Content uploaded by Yogesh Bhosale
Author content
All content in this area was uploaded by Yogesh Bhosale on Jul 04, 2022
Content may be subject to copyright.
Content uploaded by Yogesh Bhosale
Author content
All content in this area was uploaded by Yogesh Bhosale on Jul 04, 2022
Content may be subject to copyright.
IoT Deployable Lightweight Deep Learning
Application For COVID-19 Detection With Lung
Diseases Using RaspberryPi
Yogesh H. Bhosale
Dept. of CS&E, Birla Institute of Technology Ranchi, India.
yogeshbhosale988@gmail.com
Dr. K. Sridhar Patnaik
Dept. of CS&E, Birla Institute of Technology Ranchi, India.
kspatnaik@bitmesra.ac.in
Abstract- Proper assessment of COVID-19 patients has
become critical to mitigating and halting the disease's rapid
expansion during the present COVID-19 epidemic across the
nations. Due to the presence of chronic lung/pulmonary
diseases, the intensity and demise rates of COVID-19 patients
were increased. This study will analyze radiography utilizing
chest X-ray images (CXI), one of the most successful testing
methods for COVID-19 case identification. Given that deep
learning (DL) is a useful method and technique for image
processing, there have been several research on COVID-19 case
identification using CXI to train DL models. While few of the
study claims outstanding predictive outcomes, their suggested
models may struggle with overfitting, excessive variance, and
generalization mistakes due to noise, a limited number of
datasets and could not be deployed to IoT devices due to heavy
network size. Considering deep Convolutional Neural Network
(CNN) can conquer the weaknesses by getting predictions with
several diseases using a single model deployed on a real-time IoT
device. We propose a lightweight Deep Learning model (LDC-
Net) that has spearheaded an open-sourced COVID-19 case
identification technique using CNN-generated CXI by utilizing
a suggested strategy aware of distinct features learning of
different classes. Experimental results on Raspberry Pi show
that LDC-Net provides encouraging outputs for detecting
COVID-19 cases with an overall 96.86% precision, 96.78%
recall, 96.77% F1-score, and 99.28% accuracy, better than other
state-of-the-art models. By empowering the Internet of Things-
IoT and IoMT devices, this suggested framework can identify
COVID-19 from CXI and other seven lung diseases with healthy
labels.
Keywords- IoT (Internet of Things) & IoMT, Deep learning,
Lightweight CNN model, Diagnosis, Detection and classification,
COVID-19, Lung / Pulmonary diseases, Raspberry Pi.
I. INTRODUCTION
From the Chinese origin of COVID-19, as of March 14,
2022, there have been 452,201,564 affirmed cases of COVID-
19 and 6,029,852 fatalities, as described by WHO [1]. As of
March 12, 2022, a total of 10,712,423,741 vaccine doses have
been administered. The condition escalated to a worldwide
health disaster at the beginning of 2020, when the WHO
declared a COVID-19 breakout being an epidemic. To combat
this outbreak, proper recognition of COVID-19 individuals
has become necessary, and numerous approaches, such as
protein biomarkers have been at the forefront of detecting
positive patients. Whereas the RT-PCR test has an
accurateness of around 70-80% [2] [3], compelled to offer
correct detection of impacted individuals necessitates the use
of alternative ways such as radiography image analysis,
particularly in medical centers that lack other tests but could
have radiological imaging widely accessible [3] [4], thus
minimizing the shortfall. Healthcare experts are using this
approach to detect the appearance of COVID-19
infections.
With minimal radiography, radiologists have achieved greater
than 95% accuracy using this strategy. [4]. However,
according to recent research [5], the RT-PCR sensitivities of
its COVID-19 findings vary, resulting in false-negative
results.
In addition to RT-PCR, radiographic evaluation is an
excellent diagnostic tool for quick identification of COVID-
19 in which physicians examine and evaluate CXI and CT
images to determine if a suspectable individual has been
affected or not by COVID-19 [6] [7]. Even though CT scans
have a better responsivity to lung disorders, there are
significant limits to their practical application in COVID-19
case identification on a broader scale, particularly mobility,
long-term screening, and the danger of contaminating
healthcare workers. In comparison to CT scans, CXI is
portable, quicker, affordable, and widely accessible, and it
may be conducted in a controlled environment with adequate
accuracy in COVID-19 case detection [8]. Because of these
advantages, several subsequent research [7–10] have shifted
their attention to CXI analysis for COVID-19 disease
identification using machine learning(ML) and DL. With the
rapid propagation of the epidemic, research [34] has advised
that handheld CXI can be used as a reliable tool for COVID-
19 case identification. While CXI is highly rapid, it explicitly
involves specialist radiologists to reach judgment for COVID-
19 disease identification, which would be a time-consuming
operation that demands specialized understanding.
