Interdisciplinary Sciences: Computational Life Sciences

[Link]

ORIGINAL RESEARCH ARTICLE

Utilization of Deep Convolutional Neural Networks for Accurate Chest

X‑Ray Diagnosis and Disease Detection
Mukesh Mann1 · Rakesh P. Badoni2 · Harsh Soni3 · Mohammed Al‑Shehri4 · Aman Chandra Kaushik5,6 ·
Dong‑Qing Wei5

Received: 29 September 2022 / Revised: 6 March 2023 / Accepted: 6 March 2023

© International Association of Scientists in the Interdisciplinary Areas 2023

Abstract
Chest radiography is a widely used diagnostic imaging procedure in medical practice, which involves prompt reporting of
future imaging tests and diagnosis of diseases in the images. In this study, a critical phase in the radiology workflow is auto‑
mated using the three convolutional neural network (CNN) models, viz. DenseNet121, ResNet50, and EfficientNetB1 for
fast and accurate detection of 14 class labels of thoracic pathology diseases based on chest radiography. These models were
evaluated on an AUC score for normal versus abnormal chest radiographs using 112120 chest X–ray14 datasets containing
various class labels of thoracic pathology diseases to predict the probability of individual diseases and warn clinicians of
potential suspicious findings. With DenseNet121, the AUROC scores for hernia and emphysema were predicted as 0.9450
and 0.9120, respectively. Compared to the score values obtained for each class on the dataset, the DenseNet121 outperformed
the other two models. This article also aims to develop an automated server to capture fourteen thoracic pathology disease
results using a tensor processing unit (TPU). The results of this study demonstrate that our dataset can be used to train models
with high diagnostic accuracy for predicting the likelihood of 14 different diseases in abnormal chest radiographs, enabling
accurate and efficient discrimination between different types of chest radiographs. This has the potential to bring benefits to
various stakeholders and improve patient care.

2
Department of Mathematics, École Centrale School
* Mukesh Mann
of Engineering, Mahindra University, Hyderabad 500043,
[Link]@[Link]
India
* Rakesh P. Badoni 3
Department of Information Technology, Indian Institute
rakeshbadoni@[Link]
of Information Technology, Sonepat, Haryana 131029, India
* Dong‑Qing Wei 4
Department of Biology, Faculty of Science, King Khalid
dqwei@[Link]
University, Abha, Saudi Arabia
Harsh Soni 5
State Key Laboratory of Microbial Metabolism, and School
harsh.11912082@[Link]
of Life Sciences and Biotechnology, Shanghai Jiao Tong
Mohammed Al‑Shehri University, 200030 Shanghai, China
mshehri@[Link] 6
School of Biomedical Informatics, University of Texas
Aman Chandra Kaushik Health Science Centre at Houston, Houston, TX, USA
amanbioinfo@[Link]
1
Department of Computer Science and Engineering,
Indian Institute of Information Technology, Sonepat,
Haryana 131029, India

13
Vol.:(0123456789)
 Interdisciplinary Sciences: Computational Life Sciences

Graphical Abstract

Chest X-ray Diagnosis

General representation of a convolutional neural network Sample chest X-ray image for all classes in chest X-ray14

Integration Data Analysis

Keywords Chest X-ray diagnosis · Chest X-ray14 · Deep ConvNet diagnosis · TPU training

1 Introduction representation of a convolutional neural network is pro‑

vided in Fig. 1. The availability of large datasets and the
Convolutional Neural Networks, also known as CNNs or progress in deep learning have made it possible for models
ConvNets, have become a widely used method in the field to achieve human-level performance in many fields. Medi‑
of computer vision. Following the success of AlexNet [1] cal image analysis, such as the detection and segmentation
in ILSVRC 2012, CNNs have been extensively applied to of radiology images, has also yielded promising results
various Cardiovascular (CV) tasks using deep learning. using deep learning.
As time has passed, the architecture of CNNs has become The detection of many diseases can be achieved through
increasingly complex and deep, leading to the develop‑ the use of chest radiology, which is the best and most com‑
ment of deep ConvNets. With improvements in network monly used clinical imaging tool. Every year, more than two
architecture and computer hardware, we have been able billion chest radiology imaging tests are conducted [2]. This
to train these deep ConvNets, which have shown sig‑ imaging tool is crucial for identifying various thoracic dis‑
nificant performance improvements in object detection, eases, which are the leading cause of mortality worldwide. It
recognition, and image segmentation tasks. A general would be exponentially beneficial if computer systems could

