk prediction tool for aggressive tumors in clinical T1
stage clear cell renal cell carcinoma using molecular biomarkers.
Comput Struct Biotechnol J 17:371–377
19. Okyaz E, Nurettin E, Axel S, Bernhard B (2018) Diagnostic
classification of cystoscopic images using deep convolutional
neural networks. JCO Clin Cancer Inform 2:1–8
20. Atsushi I, Hirokazu N, Kochi, Hiroyuki N, Hidenori S (2018)
Objective evaluation for the cystoscopic diagnosis of bladder
cancer using artificial intelligence. European Urology Supple-
ments. e1230–e1231
21. Shkolyar E, Jia X, Xing L, Liao J (2019) Automated cystoscopic
detection of bladder cancer using deep-Learning. J Urol. https://
doi.org/10.1097/01.JU.0000557512.76700.42
22. Atsushi I, Hirokazu N, Yuta K, Hiromitsu N, Takahiro K,
Hidenori S, Masahiro M, Hiroyuki N (2020) Cystoscopic imaging
for bladder cancer detection based on stepwise organic transfer
learning with a pretrained convolutional neural network. J
endourol. https://doi.org/10.1089/end.2020.0919
23. Eric W, Lubomir M, Ravi K, Heang-Ping C, Kenny H, Caleb R,
Richard H, Elaine M, Chintana P, Ajjai A, Alon Z (2019) Deep
Learning approach for assessment of bladder cancer treatment
response,’’. Tomography 5(201):8. https://doi.org/10.18383/j.
tom.2018.00036
24. Kenny H, Lubomir M, Richard H, Heang-Ping C, Elaine M,
Matthew S, Ravi K, Alon Z, Ajjai A, Galina K-N, Kimberly S,
Nathaniel M, Daniel B, Sean W, Prasad R, Isaac R (2019)
Diagnostic accuracy of CT for prediction of bladder cancer
treatment response with and without computerized decision
support. Acad Radiol 26(9):1137–1145. https://doi.org/10.1016/j.
acra.2018.10.010
25. Takumi T, Mami H-K, Yumiko O, Satoshi I, Koji M (2019)
Prediction of prostate cancer by deep learning with multilayer
artificial neural network. Can Urol Assoc 13(5):E145–E150.
https://doi.org/10.5489/cuaj.5526
26. Andrew J, Jian C, Saum G, Paul J, Zequn L, Jessica N, Sanjay P,
Inderbir S, Yan L (2019) A deep-learning model using automated
performance metrics and clinical features to predict urinary
continence recovery after robot-assisted radical prostatectomy.
BJU Int 124(3):487–495. https://doi.org/10.1111/bju.14735
27. Junichiro I, Yoh M, Sho U, Yosuke Y, Toshiki K, Soichiro Y,
Minato Y, Kazutaka S, Kazunori K, Noboru N, Tomo K, Kosei
K, Itsuo K, Yasuhisa F (2018) ‘‘Computer-aided diagnosis of
prostate cancer on magnetic resonance imaging using a convo-
lutional neural network algorithm.’’ BJU int 122(3):411–417
28. Ma L, Guo R, Zhang G, Tade F, Schuster DM, Nieh P, Master V,
Fei B (2017) Automatic segmentation of the prostate on CT
images using deep learning and multi-atlas fusion. Med Imaging:
Image Processing 10133:760–768
29. Oberai A, Varghese B, Cen S, Angelini T, Hwang D, Gill I, Aron
M, Lau C, Duddalwar V (2020) ’’Deep learning based
Neural Computing and Applications
123

classification of solid lipid-poor contrast enhancing renal masses
using contrast enhanced CT. The Br J Radiol 93:11–31
30. Cha KH, Hadjiiski L, Chan H-P, Weizer AZ, Alva A, Cohan RH,
Caoili EM, Paramagul C, Samala RK (2017) Bladder cancer
treatment response assessment in CT using radiomics with deep-
learning. Sci Rep 7(1):8738
31. Lucas M, Liem EI, Savci-Heijink CD, Freund JE, Marquering
HA, van Leeuwen TG, de Bruin DM (2019) Toward automated
in vivo bladder tumor stratification using confocal laser.