Furthermore, the proportion of radiologists is far lower than
the rate of patients under surveillance. Artificial intelligence
(AI) assisted diagnostic system is therefore required to support
physicians in assessing COVID-19 instances in a relatively
speedy and precise method; else, contaminated individuals
may not be recognized and isolated as soon as feasible and
hence may not obtain appropriate medication [7] [12].
According to the ALA [13] and a Lancet article [14],
chronic lung/pulmonary illness increases the morbidity and
demise rates of COVID-19 confirmed individuals. As a result,
we see this as both a difficulty and a potential for additional
research. Furthermore, the studies [18-26] claims improved
outcomes despite their COVID-19 sample sizes being small
and unbalanced. As a result, this work aims to emphasize the
importance of this requirement and reduce the morbidity and
demise rates associated with chronic lung illnesses with
COVID-19. We used CXI to extend the DL-based automated
lung disease with the COVID-19 detection network (LDC-
Net) for nine-class classification. Furthermore, using COVID-
19 instances, we have studied how to employ deep CNN to
diagnose chronic illnesses on IoT and IoMT - enabled devices.
The proposed application is deployable on lightweight
platforms along with radiography machines. Therefore, the
contribution of this paper is as follows:
IEEE International Conference on IoT and Blockchain Technologies 6th-8th, May, 2022
1
978-1-6654-2416-5/22/$31.00 ©2022 IEEE
2022 International Conference on IoT and Blockchain Technology (ICIBT) | 978-1-6654-2416-5/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICIBT52874.2022.9807725
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on July 04,2022 at 04:54:33 UTC from IEEE Xplore. Restrictions apply.
•A GUI system of a DL model capable of deploying on
IoT devices automatically detects chronic obstructive
lung diseases(COLD) and COVID-19 based on CXI.
•We used five dataset repositories to detect and classify
chronic obstructive lung diseases and COVID-19,
including healthy instances.
•A lightweight LDC-Net deployed on Raspberry Pi
shows promising results. Thus, this work aims to save
time and upgrade diagnostic capabilities for healthcare
staff to minimize the extremity and demise ratios.
The remaining paper is arranged as follows. The literature
study has been discussed in section II. The materials and
methods used in this study are offered in section III.
Experimented findings and conclusion comments are
provided in sections IV and V, respectively.
II. RELATED WORK
Rajpurkar et al. [18] proposed a CNN-based technique for
COVID-19 detection that has been used on CXI because it
generates illnesses comparable to pneumonia but more
extreme. Alqudah et al. [19] suggested an automatic COVID-
19 technique in which they employed ML classifiers-SVM,
KNN with the CNN to identify with an accuracy rate of 98%.
Singh and Gupta [20] used a DL-based system to diagnose
lung tissue cancers from an imaging data set, with an 85.55%
identification ratio. Esteva et al. [21] concentrated on
demonstrating the classifying of epidermal infections,
employing a single CNN layer for detection from various
clinical images of skin. The model was evaluated for binary
classification with two crucial cases-common cancer and fatal
cancer. Furthermore, Liu et al. [22] presented a technique to
characterize TB symptoms in the chest, also detected in
tuberculosis data sets. The unbalanced X-ray imaging data set
was used. The method proved reliable for multiple CNN
models and had a good collection of optimizing strategies,
considering CNN and transfer learning. Unfortunately, it did
not take into account region-level metadata during data
preprocessing.
Butt et al. [23] employed ResNet23 and ResNet18 in their
investigation. This system attained 86.7% accuracy. CT scans
were used in model training. However, CTs are significantly
time intensive and are not offered in many clinics. Fanelli and
Piazza [24] detected patterns and forecasted COVID-19 in
several circumstances using an online public repository
dataset. They anticipated the progression of the COVID-19
epidemic and how the administration's lockdown influenced
the crisis using average kinematics of epidemic transmission.
On a collected data of 4356 CT scans, Li et al. [29] employed
AI with a 3D-DL-model to predict COVID-19 patients.
Chimmula and Zhang [30] created an automatic methodology
using DL and AI, an LSTM network, to estimate COVID-19
fluctuations and potential time stamp dates in distinct nations.
The studies [12] show that the techniques used may give up to
98% accuracy utilizing a pre-trained model and 94.1%
employing a customized CNN. The remaining COVID-19
detection techniques are summarised in Table IV.
III. MATERIALS & METHODS
A.Dataset and preprocessing
In this study, we assembled five datasets from online
public repositories, i.e., NIH[15] and Kaggle[16] [17] [33]
[35]. The collected CXIs of COVID-19 from [17], viral,
bacterial pneumonia from[16], emphysema, fibrosis,
cardiomegaly, healthy, atelectasis from[15], and tuberculosis
from[33] [35] have been used for experimentation, including
10800 CXI out of which 1200 CXIs allocated for every nine
classes. The dataset was partitioned based on a 76:12:12
ratio(%) for train, validation, and test set. The training set
contains 9000 CXI, the validation set includes 900 CXI, and
the test set includes 900 CXI. For normalization, the pixel
values of the input instances were standardized between 0 to
1. The CXR utilized in the datasets under consideration are
grayscale, and the rescale was accomplished by 1./255 to pixel
values. Online augmentation was performed on the training
set with 'imagedatagenerator', which expands the dataset and
provides resilience to the trained network, reducing the
incidence of overfitting issues. Shear-range=0.2, zoom-
range=0.2, and horizontal-flipping are augmentation
approaches utilized to create the network and improve test
image efficiency. All CXIs were rescaled to 1024×1024.