13
Interdisciplinary Sciences: Computational Life Sciences

Fig. 1  General representation of

a convolutional neural network

interpret chest X-rays with the same efficiency as a prac‑ approaches used for COVID-19 diagnosis at both the image
ticing radiologist [3, 4]. Over the last few years [5–9], the and region levels. Khan et al. [33] introduced a new method
diagnosis of chest radiology images has received increased called deep hybrid learning (DHL) and deep boosted hybrid
attention, and several algorithms have been developed for learning (DBHL) for accurately detecting COVID-19 in
pulmonary tuberculosis classification [10–13] and pneumo‑ chest X-ray images. The DBHL technique involves using
nia detection [3, 14, 15]. During the pandemic, deep learning data augmentation, transfer learning-based fine-tuning,
has also found its usage in COVID-19 detection [16–21]. deep features boosting, and hybrid learning to improve
Our proposed study aims to identify multiple pathology the performance of the COVID-RENets models (COVID-
diseases by re-implementing the CheXNet model and con‑ RENets-1 and COVID-RENets-2). In their experiments,
structing several additional models with nearly identical the DBHL framework outperformed other well-established
hyperparameters. These models will be compared side-by- CNN models.
side. Additionally, we aim to train these models on a Ten‑ Stirenko et al. [11] conducted a study on the use of deep
sor Processing Unit (TPU) to reduce training time. Deep learning-based computer-aided diagnosis (CADx) to predict
learning has made significant advancements in the field of the presence of tuberculosis by statistically analyzing 2D
medicine due to the availability of vast datasets, enabling chest X-ray images. They demonstrated the effectiveness of
the development of models that surpass the performance of deep CNN for CADx of tuberculosis, particularly through
medical professionals. For instance, pneumonia detection [3, techniques like lung segmentation and data augmentation,
14, 15], skin cancer classification [22–24], and lung cancer both lossless and lossy, on a small and unbalanced dataset.
screening [25–27] have all benefited from deep learning. Rahman et al. [14] presented a study that aimed to develop
CheXNet [3], an algorithm that can detect pneumonia from an automated method for identifying bacterial and viral
chest X-rays, performs better than practicing radiologists. pneumonia by analyzing digital X-ray images. They gave
CNN models, such as CheXNeXt [4], can identify various a comprehensive overview of current approaches to detect‑
pathologies diseases with a performance similar to that of ing pneumonia and described the specific techniques used
practicing board-certified radiologists using frontal-view in their research. Alakus and Turkoglu [17] created clinical
chest X-rays. predictive models using deep learning and laboratory data
In recent years, various life-threatening diseases have to forecast which patients were likely to contract COVID-
been detected and diagnosed using deep learning tech‑ 19. They evaluated the models’ effectiveness using various
niques by a number of researchers [28–32]. Baltruschat et al. performance metrics like precision, F1-score, recall, AUC,
[28] conducted a study comparing multiple deep-learning and accuracy with data from 600 patients and 18 labora‑
approaches to classify chest X-Ray images with multiple tory findings and validated them through ten-fold cross-
labels. They analyzed various methods for using CNNs to validation and train-test split methods. The results showed
classify X-ray images from the Chest X-ray14 dataset and that the predictive models accurately identified patients with
found that fine-tuned networks using the ImageNet dataset COVID-19. Dey et al. [15] designed a Deep-Learning Sys‑
produced satisfactory results. However, the most effective tem (DLS) for diagnosing lung abnormalities by using chest
model was specifically trained using only X-ray images and X-ray images. They tested the system with conventional and
incorporated non-image data. A systematic survey of deep filtered chest radiographs and conducted an initial evaluation
learning techniques for the analysis of COVID-19 and their using a SoftMax classifier. The outcomes indicated that the
usability for detecting Omicron has been provided by [32]. VGG19 method provided higher classification accuracy than
The COVID-19 pandemic has caused a shift towards uti‑ other methods.
lizing deep learning methods for analyzing and identifying Khan et al. [34] have proposed a new diagnostic system
infected areas in radiology images. These techniques can that employs deep CNNs to detect and analyze COVID-19
be divided into classification, segmentation, and multi-stage infections by identifying minor irregularities. This system

