Detecting Lung Cancer from Histopathological Images using Convolution Neural Network

Abstract

Lung cancer is one of the leading causes of
mortality in both men and women throughout the world. That
is why early identification and treatment of lung cancer patients
bear a huge significance in the recovery procedure of such
patients. A lot of time, pathologists use histopathological
pictures of tissue biopsy from possibly diseased regions of the
lungs to detect the probability and type of cancer. However, this
procedure is both tedious and sometimes fallible too. Machine
learning based solutions for medical image analysis can help a
lot in this regard. The aim of this work is to provide a
convolution neural network (CNN) model that can accurately
recognize and categorize lung cancer types with superior
accuracy which is very important for treatment. We propose a
CNN model with 15000 images split into 3 categories: Training,
validation, and testing. Three different types of lung tissues
(Benign tissue, Adenocarcinoma, and squamous cell carcinoma)
have been examined. 50 instances from every class were kept for
testing procedure. The rest of the data was split as: about 80%
and 20% for training and validation respectively. Eventually,
our model obtained 98.15% training accuracy and 98.07%
validation accuracy.

Full text

Detecting Lung Cancer from Histopathological
Images using Convolution Neural Network
Dewan Ziaul Karim
Department of Computer Science and Engineering
Brac University
Dhaka, Bangladesh
ziaul.karim@bracu.ac.bd
Tasfia Anika Bushra
Department of Computer Science and Engineering
Daffodil International University
Dhaka, Bangladesh
anika.cse@diu.edu.bd
Abstract— Lung cancer is one of the leading causes of
mortality in both men and women throughout the world. That
is why early identification and treatment of lung cancer patients
bear a huge significance in the recovery procedure of such
patients. A lot of time, pathologists use histopathological
pictures of tissue biopsy from possibly diseased regions of the
lungs to detect the probability and type of cancer. However, this
procedure is both tedious and sometimes fallible too. Machine
learning based solutions for medical image analysis can help a
lot in this regard. The aim of this work is to provide a
convolution neural network (CNN) model that can accurately
recognize and categorize lung cancer types with superior
accuracy which is very important for treatment. We propose a
CNN model with 15000 images split into 3 categories: Training,
validation, and testing. Three different types of lung tissues
(Benign tissue, Adenocarcinoma, and squamous cell carcinoma)
have been examined. 50 instances from every class were kept for
testing procedure. The rest of the data was split as: about 80%
and 20% for training and validation respectively. Eventually,
our model obtained 98.15% training accuracy and 98.07%
validation accuracy.
Keywords—Lung Cancer, Histopathological Images, Deep
Learning, CNN, Classification.
I. INTRODUCTION
Lung cancer is regarded as one of the most prominent
cancers in the whole wide world. It makes up around 25% of
all cancer related deaths [1]. The most common reason behind
lung cancer is smoking. However, in the case of non-smokers,