TABLE I. LDC-NET SUMMARY.
___________________________________________________
Layer (type) Output Shape Param
=============================================
conv2d (Conv2D) (None, 1022, 1022, 16) 448
max_pooling2d_0 (None, 511, 511, 16) 0
conv2d_1 (Conv2D) (None, 511, 511, 32) 4640
max_pooling2d_1 (None, 256, 256, 32) 0
conv2d_2 (Conv2D) (None, 256, 256, 64) 18496
max_pooling2d_2 (None, 128, 128, 64) 0
conv2d_3 (Conv2D) (None, 128, 128, 64) 36928
max_pooling2d_3 (None, 64, 64, 64) 0
conv2d_4 (Conv2D) (None, 64, 64, 64) 36928
max_pooling2d_4 (None, 32, 32, 64) 0
conv2d_5 (Conv2D) (None, 32, 32, 64) 36928
max_pooling2d_5 (None, 16, 16, 64) 0
conv2d_6 (Conv2D) (None, 16, 16, 64) 36928
max_pooling2d_6 (None, 8, 8, 64) 0
conv2d_7 (Conv2D) (None, 8, 8, 64) 36928
max_pooling2d_7 (None, 4, 4, 64) 0
flatten (Flatten) (None, 1024) 0
dense (Dense) (None, 512) 524800
dense_1 (Dense) (None, 64) 32832
dense_2 (Dense) (None, 9) 585
=============================================
Total params: 766,246
Trainable params: 766,246
Non-trainable params: 0
B. Proposed methodology
CNN's are multi-layer channels made up of overlapping
convolutional layers(CL) for feature mining and
downsampling layers for feature processing. The structure of
a typical CNN is shown in Fig. 1. CNN can retrieve features
from CXI, making them a popular research area. CNN's are
perceptron-based neural networks [24]. Their own benefit is
obtaining an actual image by avoiding extreme image
preprocessing. CNN's can reduce network costs by integrating
effective use of the image's local and global information via
the local receptive field, weight sharing, and pooling.
Therefore, they are capable of robust interpretation and spin.
Table I depicts the proposed LDC-Net structure, a
straightforward CNN for feature extraction. Generally, CNN
is composed of 3 stacks: convolutional layer (CL), pooling
layer (PL), and fully connected layer (FC). CXI of dimension
1024x1024x3 is inputting for CNN to train it. The range of
channels inside the supplied CXIs is three in this case. The CL
is the first layer. Convolution accepts filters(kernels).
Normally, the kernel's height(h) and width(w) stay constant
as a feature descriptor. This layer collects low-level
characteristics using these kernels. More CL is applied for
feature mining from CXI. Convolution is performed on a sub-
IEEE International Conference on IoT and Blockchain Technologies 6th-8th, May, 2022
2
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on July 04,2022 at 04:54:33 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. Flowchart of the suggested system
ALGORITHM 1. PROPOSED LIGHTWEIGHT FRAMEWORK(LDC-NET)
Input: Ix=X-Ray image(input).
Parameters: m=model, tr=train, v=valid, te=test, a=argmax (minput-layer,
maveragepooling2d, mflatten, mdense, fl=feature-layer, α=learning-rate, es=early-
stopping, b=dataloader_batch, e=epoch), Ac= accuracy, image-shape=
1024x1024, class=Image-label.
Output: Pci [Pclass0,…, Pclass8] Predicted class index
Preprocessing:
if set == training_set:
resize Ix to 1024x1024, crop and resize(Ix) to 1024x1024, fliping, normalize
pixel to (0, 1), pixels /= 255.0
else: resize Ix to 1024x1024, normalize pixels to (0, 1), pixels /= 255.0
Models training:
for x=1 to 50
[tr, v]=partition(tr, v)
for y= (tr/b, v/b)
if Acv < 99:
continue;
if Acv is not improving
for next 5 epochs, then
increase α = α x 0.1:
t(m, c, j)=train[m(c), a , tr(x), v(x)]
v(m, c, j)=valid[(m(c), a, x), v(x)]
unfreeze fl
else:
call es(m);
else:
break;
end y;
end x;
te(m(c), a)=[test(m(c), a, te)]
Pci[Pclass] = te(m(c), a)
-region of the CXI by the kernel. Weights/parameters are the
names given to kernel values. These weights should be
learned using the training. The subregion is known as the
perceptron. The kernel initiates convolution at the
commencement of the CXI. It will be constantly moved over
the entire CXI by a given number of pixels, and compression
is performed till an entire CXI is captured. A solitary process
generates a unique outcome. Convolution across the
CXI generates an array of data. An additional option termed
stride determines the degree to wherein the kernel is moved.