13
 Interdisciplinary Sciences: Computational Life Sciences

Atelectasis Cardiomegaly Consolidation Edema Effusion Emphysema Fibrosis

Hernia Infiltration Mass Nodule Pleural Thickening Pneumonia Pneumothorax

Fig. 2  Sample chest X-ray image for all classes in the chest X-ray14

comprises two phases. In the first phase, a new CNN named First, many studies have used small and potentially biased
SB-STM-BRNet is utilized to identify COVID-19 infections datasets, which negatively impacts the generalizability and
in lung CT images. This is achieved using a Squeezed and accuracy of these models. Therefore, the need for more
Boosted (SB) channel and a Split-Transform-Merge (STM) extensive and diverse datasets is essential. Second, there
block with dilated convolutions. In the second phase, the has been a lack of research on the ability of deep learning
COVID-CB-RESeg CNN is employed to detect and analyze models to diagnose rare or less common diseases in chest
COVID-19-infected areas in the images. This CNN incor‑ X-rays accurately. Finally, previous studies have relied on
porates region-homogeneity and heterogeneity operations simple accuracy metrics, which is inadequate for evaluating
in each encoder-decoder block and auxiliary channels in the performance of these models. More robust evaluation
the boosted-decoder to learn about low illumination and the methods, such as sensitivity, specificity, and area under the
boundaries of the infected regions. The proposed diagnostic curve, are required to understand these models’ performance
system has shown promising results in identifying COVID- better.
19 infections. Additionally, Khan et al. [35] have introduced
a new CNN architecture called STM-RENet, which utilizes
a split-transform-merge approach to analyze X-ray images
and identify radiographic patterns associated with COVID- 2 Materials and Methods
19 infection. This block-based CNN includes a new con‑
volutional block named STM, which separately and jointly 2.1 Dataset
performs region and edge-based operations. By combining
these operations with convolutional techniques, STM-RENet A significant amount of research has been done using the
can analyze the homogeneity of regions, intensity inhomoge‑ Chest X-ray14 dataset [3, 7–9, 36–38]. The dataset has been
neity, and features that define boundaries. The authors have collected and made openly available by the National Institute
also presented an improved version of STM-RENet called of Health. It consists of 112120 frontal view chest X-ray
CB-STM-RENet, which utilizes channel boosting and learns images of 30805 unique patients. When loaded, these images
textural variations to enhance its performance. When evalu‑ are single-channel gray-scale images and need to be con‑
ated on three datasets, CB-STM-RENet demonstrated sig‑ verted to 3-channel images to allow our pre-trained model
nificantly superior results than conventional CNNs. to process them. Each image in the dataset is annotated with
A major limitation of previous research in the field of deep up to 14 different thoracic pathology labels. Figure 2 shows
CNNs for chest X-ray diagnosis is that many studies have a sample for each of these diseases from the dataset itself.
only examined their ability to perform binary classification Table 1 lists all the diseases in chest X-ray14.
tasks. These tasks involve detecting the presence or absence Wang et al. [39] used NLP to text-mine disease classi‑
of a specific disease. There is a need for more research on fications from the related radiological reports to label the
the ability of these models to simultaneously detect and images. The labels are expected to have an accuracy greater
classify multiple diseases or conditions in a single image. than 90% [40]. The dataset also consists of images labeled
Moreover, there are two additional issues in previous studies. as No Finding, which simply indicates that the NLP system