exposure to radon, second-hand smoking, air pollution, or
certain other substances can all cause lung cancer [2].
Unfortunately, the mortality rate of lung cancer is on the rise
and it is supposed to become about 17 million worldwide in
the year 2030 [3]. In case of developing countries, at the
current growth rate, people's odds of acquiring cancer during
their lifespan may rise up to 50%-60% by 2050 [4].
There are many medical tests (CT scan, X-rays, biopsy,
etc.) done to find out potential cancerous cells. In a biopsy,
histopathology slides are evaluated by pathologists to
establish the potential diagnosis [5,6,7] and determine the
type of lung cancers [8]. But it is a time-consuming procedure
and there is always a chance that cancer types could be
misdiagnosed, which eventually results in incorrect treatment
and puts a toll on patients’ lives.
For the reason mentioned above, it is essential to
implement an automated system for assisting doctors in the
diagnosis of lung cancers as early as possible with high
accuracy. Due to advancements in the technological sector, it
is now possible to build such an automated system using
artificial intelligence (AI) and machine learning (ML).
ML is considered as a branch of AI that concentrates on
using algorithms and data to emulate the way that people
learn and improve accuracy over time [9]. In recent years,
many researchers considered combining different machine
learning techniques with x-rays and CT images to provide a
workable system for identifying types of lung cancer. These
techniques involve Random Forest (RF), Support Vector
Machine (SVM), Bayesian Networks (BN), and Convolution
Neural Network (CNN) for detecting and recognizing lung
cancers. Recently, some authors considered using
histopathological images to differentiate between carcinomas
and non-carcinomas images using CNN.
CNN is an approach under deep learning that is widely
used in image recognition and classification [10,11,12]. It
usually considers an input image, allocates biases and
weights to the images and distinguishes one image from
another. CNN is superior to other conventional approaches in
a sense that it needs a very low amount of preprocessing.
Meaning that in other traditional techniques, filters have to be
set up manually, whereas the neural network obtains the
information itself.
CNN is frequently used for image-related tasks including
classification, segmentation, medical image analysis,
recognition, etc. because it has numerous benefits over other
methods. After providing the input images in CNN, they go
through several convolution layers like flattening, pooling,
and fully-connected (FC) layers. Some types of activation
functions are also used in order to perfectly identify an
image.
The primary aim of our research is to provide a feasible,
efficient and accurate ML model to detect lung cancer from
histopathological images by classifying benign tissue,
adenocarcinoma, and squamous cell carcinomas using CNN
architecture.
II. RELATED WORK
The authors Bijaya Kumar Hatuwal and Himal Chand
Thapa [13] created a deep CNN model to identify benign
tissue, adenocarcinoma, and squamous cell carcinoma where
there were three hidden layers, one input layer, and one fully
connected layer. A dropout value of 0.1 and max-pooling were