Stride is set to 3 for every CL. Padding is set to the same. In
the first layer 16 filters, the second layer 32 filters, and 64
filters for the remaining filters were used. The result region is
referred to as an activation map[11]. On the convolutional
results, ReLU is used. Following every CL, the max-pooling
layer is utilized to decrease the input's temporal dimensions(h
x w) with a 2x2 kernel. The flattening layer transforms a 2D
activation into a 1D vector that is then fed inside an FC. Deep
features have been retrieved so far. The classifications
problem of lung illnesses is handled by FC using a 1D
continuous vector. The flattening layer comprises 1024
neurons, followed by 512, 64, and 9(params) dense layers.
The final dense receives nine outputs activation from the 64
neurons of the dense layer. Finally, the system utilizes softmax
to detect the classification labels from 0 to 8 classes in the
output layer.
Fig. 1 shows the proposed flow chart of LDC-Net from
input CXI to the last results phase. Before proceeding with
model training, the collected datasets are assembled and
preprocessed. Once the model is trained with specific argmax
as shown in Algorithm 1, the obtained weights/features are
organized in the .h5 file. With network connectivity, the
proposed LDC-Net structure and weights are deployed in IoT-
enabled Raspberry Pi. Clinicians and radiologists can choose
an individual CXI from an X-ray machine database, a hospital
patient X-ray record database, or a patient personal health
record. As a consequence of uploading CXI to the suggested
lung illness detection system, the system will detect probable
lung disease as a final result. The result includes the lung
disease class index as a disease name.
IV. EXPERIMENTS AND RESULTS
A. Experimental setup
Our proposed approach is twofold: training LDC-Net and
deploying on IoT devices. The first fold was implemented on
HPC with the parameters mentioned above in Algorithm 1.
The CNN input CXR was first set to 1024x1024. The training
model was set to 50 epochs based on early-stopping criteria to
avoid overfitting. The standards utilized for early
stopping(ES) and callback are verbose of 1, the patience of 5,
mode to auto, and "restore-best-weights" to true. The actual
learning rate(LR) was set to 0.0001. The epoch size has been
determined internally based on ES parameters. The training
phase of the model terminated automatically based on
callback criteria for ES to avoid overfitting. Model training
automatically halted at 35 epochs with 97.18% accuracy and
95.55% val_acc, as shown in Fig. 2. The trainable params of
LDC-Net are specified in Table I. The adam optimizer was
applied for training, and the LR[11] was controlled
empirically. The batch size of 32 was set. Finally, 9 class
prediction results for LDC-Net from the softmax layer were
achieved, including chronic lung diseases with COVID-19.
Ultimately, DL model experimental performance
accumulated.
Input X-Ray images
Results
Transformation
Integration
Partitioning
Normalization
emphysema
Covid-19
cardiomegaly
fibrosis
viral pneu.
bacterial Pneu.
Deep CNN(LDC-Net) architecture
Disease detection model deployed on Raspberry Pi
CXI Pre-processing
Conv+Relu(1024x1024x16) 1
Pooling(511x511x16) 1
Conv+Relu(8x8x64) 8
Pooling(4x4x64) 8
7 . . . 2
Softmax(9)
Dense(64)
Dense(512)
Flattern(1024)
. . . . .
X-ray Machine Database
Hospital Patient
X-Ray Record
Patient Personal
X-Ray Record
Network
atelectasis
healthy
tuberculosis
IEEE International Conference on IoT and Blockchain Technologies 6th-8th, May, 2022
3
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on July 04,2022 at 04:54:33 UTC from IEEE Xplore. Restrictions apply.
Table II. CLASS-WISE PERFORMANCE ANALYSIS OF LDC-NET.
Metrics
Atelectasis
Bacterial
pneu.
Cardio-
megaly
Covid-19
Emphy-
sema
Fibrosis
Healthy
Tuberc-
ulosis
Viral pneu.