J Endurol 33(11):930–937
32. Cha KH, Hadjiiski LM, Samala RK, Chan H-P, Cohan RH, Caoili
EM, Paramagul C, Alva A, Weizer AZ (2016) Bladder cancer
segmentation in CT for treatment response assessment: applica-
tion of deep-learning convolution neural network: a pilot study.
Tomography 2(4):421–429
33. Harmon SA, Sanford TH, Brown T, Yang C, Mehralivand S,
Jacob JM, Valera VA, Shih JH, Agarwal PK, Choyke PL,
Turkbey B (2020) Multiresolution application of artificial intel-
ligence in digital pathology for prediction of positive lymph
nodes from primary tumors in bladder cancer. Orig Rep
4:367–382
34. Hadjiiski LM, Cha KH, Cohan RH, Chan H-P, Caoili EM,
Davenport MS, Samala RK, Weizer AZ, Alva A, Kirova-
Nedyalkova G, Shampain K, Meyer N, Barkmeier D, Woolen SA,
Shankar PR (2020) Intraobserver variability in bladder cancer
treatment response assessment with and without computerized
decision support. TomoGraphy 6(2):194–202
35. Ilaria J, Marit L, Judith B, Onno J, Sybren L, Ton G, Henk A,
Jakko A, Daniel M (2020) Automated detection and grading of
non-muscle-invasive urothelial cell carcinoma. Am J Pathol
190(7):1483–1490
36. Wu E, Hadjiiski LM, Samala RK, Chan HP, Cha KH, Richter C,
Cohan RH, Caoili EM, Paramagul C, Alva A, Weizer AZ (2019)
Deep learning approach for assessment of bladder cancer treat-
ment response. TomoGraphy 5(1):201–208
37. Cha KH, Hadjiiski LM, Cohan RH, Chan HP, Caoili EM,
Davenport MS, Samala RK, Weizer AZ, Alva A, Kirova-
Nedyalkova G, Shampain K (2019) Diagnostic accuracy of CT
for prediction of bladder cancer treatment response with and
without computerized decision support. Acad Radiol
26(9):1137–1145
38. Atsushi I, Hirokazu N, Yuta K, Takahiro K, Koji K, Hidenori S,
Masahiro M, Hiroyuki N (2020) Support system of cystoscopic
diagnosis for bladder cancer based on artificial. J Endurol
34(3):352–358
39. Alexander U, Michelle B, Justin G-B, Edward U, Chanon C,
Michael N, Daniel C, Peter D, Roozbeh H (2020) A 3D/2D
Hybrid U-Net CNN approach to prostate organ segmentation of
mpMRI. Am J Roentgenol 216(1):111–116
40. dBJ and WA (2020) ‘‘Deep learning in MIBC’’ Research
Hihglights
41. Eirini A, Kim S, Michael M, Niels R, Thomas H, Christian F,
Norbert W, Peter J, Jan H, Manfred C (2018) Automated gleason
grading of prostate cancer tissue microarrays via deep learning.
Sci Rep 8(1):1–11
42. Ke Y, Xiuying W, Jinman K, Mohamed K, Michael F, Dagan F
(2019) A propagation-DNN: deep combination learning of multi-
level features for MR prostate segmentation. Comput Methods
Progr Biomed 170:11–21
43. Patrick S, Simon K, Jan P, Manuel W, Philipp K, Sebastian B,
Tristan A, Albrecht S, Markus H, Heinz-Peter S, Klaus H-H,
David B (2019) Classification of cancer at prostate MRI: deep
learning versus clinical PI-RADS assessment. Radiology
293(3):607–617
44. Reda I, Shalaby A, Elmogy M, Elfotouh AA, Khalifa F, El-Ghar
MA, Hosseini-Asl E, Gimel’farb G, Werghi N, El-Baz A (2017)
A comprehensive non-invasive framework for diagnosing pros-
tate cancer. Comput Biol Med 81:148–158
45. Rubinstein E, Salhov M, Nidam-Leshem M, White V, Golan S,
Baniel J, Bernstine H, Groshar D, Averbuch A (2019) Unsuper-
vised tumor detection in dynamic PET/CT imaging of the pros-
tate. Med Image Anal 55:27–40
46. Thomas S, Stephanie A, Evrim B, Deepak K, Sena T, Manuel M,
Chris Y, Jonathan S, Sherif M, Pingkun Y, Sheng X, Bradford J,
Maria J, Peter A, Peter L, Baris T (2020) Deep-learning-based
artificial intelligence for PI-RADS classification to assist multi-
parametric prostate MRI interpretation: a development study.