used in their research. They used “Adam” optimizer and
eventually got 96.11% accuracy in training and 97.2%
accuracy in validation.
Muayed S AL-Huseiny et al. [14] proposed the approach
of deploying a transfer learning based deep neural network
(DNN) to detect lung nodules that are malignant using CT
images. They performed a fast pre-processing technique to
find out the ROI (Region of Interest) from the images. In this
work, GoogLeNet DNN was used and modified for their
dataset. The code was run in a machine having a processor of
2.5 GHz (Core-i3) with 16 GBs of ram and eventually
achieved an accuracy of 94.38%.
Another paper [15] described a lung cancer detection
system using Alexnet CNN. This work only distinguished
between malignant and benign lung tumors with the help of a
model based on convolution neural network and AlexNet. It
is to be mentioned that AlexNet is made up of 25 layers (with
a scale of 227x227x3). SGDM optimization model and an
initial learning rate of 0.0003 were used. MATLAB 2021a
software was used to run the code and the proposed method
achieved 96% accuracy in the end.
Ying Su et al. [16] proposed an approach for detecting
lung nodules using Faster R-CNN. They experimented on the
LIDC-IDRI dataset [17]. They used 0.001 as learning rate and
70000 as step size. Their attenuation coefficient, dropout rate,
and batch size were 0.1, 0.5, and 64 respectively. The
researchers achieved an accuracy of 91.2% with their
optimized and improved Faster R-CNN method.
Mehedi Masud et al. [18] suggested a classification
framework that differentiates among 5 different types of
colon and lung tissues by analyzing their histopathological
images. Among those 5 classes, 2 are benign and 3 are
malignant. A total of 25000 pictures were included in the
dataset. The authors used DFT and DWT techniques for
feature extraction from images. Later they used a CNN based
technique to identify cancer tissues with an accuracy of
96.33%. Satvik Garg et al. [19] conducted another research
that demonstrated the results of various pre-trained CNN
models.
Another work [20] suggested an automated system for
detecting lung malignancies in WSI (Whole Slide Images) of
lung tissues using two CNN architectures - ResNet and
VGG16. The target was to identify image patches into normal
and tumor cells. The authors used SGD as the optimizer.
Binary crossentropy was assigned as the loss function and a
learning rate of 0.0001 was chosen. Finally, it was observed
that VGG16 (75.41%) outperformed ResNet (72.05%) in
terms of patch level accuracy.
Albert Chon [21] et al. presented a Googlenet-based 3D
CNN model for lung cancer detection. The dataset contained
labeled data for 2101 patients, which the authors divided into
training, validation and test set size of 1261, 420, and 420. A
dropout with 0.3 probability was used after each convolution
and inception layers during training. They used “Adam”
optimizer with 0.0001 learning rate. It was seen that the
suggested model achieved an accuracy of 75.1% with an
AUC score of 0.757.
III. DATASET DESCRIPTION
The dataset used in this study contains 15000 lung
histopathology images. This dataset is obtained from
LC25000 Lung and colon histopathological image dataset
[22]. Those 15000 images are divided into 3 different
categories: benign tissue, adenocarcinoma, and squamous cell
carcinoma. Among those 15000 images, 11850 were put into
training, 3000 were used for validation and 150 were kept for
testing purposes. The pictures were all in RGB format, with
256 X 256 pixel sizes. Some samples from different classes is
shown below:
adenocarcinoma
squamous cell
carcinoma
benign
Fig. 1. Dataset Sample
IV. METHODOLOGY
In this study, a CNN model has been created to detect 3
classes of lung cancers. Fig. 2 indicates the complete
workflow of this research.
Fig. 2. Methodology of Detecting Lung Cancer
The procedure can be divided into 2 main steps: i)
Preparation of dataset ii) Implementing CNN model
A. Dataset Preparation
To avoid getting a disappointing result, it is always better
to pre-process the dataset to increase efficiency [23]. In our
work, various steps were considered to prepare the training
dataset.
• Outliers Removal: The dataset was examined
rigorously for any outliers as outliers can affect the
performance of our model.
• Resizing Images: All the images were scaled to a pixel
size of 256 x 256 as CNN models tend to take a fixed
dimension as inputs.
• Dataset Normalization: Normalized data can help deep
learning based models gain more stability and provide
a better chance of convergence. The range of pixel
values in a picture is 0 to 255. So we used Minmax (1)
normalizer to normalize the pixel values of our
images.
   󰇛 󰇜 (1)