Accuracy(%)
98.11
99.88
99.11
99.66
98.77
98.55
99.88
99.77
99.77
Precision-PPV(%)
92.78
100
93.40
98.99
90.83
96.77
100
99
100
Recall-TPR (%)
90
99
99
98
99
90
99
99
98
Specificity-TNR(%)
99.12
100
99.12
99.87
98.75
99.62
100
99.87
100
F1-score(%)
91.37
99.50
96.12
98.49
94.74
93.26
99.50
99
98.99
AUC(%)
94.56
99.5
99.06
98.93
98.87
94.81
99.5
99.43
99
False Negative Rate
0.1
0.01
0.01
0.02
0.01
0.1
0.01
0.01
0.02
Type 1 error (FP)
7
0
7
1
10
3
0
1
0
Type 2 error (FN)
10
1
1
2
1
10
1
1
2
MCC
0.9032
0.9943
0.9566
0.9830
0.9415
0.9253
0.9943
0.9887
0.9887
Confusion entropy
0.1214
0.0095
0.0605
0.0287
0.0861
0.09481
0.0095
0.0191
0.0192
True Negative(TN)
793
800
793
799
790
797
800
799
800
True Positive(TP)
90
99
99
98
99
90
99
99
98
TABLE III. OBTAINED MACRO AVERAGE(OVERALL) RESULTS.
Macro avg.
Precision (%)
Macro avg.
Recall(%)
Macro avg.
F1-score(%)
Macro avg.
Accuracy( %)
Zero-one
Loss
AUC-ROC
-Score(%)
Overall Error
Rate(%)
Test Time
95% CI
96.86
96.78
96.77
99.28
29
98.18
4.83
136.43 ms
(0.95624, 0.97931)
Fig. 2. Accuracy and loss were obtained during training, validation phase.
B. Performance Metrics and Results
We utilized a confusion matrix to evaluate performance. The
confusion matrix, depicted in Fig. 3, measures the model
efficacy. It compares real (X-axis) and projected (Y-axis)
values: TP(true positive), FP(false positive), TN(true
negative), and FN(false negative). These measurements
include accuracy, precision, recall, specificity, F1-score,
error-rate, MCC, zero-one loss, and confusion entropy. We
required the counts of the essential elements to evaluate these
metrics: TP, FP, TN, and FN.
The evaluation criteria listed in Tables II and III are the
most frequently utilized to assess the performance of
classification systems. From Table II the highest 99.88%
accuracy obtained by bacterial pneumonia and healthy class
class followed by 98.11%, 99.11%, 99.66%, 98.77%, 98.55%,
99.77%, 99.77% by atelectasis, cardiomegaly, Covid-19,
emphysema, fibrosis, tuberculosis, viral pneumonia
respectively. Attained precision, recall, specificity, F1-score,
and AUC of 90.83%, 99%, 98.75%, 94.74%, 98.87% by
emphysema; 96.77%, 90%, 99.62%, 93.26%, 94.81% by
fibrosis; 98.99%, 98%, 99.87%, 98.49%, 98.93% by Covid-
19 class; 100%, 99%, 100%, 99.50%, 99.5% by bacterial
class; 100%, 98%, 100%, 98.99%, 99% by viral class; 100%,
99%, 100%, 99.50%, 99.5% by healthy; 92.78%, 90%,
99.12%, 91.37%, 94.56% by atelectasis; 93.40%, 99%,
99.12%, 96.12%, 99.06% by cardiomegaly; 99%, 99%,
99.87%, 99%, 99.43 by tuberculosis class respectively.
However, the lowest 0.01 FNR attained by tuberculosis,
healthy, emphysema, cardiomegaly, bacterial pneumonia
classes. 0.9943 MCC, 0.0095 confusion entropy(CEN), 800
TN attained by bacterial and healthy class.
Table III shows the overall(average) results obtained by
LDC-Net, including precision, recall, F1-score, accuracy,
zero-one loss (out of 900 test samples), AUC-ROC-score,
error rate, testing time taken for each image, and confidence
interval. LDC-Net achieved the average macro performance
of 99.28% accuracy, 98.18% auc-roc score, 96.86% precision,
96.78% recall, 96.77% F1-score, lowest 29 zero-one loss,
4.83% overall error-rate, (0.9562, 0.9793)confidence interval
(95% CI), and 4.3 seconds to test individual CXI on Raspberry
and 0.136 seconds on desktop. Hence, based on the obtained
superior results, we recommend that the LDC-Net with GUI
is expected to be deployed on all X-ray machines to detect
lung diseases with COVID-19 instances to assist radiologists.
Fig. 3. Confusion matrix acquired for LDC-Net at the testing phase.
[Confusion matrix x and y-axis labels from 0 to 8(0:atelectasis, 1:bacterial-
pneumonia, 2:cardiomegaly, 3:Covid19, 4:emphysema, 5:fibrosis 6:healthy,
7:tuberculosis, 8:viral-pneumonia)].
Fig. 3 shows the AUC-ROC curve of LDC-Net. ROC is a
2D chart that equates to a true-positive rate(TPR) as opposed
to the false-positive rate (FPR). The ROC curve represents -
IEEE International Conference on IoT and Blockchain Technologies 6th-8th, May, 2022
4
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on July 04,2022 at 04:54:33 UTC from IEEE Xplore. Restrictions apply.
TABLE IV. COMPARATIVE ANALYSIS OF PREVIOUSLY DEVELOPED COVID-19 CLASSIFICATION SYSTEMS.