J Magn Reson Imaging 52(5):1499–1507
47. Jun A, Yoichiro Y, Tetsuro S, Yasushi N, Hiromu M, Kotaro T,
Masato Y, Yuki E, Hayato T, Tatsuro H, Masao U, Gen T, Ichiro
M, Manabu F, Akira S, Toyonori T, Go K (2019) illuminating
clues of cancer buried in prostate MR image: deep learning and
expert approaches. Biomolecules 9(11):673
48. Satheesh K, Krishna S, Nicola S, Fernando M, Benhabib H,
Girish K, Matthew D, Ronaldo H (2020) Role of MRI in staging
of penile cancer. J Magn Reson Imaging 51(6):1612–1629
49. Yang G, Wang C, Yang J, Chen Y, Tang L, Shao P, Dillenseger
J-L, Shu H, Luo L (2020) Weakly-supervised convolutional
neural networks of renal tumor segmentation in abdominal CTA
images. BMC Med Imaging 20:20–37
50. Guanyu Y, Chuanxia W, Jian Y, Yang C, Lijun T, Pengfei S,
Jean-Louis D, Huazhong S, Limin L (2020) Weakly-supervised
convolutional neural networks of renal tumor segmentation in
abdominal CTA images. BMC med imaging 20:11–37
51. Prathamesh M, Eric J, Jaya S, Robyn D-C, Bethany R, Megan H,
Larisa J, Harriet M, Pok F, Balazs A, Emanuelle M, Chen Y,
Manas M, Michael R (2020) Deep learning based on standard
H&E images of primary melanoma tumors identifies patients at
risk for visceral recurrence and death. Clin Cancer Res
26(5):1126–1134
52. Lindgren S, May S, Reza K, Olof E, Johannes U, Mads H, Jane S,
Poul F-C, Lars E, Elin T (2019) Deep learning for segmentation
of 49 selected bones in CT scans: first step in automated PET/CT-
based 3D quantification of skeletal metastases. Eur J Radiol
113:89–95
53. Ruud J, Rogier R, Christophe K, Arnoud W, Maudy G, Harrie P,
Georg S, Hessel W, Massimo M (2019) Deep learning for real-
time, automatic, and scanner-adapted prostate (zone) segmenta-
tion of transrectal ultrasound, for example, magnetic resonance
imaging–transrectal ultrasound fusion prostate biopsy. Eur urol
focus 7(1):78–85
54. Arif M, Ivo G, Jose C, Chris H, Gabriel P, Monique J, Wiro N
(2020) Clinically significant prostate cancer detection and seg-
mentation in low-risk patients using a convolutional neural net-
work on multi-parametric MRI. Eur Radiol. https://doi.org/10.
1007/s00330-020-07008-z
55. Jae K, Mun J, Jong S, Jun H, Choung-Soo K, Seong I, Chang W,
Seok-Soo B, Kyo C, Byung H, Yong H, Ji Y (2018) A Deep
belief network and dempster-shafer-based multiclassifier for the
pathology stage of prostate cancer. J Healthc Eng. https://doi.org/
10.1155/2018/4651582
56. Lucas M, Jansen I, Savci-Heijink CD, Meijer SL, de Boer OJ, van
Leeuwen TG, de Bruin DM, Marquering HA (2019) Deep
learning for automatic Gleason pattern classification for grade
group determination of prostate biopsies. Virchows Arch
475:77–83
57. Eirini P, May S, Reza K, Pablo B, Olof E, Johannes U, Mattias O,
Elin T, Mads H, Jane A, Poul F-C, A
˚se A, Lars E (2019) Deep
learning-based quantification of PET/CT prostate gland uptake:
association with overall survival. Wiley
58. Ohad K, Drew L, Ali A, Andreas K, Carleen J, Dragan G,
Thomas S, Boris G (2019) Development of a deep learning
Neural Computing and Applications
123

algorithm for the histopathologic diagnosis and gleason grading
of prostate cancer biopsies: a pilot study. Eur Urol Focus. https://
doi.org/10.1016/j.euf.2019.11.003
59. Yoichiro Y, Toyonori T, Jun A, Masao U, Hiromu M, Yasushi N,
Taishi T, Takuji T, Kotaro T, Ryuto N, Akira S, Ichiro M, Shi-
nichi T, Hiroyuki K, Yukihiro K (2019) Automated acquisition of
explainable knowledge from unannotated histopathology images.