• Data Augmentation: Usually CNN models perform
better with more images. Hence, we applied some data
augmentation methods to expand our training data.
Techniques such as shearing, rotating, shifting,
flipping, etc. were applied to bring variety to the
dataset and make the model more robust.
B. Proposed Model’s Architecture
In this work, we suggest a multi-layered CNN model to
classify different types of lung cancers from histopathological
images. There are 6 convolution layers and 3 dense layers in
our CNN model. There are 32,64,128,128,128, and 64 filters
respectively in those 6 convolution layers with 3 x 3 kernel
size. All the convolution operations are followed by Batch
Normalization [24] operation (2) which helps to make the
learning procedure faster. Following that, a Max-pooling [25]
procedure with a pool size of 2 x 2 was performed.



 (2)
Since convolution networks work better with ReLU [26],
all the convolution layers use “ReLU” as activation function
(3).
  󰇛󰇜 (3)
A flatten layer is designed just after those 6 convolution
layers. It helps in the process of converting data into a one-
dimensional array for usage in the next layer. After this, 3
consecutive dense layers are implemented with 512, 64, and 3
units respectively. There are 3 nodes in the last dense layer as
we are trying to classify 3 different types of lung tissues. A
softmax activation function (4) was applied in the last dense
layer.



 (4)
TABLE I. PROPOSED MODEL SUMMARY
Layers
Shape of Output
conv2d_0
(None,254,254,32)
batch_normalization_0
(None,254,254,32)
max_pooling2d_0
(None,127,127,32)
conv2d_1
(None,125,125,64)
batch_normalization_1
(None,125,125,64)
max_pooling2d_1
(None,62,62,64)
conv2d_2
(None,60,60,128)
batch_normalization_2
(None,60,60,128)
max_pooling2d_2
(None,30,30,128)
conv2d_3
(None,28,28,128)
batch_normalization_3
(None,28,28,128)
max_pooling2d_3
(None,14,14,128)
conv2d_4
(None,12,12,128)
batch_normalization_4
(None,12,12,128)
max_pooling2d_4
(None,6,6,128)
conv2d_5
(None,4,4,64)
batch_normalization_5
(None,4,4,64)
max_pooling2d_5
(None,2,2,64)
flatten_1
(None,256)
dense_0
(None,512)
batch_normalization_6
(None,512)
dense_1
(None,64)
batch_normalization_7
(None,64)
dense_2
(None,3)
activation
(None,3)
Total params: 631,299
Trainable params: 629,059
Non-trainable params: 2,240
C. Parameters used in Training
For our proposed model, we tried to use multiple
parameters e.g., optimizer, learning rate, metrics, batch size,
epoch numbers, callbacks, etc. Table II indicates the various
training parameters used in our model:
TABLE II. TRAINING PARAMETERS USED IN THE MODEL
Name of Parameter
Value
Used Optimizer
Adam
Learning Rate (Initial)
0.01
Learning Rate (Minimum)
.000001
Regularizer
L1 (0.000001)
Batch Size
20
Epochs
60
Steps per Epoch
593
Loss Function
Categorical Crossentropy
Metrics
Accuracy, Precision,
Recall, Loss
Callbacks
ReduceLROnPlateau
D. Evaluation Tools
Python version 3.X was used for the whole experiment
including dataset preparation, model implementation, and
evaluation.

V. RESULT ANALYSIS
From the whole dataset, 50 images from each class were
kept aside for testing purposes. The remaining images were
split in such a way that about 80% data went into training and
20% went into validation. The model finally achieved a
training and validation accuracy of 98.15% and 98.07%
respectively. Fig. 3 and 4 indicate accuracy and loss graphs
for both training and validation respectively.
Fig. 3. Accuracy for Both Training and Validation
Fig. 4. Loss for Both Training and Talidation
Moreover, we also calculated the accuracy of different
pre-trained CNN models for the same dataset along with same
hyperparameters and compared with the result of our proposed
CNN model. The different models that we tried out are
DenseNet201, ResNet152V2, MobileNetV2, InceptionV3,
Xception, InceptionResNetV2, VGG16, VGG19 and
ResNet50. It was seen that all of those models performed
poorer than our proposed model. Among the pre-trained
models, DenseNet201 and MobileNetV2 achieved the highest
training and validation accuracy of 95.41% and 95.03%
respectively. Nevertheless, both of these are lower than the
training and validation accuracy achieved by our proposed
model. As a result, we came to the conclusion that compared
to the different transfer learning approaches, our approach to
lung cancer diagnosis has demonstrated better results with
greater accuracy rates.
Table III shows the comparison between the accuracy rate
of pretrained models and our suggested CNN model against
the same dataset.
TABLE III. TRAINING AND VALIDATION ACCURACY COMPARISON OF
PROPOSED AND PRE-TRAINED CNN MODELS
Model
Training
Accuracy
Validation
Accuracy
Proposed Model
98.15%
98.07%
DenseNet201
95.41%
94.10%
ResNet152V2
94.55%
93.53%
MobileNetV2
94.23%
95.03%
InceptionV3
93.79%
93.20%
Xception
93.72%
92.30%
InceptionResNetV2
93.00%
92.60%
VGG16
91.91%
91.77%
VGG19
90.62%
82.50%
ResNet50
74.68%
51.50%
Fig. 5 and 6 indicate training and validation accuracy
graphs for different models respectively.
Fig. 5. Training Accuracy Comparison of Different Models
Fig. 6. Validation Accuracy Comparison of Different Models