Author
Model
Class with Test Sample Size
Performance Metrics
Test Time
Alqudah et al.[19]
AOCT-NET
COVID19 -, Normal -
Accuracy:95.2%, Specificity:100%, Sensitivity: 93.3%
6.3 s
Butt et al. [22]
ResNet
COVID19(30), Influenza-A viral
pneumonia(30), Healthy(30)
Specificity:92.2%, Recall:86.66%, Sensitivity of 98.2%,
Precision:86.86%, F1-score:86.7%
-
Li et al.[25]
COVNet
COVID19(68), CAP(28), Healthy(51)
AUC:96.33%, Specificity:94.66%
-
Ismael and
Şengür[28]
ResNet50+
SVM
COVID19(40), Healthy(47)
Accuracy:95.79%, Sensitivity:94.00%,
Specificity:97.78%, F1-score95.92%, AUC:99.87%
48.9 s
Hemdan et al.[29]
VGG19
COVID19(25), Normal(25)
Accuracy:90%, Precision:83%
90.0 s
Narin et al.[30]
ResNet50
COVID19(50), Normal(50)
Accuracy:95.38%
-
Sethy and Behera
[31]
ResNet50+
SVM
COVID19(25), Normal(25)
Accuracy:95.38%, FPR:95.52%, F1 -score:91.41%,
MCC/Kappa-measure:90.76%
-
Proposed Model
Lightweight
LDC-Net
Fibrosis, BactPneu., Covid-19,
Atelectasis, Healthy, Cardiomegaly,
Emphysema, Tuberculosis, ViralPneu.
Accuracy:99.28%, Precision:96.86%, Recall:96.78%,
F1-score:96.77%, Zero-one Loss:29(Out of 900 CXI),
AUC-ROC-score:98.18%, Error rate:4.83%
0.136 s to
4.3 s
the commutations across sensitivity and specificity. The ROC
curve has been shown using TPR on the y-axis and FPR on
the x-axis. Superior Area Under Curve (AUC) values are
essential in healthcare diagnostics. Therefore, its
computations in healthcare analysis help the data analyst ease
the analysis with forecasting on clinical research. From Fig. 4,
the highest ROC-AUC-score of 0.99% was attained by
bacterial, viral pneumonia, tuberculosis, Covid-19,
emphysema, healthy, tuberculosis, followed by 0.95 for
fibrosis and atelectasis class, respectively. Fig. 5 shows the
GUI application of the proposed framework deployed on IoT-
enabled RaspberryPi 3 for COVID-19 detection with other
lung diseases. Where clinicians and radiologists can select the
individual CXI from the X-ray machine database, hospital
patient X-Ray record database, or patient personal health
record. After uploading the CXI to the proposed lung disease
detection system, it will detect the possible disease label as a
Fig. 4. ROC curve obtained at the testing phase.
Fig. 5. Raspberry Pi deployed a GUI application recommended framework
(LDC-Net) for COVID-19 with other lung disease detection.
- a final result.
C. Experimental environment
The LDC-Net was trained on High-Performance
Computing (HPC). The hardware used for the
experimentation was an HPC Intel(R) Xeon(R) CPU E5-2630
v3 @ 2.40GHz. CentOS 6 operating system, Python 3.7,
Tensorflow 2.3.1, keras, theano, matplotlib, numpy, pandas,
and Tkinter are used for LDC-Net experimentation and finally
obtained lightweight CNN network structure and features
deployed on Raspberry Pi 3 with 1 GB Ram and Buster OS.
D. Results Comparison
The comparative performance analysis of the LDC-Net
with other existing classification techniques has been shown
in Table IV. Existing Covid-19 classification approaches
tried to implement binary[19] [28] [29] [30] [31] and three-
class[22] [25] classification. The highest accuracy of 95.79%
was attained by Ismael and Şengür [28], but their used CXI
samples were very tiny. In comparison, LDC-Net achieves
99.28% overall accuracy. However, it has been observed that
because of the low computational power of the Raspberry Pi
device, the proposed LDC-Net took 4.3 seconds. In contrast,
when it was tested on a desktop with 8 GB of RAM, it took
0.136 seconds to test individual CXI.
E. Limitations
However, the proposed LDC-Net structure and weights
have been deployed in IoT-enabled Raspberry Pi with network
connectivity. The implemented construction is assumed as an
ideal scenario, but network connectivity can incorporate
delay, loss of data, other security aspects, etc., as a limitation
of the study.