Nat Commun 10(1):1–9
60. Amogh H, Rakesh S, Harri M, Prateek P, Otto E, Pekka T, Hannu
J, Peter J, Ivan J, Anant M (2020) Test-retest repeatability of a
deep learning architecture in detecting and segmenting clinically
significant prostate cancer on apparent diffusion coefficient
(ADC) maps. Eur Radiol. https://doi.org/10.1007/s00330-020-
07065-4
61. Patrick S, Xianfeng W, Jan P, Manuel W, Philipp K, Albrecht S,
Markus H, Heinz-Peter S, Klaus H-H (2020) Simulated clinical
deployment of fully automatic deep learning for clinical prostate
MRI assessment. Eur radiol 31:1–12
62. Stro
¨m P, Kartasalo K, Olsson H, Solorzano L, Delahunt B,
Berney D, Bostwick D, Evans A, Grignon D, Humphrey P, Icz-
kowski K (2020) Artificial intelligence for diagnosis and grading
of prostate cancer in biopsies: a population-based, diagnostic
study. Lancet Oncol 21(2):222–232
63. Zhang H, Ji J, Liu Z, Lu H, Qian C, Wei C, Chen S, Lu W, Wang
C, Xu H, Xu Y (2023) Artificial intelligence for the diagnosis of
clinically significant prostate cancer based on multimodal data: a
multicenter study. BMC Med 21(1):270
64. Bashashati A, Goldenberg S (2022) AI for prostate cancer diag-
nosis: hype or today’s reality? Nature Rev Urol 19(5):261–262
65. Bulten WEA (2022) Artificial intelligence for diagnosis and
Gleason grading of prostate cancer: the PANDA challenge.
Nature Med 28:154–163
66. Rho M, Park J, Moon H, Kim C, Jeon S, Kang M, Lee J (2022)
Dr. Answer AI for prostate cancer: intention to use, expected
effects, performance, and concerns of urologists. Prostate Int
10(1):38–44
67. Rodrigo S-I, Simon H, Gerd R, Christian G, Arkadiusz M (2020)
Current and future applications of machine and deep learning in
urology: a review of the literature on urolithiasis, renal cell car-
cinoma, and bladder and prostate cancer. World J Urol
38(10):2329–2347. https://doi.org/10.1007/s00345-019-03000-5
68. Milap S, Nithesh N, Bhaskar S, H. BM, (2020) Artificial intel-
ligence (AI) in urology-current use and future directions: An
iTRUE study. Turk J Urol 46:27–39. https://doi.org/10.5152/tud.
2020.20117
69. Misgana N, Rodrigo S, Simon H, Arkadiusz M, Alexander R
(2020) Application of artifcial neural networks for automated
analysis of cystoscopic images: a review of the current status and
future prospects. World J Urol 38:2349–2358. https://doi.org/10.
1007/s00345-019-03059-0
70. Checcucci E, Autorino R, Cacciamani GE, Amparore D, De Cillis
S, Piana A, Piazzolla P, Vezzetti E, Fiori C, Veneziano D, Tewari
A (2020) Artificial intelligence and neural networks in urology:
current clinical applications. Minerva Urol Nefrol 72(1):49–57
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the
accepted manuscript version of this article is solely governed by the
terms of such publishing agreement and applicable law.
Neural Computing and Applications
123