To understand our results better, we noticed the confusion
matrix based on test samples. It is important in the sense that it
provides a clear overview of samples being classified
correctly or incorrectly [27].
Fig. 7. Confusion Matrix on Test Samples
Fig. 7 exhibits the model’s confusion matrix on our
selected test(unseen) samples that include lung
adenocarcinoma (lung_aca), lung squamous cell carcinoma
(lung_scc), and lung benign tissue (lung_n).
If we look at the confusion matrix, we notice that our
model identified all of the samples from lung adenocarcinoma
and lung benign tissues with an accuracy of 100%. However,
for the class squamous cell carcinoma, 48 instances were
classified correctly and 2 were wrongly classified.
Observing the value of Recall (R), Precision (P), and F1
score on test samples is another good idea to check the
reliability of any model [28]. The formula for F1 is = 2*((R *
P) / (R + P)). Precision is calculated using the formula = TP /
(TP + FP). Dividing TP by the addition of TP and FN gives us
Recall. Here, TP is True Positive, FP is False Positive and FN
is False Negative. Table IV illustrates precision, recall and F1
score for every category in our test sample dataset.
TABLE IV. CLASSIFICATION REPORT BASED ON TEST SAMPLES
VI. FUTURE WORK
In the future, different CNN architecture with some
hyperparameters tuning may result in better accuracy than the
current one. This work may be extended to CT scan imaging
problems too. It may also be possible to build a mobile
application that will provide real time detection and eventually
widen the utilization of our technique.
VII. CONCLUSION
This work represents a CNN model to detect lung cancer
using histopathological images. The whole dataset consisted
of 15000 images and our experimental findings indicated
training and validation accuracy of 98.15% and 98.07%
respectively. It is expected that this model will help
pathologists to identify lung cancer (benign, adenocarcinoma,
squamous cell adenocarcinoma lung tissues) with less time,
effort and cost.
REFERENCES
[1] (2020) “American Cancer Society, Lung Cancer
Statistics.[Online]”.Available:
https://www.cancer.org/cancer/lungcancer/about/key-
statistics.html
[2] (2019) “American Cancer Society, Lung Cancer
Causes. [Online].” Available:
https://www.cancer.org/cancer/lungcancer/causes-
risks-prevention/what-causes.html
[3] Nie, L., Zhang, L., Yang, Y., Wang, M., Hong, R. and
Chua, T.S., 2015, October. Beyond doctors: Future
health prediction from multimedia and multimodal
observations. In Proceedings of the 23rd ACM
international conference on Multimedia (pp. 591-600).
[4] Kumar, P., Bhattacharyya, G.S., Dattatreya, S. and
Malhotra, H., 2009. Tackling the cancer Tsunami.
Indian journal of cancer, 46(1), p.1.
[5] Silvestri, G.A., Gould, M.K., Margolis, M.L., Tanoue,
L.T., McCrory, D., Toloza, E. and Detterbeck, F., 2007.
Noninvasive staging of non-small cell lung cancer:
ACCP evidenced-based clinical practice
guidelines. Chest, 132(3), pp.178S-201S.
[6] Travis, W.D., Brambilla, E., Noguchi, M., Nicholson,
A.G., Geisinger, K.R., Yatabe, Y., Beer, D.G., Powell,
C.A., Riely, G.J., Van Schil, P.E. and Garg, K., 2011.
International association for the study of lung
cancer/american thoracic society/european respiratory
society international multidisciplinary classification of
lung adenocarcinoma. Journal of thoracic
oncology, 6(2), pp.244-285.
[7] Collins, L.G., Haines, C., Perkel, R. and Enck, R.E.,
2007. Lung cancer: diagnosis and
management. American family physician, 75(1), pp.56-
63.
[8] Yu, K.H., Zhang, C., Berry, G.J., Altman, R.B., Ré, C.,
Rubin, D.L. and Snyder, M., 2016. Predicting non-
small cell lung cancer prognosis by fully automated
microscopic pathology image features. Nature
communications, 7(1), pp.1-10.
[9] Michie, D., Spiegelhalter, D.J. and Taylor, C.C., 1994.
Machine learning, neural and statistical classification.
[10] O'Shea, K. and Nash, R., 2015. An introduction to
convolutional neural networks. arXiv preprint
arXiv:1511.08458.
[11] Hijazi, S., Kumar, R. and Rowen, C., 2015. Using
convolutional neural networks for image recognition.
Cadence Design Systems Inc.: San Jose, CA, USA,
pp.1-12.
[12] Sultana, F., Sufian, A. and Dutta, P., 2018, November.
Advancements in image classification using
convolutional neural network. In 2018 Fourth
International Conference on Research in Computational
Name of Classes
Precision
Recall
F1
Score
Support
Lung
Adenocarcinoma
.96
1
.98
50
Lung Benign
1
1
1
Lung Squamous
Cell Carcinoma
1
.96
.98
Accuracy
.99
50*3=150
Marco Average
.99
.99
.99
Weighted
Average
.99
.99
.99