V. CONCLUSION
Significant advancements and the accessibility of
intelligent devices necessitate the emergence of newer
solutions to fulfill the demands of growing nations. This work
has developed an IoT-enabled COVID-19, chronic obstructive
lung diseases, and healthy cases detection system to reduce
mortality. This will assist in lessening the effort of the health
professionals while also offering an additional level of
prevention towards the intensity of COVID-19. The suggested
model employs a real-time DL system based on a Raspberry
Pi to diagnose illnesses. When detecting all nine labels, the
deployed LDC-Net on the Raspberry device operates
impressively even in low computation; the tested LDC-Net
achieved a macro average (overall) accuracy of 99.28%. The
IEEE International Conference on IoT and Blockchain Technologies 6th-8th, May, 2022
5
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on July 04,2022 at 04:54:33 UTC from IEEE Xplore. Restrictions apply.
test results show a high degree of a confidence interval, i.e.
(0.9562, 0.9793). Additionally, there are several strategies for
improving performance and achieving better outcomes.
Future research will entail increasing the deployments of IoT
devices with radiography machines, combining multiple
capabilities, improving performance, and creating a web and
mobile application with a user-friendly interface for
monitoring. Consequently, authorities will be able to take
rapid action under pandemic safety guidelines.
More COVID-19 CXRs will be accumulated in the future,
and more models for COVID-19 classification will be
examined. In addition, 14 lung illnesses from[15],
bronchiectasis, severe asthma, extrinsic allergic alveolitis,
etc., will be studied in the future. The aspect will be
considered in upcoming research, such as creating and
deploying a web and smartphone application to assist
radiologists in detecting COVID-19 and pulmonary disorders
from a remote location. We hope that future studies will
identify various COVID-19 variants like Beta, Delta,
Omicron, and IHU[32] from radiography images.
REFERENCES
[1] WHO, WHO Coronavirus (COVID-19) Dashboard, Accessed March
15, 2022, https://covid19.who.int/.
[2] WHO, WHO Director-general's opening remarks at the media briefing
on COVID-19-27 2020, 2020, Accessed January 20, 2022,
https://www.who.int/dg/speeches/detail/who-director-general-s-
opening-remarks-at-the-media-briefing-on-covid-19-27-july-2020
[3] Rubin R, Abbasi J, Voelker R. Latin America, and its global partners
Toil to procure medical Supplies as COVID-19 Pushes the Region to
its limit. J Am Med Assoc 2020; 324(3):217–9.
https://doi.org/10.1001/jama.2020.11182.
[4] Dangis, A., Gieraerts, C., de Bruecker, Y., Janssen, L., Valgaeren, H.,
Obbels, D.,Gillis, M., Ranst, M. van, Frans, J., Demeyere, A., &
Symons, R. (n.d.). Accuracy and reproducibility of low-dose
submillisievert chest CT for the diagnosis of COVID-19.
[5] A. Tahamtan and A. Ardebili, "Real-time RT-PCR in COVID-19
detection: Issues affecting the results," Expert Rev. Mol. Diagn., vol.
10, no. 5, pp. 453–454, 2020
[6] M. Abdel-Basset, V. Chang, H. Hawash, R. K. Chakrabortty, and M.
Ryan, "FSS-2019-nCov: A deep learning architecture for semi-
supervised fewshot segmentation of covid-19 infection," Knowl.-
Based Syst. vol. 212, 2021, Art. no. 106647.
[7] L. Wang and A. Wong, "COVID-Net: A tailored deep convolutional
neural network design for detection of COVID-19 cases from chest x-
ray images," 2020, arXiv:2003.09871.
[8] S. Rajaraman, J. Siegelman, P. O. Alderson, L. S. Folio, L. R. Folio,
and S. K. Antani, "Iteratively pruned deep learning ensembles for
COVID-19 detection in chest x-rays," 2020, arXiv:2004.08379.
[9] A. Borghesi and R. Maroldi, "COVID-19 outbreak in italy:
Experimental chest x-ray scoring system for quantifying and
monitoring disease progression," La Radiologia Med., vol. 125, no. 5,
pp. 509–513, 2020
[10] A. Jacobi, M. Chung, A. Bernheim, and C. Eber, "Portable chest x-ray
in coronavirus disease-19 (COVID-19): A pictorial review," Clin.
Imag., vol. 64, pp. 35–42, 2020.
[11] Y. H. Bhosale. "Digitization Of Households With Population Using
Cluster And List Sampling Frame In Aerial Images," www.oaijse.com,
vol 5., issue 2, pp. 22–26, Feb 2020.
[12] H. S. Maghdid, A. T. Asaad, K. Z. Ghafoor, A. S. Sadiq, and M. K.
Khan, "Diagnosing COVID-19 pneumonia from x-ray and CT images
using deep learning and transfer learning algorithms," 2020,
arXiv:2004.00038
[13] American Lung Association. Accessed On September 12, 2021.
[Online]. Available: https://www.lung.org/lung-health-diseases/lung-
disease-lookup/covid-19/chronic-lung-diseases-and-covid.
[14] Paul Aveyard, Min Gao, Nicola Lindson, Jamie Hartmann-Boyce,
Peter Watkinson, Duncan Young et al. Association between pre-
existing respiratory disease and its treatment, and severe COVID-19: a
population cohort study. https://doi.org/10.1016/S2213-
2600(21)00095-3.