Intelligence and Communication Networks (ICRCICN)
(pp. 122-129). IEEE.
[13] Hatuwal, B.K. and Thapa, H.C., 2020. Lung Cancer
Detection Using Convolutional Neural Network on
Histopathological Images. Int. J. Comput. Trends
Technol, 68, pp.21-24.
[14] AL-Huseiny, M.S. and Sajit, A.S., 2021. Transfer
learning with GoogLeNet for detection of lung
cancer. Indonesian Journal of Electrical Engineering
and Computer Science, 22(2), pp.1078-1086.
[15] Agarwal, A., Patni, K. and Rajeswari, D., 2021, July.
Lung Cancer Detection and Classification Based on
Alexnet CNN. In 2021 6th International Conference on
Communication and Electronics Systems (ICCES) (pp.
1390-1397). IEEE.
[16] Su, Y., Li, D. and Chen, X., 2021. Lung nodule
detection based on faster R-CNN
framework. Computer Methods and Programs in
Biomedicine, 200, p.105866.
[17] Armato III, S.G., McLennan, G., Bidaut, L., McNitt‐
Gray, M.F., Meyer, C.R., Reeves, A.P., Zhao, B.,
Aberle, D.R., Henschke, C.I., Hoffman, E.A. and
Kazerooni, E.A., 2011. The lung image database
consortium (LIDC) and image database resource
initiative (IDRI): a completed reference database of
lung nodules on CT scans. Medical physics, 38(2),
pp.915-931.
[18] Masud, M., Sikder, N., Nahid, A.A., Bairagi, A.K. and
AlZain, M.A., 2021. A machine learning approach to
diagnosing lung and colon cancer using a deep learning-
based classification framework. Sensors, 21(3), p.748.
[19] Garg, S. and Garg, S., 2020, December. Prediction of
lung and colon cancer through analysis of
histopathological images by utilizing Pre-trained CNN
models with visualization of class activation and
saliency maps. In 2020 3rd Artificial Intelligence and
Cloud Computing Conference (pp. 38-45).
[20] Šarić, M., Russo, M., Stella, M. and Sikora, M., 2019,
June. CNN-based method for lung cancer detection in
whole slide histopathology images. In 2019 4th
International Conference on Smart and Sustainable
Technologies (SpliTech) (pp. 1-4). IEEE.
[21] Chon, A., Balachandar, N. and Lu, P., 2017. Deep
convolutional neural networks for lung cancer
detection. Standford University.
[22] Borkowski, A.A., Bui, M.M., Thomas, L.B., Wilson,
C.P., DeLand, L.A. and Mastorides, S.M., 2019. Lung
and colon cancer histopathological image dataset
(lc25000). arXiv preprint arXiv:1912.12142.
[23] Pal, K.K. and Sudeep, K.S., 2016, May. Preprocessing
for image classification by convolutional neural
networks. In 2016 IEEE International Conference on
Recent Trends in Electronics, Information &
Communication Technology (RTEICT) (pp. 1778-
1781). IEEE.
[24] Ioffe, S. and Szegedy, C., 2015, June. Batch
normalization: Accelerating deep network training by
reducing internal covariate shift. In International
conference on machine learning (pp. 448-456). PMLR.
[25] Scherer, D., Müller, A. and Behnke, S., 2010,
September. Evaluation of pooling operations in
convolutional architectures for object recognition. In
International conference on artificial neural networks
(pp. 92-101). Springer, Berlin, Heidelberg.
[26] Agarap, A.F., 2018. Deep learning using rectified linear
units (relu). arXiv preprint arXiv:1803.08375.
[27] Ting, K.M., 2017. Confusion matrix. Encyclopedia of
Machine Learning and Data Mining, 260.
[28] Goutte, C. and Gaussier, E., 2005, March. A
probabilistic interpretation of precision, recall and F-
score, with implication for evaluation. In European
conference on information retrieval (pp. 345-359).
Springer, Berlin, Heidelberg.