[15] Summers, Ronald (Lung Disease X-Ray Dataset). CXR8 - National
Institutes of Health - Clinical Center. Accessed On September 12 2021.
Online Available https://nihcc.app.box.com/v/ChestXray-
NIHCC/folder/37178474737.
[16] Paul Mooney (Chest X-Ray Pneumonia Dataset). Accessed On
September 12 2021. [Online]. Available:
https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia.
[17] SIIM-FISABIO-RSNA COVID-19 Detection Challenge. Accessed On
September 12 2021. [Online]. Available:
https://www.kaggle.com/c/siim-covid19-detection.
[18] Rajpurkar P., Irvin J., Zhu K., Yang B., Mehta H., Duan T., et al.
Chexnet: radiologist-level pneumonia detection on chest x-rays with
deep learning. arXiv preprint arXiv:1711052252017.
[19] Alqudah A., Qazan S., Alqudah A.. Automated systems for detection
of covid-19 using chest x-ray images and lightweight convolutional
neural networks2020; 10.21203/rs.3.rs-26500/v1.
[20] Singh GAP, Gupta P. Performance analysis of various machine
learning-based approaches for detection and classification of lung
cancer in humans. Neural Comput Appl 2019;31(10):6863–77.
[21] Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al.
Dermatologist-level classification of skin cancer with deep neural
networks. Nature 2017;542(7639):115–18 5]
[22] Liu C, Cao Y, Alcantara M, Liu B, Brunette M, Peinado J, et al. Tx-
cnn: detecting tuberculosis in chest x-ray images using convolutional
neural network. In: 2017 IEEE International Conference on Image
Processing (ICIP). IEEE; 2017. p. 2314–18.
[23] Butt C, Gill J, Chun D, Babu BA. Deep learning system to screen
coronavirus disease 2019 pneumonia. Appl Intell 2020:1.
[24] Fanelli D, Piazza F. Analysis and forecast of covid-19 spreading in
China, Italy and France. Chaos Solitons Fractals 2020;134:109761.
[25] Li L, Qin L, Xu Z, Yin Y, Wang X, Kong B, et al. Artificial intelligence
distinguishes covid-19 from community acquired pneumonia on chest
ct. Radiology 2020:200905.
[26] Chimmula VKR, Zhang L. Time series forecasting of covid-19
transmission in canada using lstm networks. Chaos Solitons Fractals
2020:109864. doi:10.1016/ j.chaos.2020.109864.
[27] Y. H. Bhosale. "Digitization Of Households with Population Using
Cluster and List Sampling Frame In Aerial Images," www.oaijse.com,
vol 5., issue 2, pp. 22–26, Feb 2020.
[28] Aras M. Ismael, Abdulkadir Şengür, Deep learning approaches for
COVID-19 detection based on chest X-ray images, Expert Systems
with Applications, Volume 164, 2021, 114054, ISSN 0957-4174,
https://doi.org/10.1016/j.eswa.2020.114054.
[29] E.E.-D. Hemdan, M.A. Shouman, M.E. Karar, COVIDX-Net: A
framework of deep learning classifiers to diagnose covid-19 in X-ray
images, 2020, arXiv preprint arXiv:2003.11055.
[30] A. Narin, C. Kaya, Z. Pamuk, Automatic detection of coronavirus
disease (COVID19) using X-ray images and deep convolutional neural
networks, 2020, arXivpreprint arXiv:2003.10849.
[31] P.K. Sethy, S.K. Behera, Detection of coronavirus disease (COVID-
19) based on deep features, 2020,
http://dx.doi.org/10.20944/preprints202003.0300.v1, Preprints.
[32] Covid-19: Omicron less severe than Delta variant, UK studies find.
Accessed On September 12, 2021. [Online]. Available:
https://www.business-standard.com/article/current-affairs/covid-19-
omicron-less-severe-than-delta-variant-uk-studies-find-
121122300641_1.html
[33] Dataset tuberculosis. Accessed: January 26, 2022, Online. Available:
http://www.kaggle.com/kmader/pulmonary-chest-xray-abnormalities
[34] A. Nair et al., "A British Society of Thoracic Imaging statement:
Considerations in designing local imaging diagnostic algorithms for the
COVID-19 pandemic," Clin. Radiol., vol. 75, no. 5, pp. 329–334, 2020
[35] Tuberculosis (TB) Chest X-ray Database. Accessed On January 27
2022, Online. Available: https://www.kaggle.com/tawsifurrahman/
tuberculosis-tb-chest-xray-dataset.
IEEE International Conference on IoT and Blockchain Technologies 6th-8th, May, 2022
6
Authorized licensed use limited to: Birla Institute of Technology. Downloaded on July 04,2022 at 04:54:33 UTC from IEEE Xplore. Restrictions apply.