13
Interdisciplinary Sciences: Computational Life Sciences

Table 1  List of all diseases in chest X-ray14 2.3 Dataset Preprocessing

Pathology Total positives
The dataset needs to be preprocessed before building and
Atelectasis 11559 training the model. The X-ray images in the dataset were
Cardiomegaly 2776 preprocessed. For the training set, each image was standard‑
Effusion 13317 ized by subtracting the mean and standard deviation of that
Infiltration 19894 image from each pixel of that image.
Mass 5782
Nodule 6331 Xj[i] − X̄ [i]
Pneumonia 1431 X̂ j[i] =
𝜎 [i]
Pneumothorax 5302
Consolidation 4667 Here, i refers to the ith image in the training set, j refers to the
Edema 2303 jth pixel in the ith image, and X̂ represents the standardized
Emphysema 2516 image. Similarly, X̄ represents the mean of an image, and 𝜎
Fibrosis 1686 represents the standard deviation of an image.
Pleural Thickening 3385 For standardization of validation and test set, the mean
Hernia 227 and standard deviation for each channel were calculated
using a single batch of the training set, which was hypoth‑
esized to represent the mean and standard deviation of the
was unable to find any diseases for that particular image. It entire training set. Then the individual images in the vali‑
does not necessarily imply a healthy chest X-ray image. To dation and test were standardized (feature-wise) using the
be able to train on a TPU, the entire dataset was converted above-stated formula. After standardizing, the images in the
to tfrecord and made publicly available at NIH Chest X-ray dataset were re-scaled to 224 × 224; this was done to remain
TFRecords.1 consistent with the pre-trained models. The models used
were pre-trained on the ImageNet dataset with an input size
2.2 Dataset Distribution of 224 × 224 per image. After re-scaling, the images were
batched, with each batch containing 16 × 8 (for DenseNet
Table 1 provides a breakdown of the total positive labels for and ResNet50 models) and 8 × 8 (for EfficientNetB1 model)
each of the 14 pathology diseases. To train and evaluate the examples. The batch size was kept large to utilize TPU effi‑
models, the entire dataset was divided into three sets: a train‑ ciently. Figure 3 shows the generalized workflow taken while
ing set, a validation set, and a test set, in accordance with the developing models for this paper.
recommendations of [41–44]. It is important to note that the
dataset was not simply split into three parts. This is because
the dataset contains follow-up images for each patient, and a 3 Model Development
direct split could result in data leaks, leading to misleading
results. Instead, the dataset was first grouped by patient IDs, 3.1 Transfer Learning
treating each patient as a separate entity, then split into the
following three sets. Transfer learning refers to the idea of taking a model trained
on a different task and using this pre-trained model for a
• Training dataset: 85% of the total groups. downstream task. Figure 4 briefly describes the idea behind
• Validation dataset: 10% of the total groups. transfer learning. This paper uses a handful of models
• Test dataset: 5% of the total group. trained on the ImageNet [45] dataset, trains, or fine-tunes
them for the task of chest X-ray diagnosis using transfer
Hence, the distribution of the dataset used is given as learning. All model variants are trained on a TPU, and their
follows. AUROC scores are recorded on the held-out test set.

• Training dataset: 95466 examples. 3.2 DenseNet121

• Validation dataset: 11265 examples.
• Test dataset: 5389 examples. The first model used as the backbone for this task was
the DenseNet121 [46]. Densely Connected Convolutional

1
[Link]

13
 Interdisciplinary Sciences: Computational Life Sciences

Fig. 3  Generalized workflow for Data Pre-processing

the process
Training Set
(95,466 images)

Validation Set Rescale Images Batch Images

Standardization
(11,265 images) (224 × 224) 8 × 8⁄16 × 80

Test Set
(5,389 images)

Model Development
DenseNet121
ResNet50
EfficientNetB1

Model Training
Hardware: TPU
Epochs: 100
Early Stopping

Evaluation
Metric: AUROC

Result

input: ? , 224, 224, 3

input_18: InputLayer
output: ? , 224, 224, 3

input: ? , 224, 224, 3

densenet121: Functional
output: ? , 7, 7, 1024

input: ? , 7, 7, 1024
global_average_pooling2d_8: GlobalAveragePooling2D
output: ? , 1024

input: ? , 1024
dense_20: Dense
output: ? , 14

input: ? , 14
Sigmoid_1: Activation
output: ? , 14

Fig. 4  Transfer learning process

Fig. 5  Model architecture using DenseNet121 as a backbone

Networks, or simply DenseNet, is another way to keep
increasing the depth of a deep convolutional network. The
problem in deep convnets arises when they become so deep CheXNet model and is a re-implementation of the same. To
that the gradients vanish on their way back. Huang et al. [46] use transfer learning on the DenseNet121, the final dense
designed an architecture to ensure maximum gradient flow layer of the pre-trained model was replaced by a Dense layer
during back-propagation to resolve this problem. DenseNet with 14 units, and a sigmoid activation was applied to it. The
exploits the potential of the network through feature reuse. resultant model architecture is shown in Fig. 5. In addition,
The architecture for DenseNets is exhibited in Fig. 6(a). This the training parameters for the DenseNet121 backbone are
paper uses a DenseNet121, a 121-layered convolutional neu‑ given in Table 2.
ral network model. This model derives inspiration from the

13
Interdisciplinary Sciences: Computational Life Sciences

Fig. 6  Model architecture on ImageNet for a DenseNet [46]; b ResNet [47]; and c EfficientNet-B0 baseline network [48]

13
 Interdisciplinary Sciences: Computational Life Sciences

Table 2  Training parameters for DenseNet121 backbone

input: ? , 224, 224, 3
input_16: InputLayer
Training parameter Value output: ? , 224, 224, 3

Input shape (224, 224) input: ? , 224, 224, 3

resnet50V2: Functional
Output shape (14, ) output: ? , 7, 7, 2048
Batch size 16 × 8 (16 batches per TPU core)
input: ? , 7, 7, 2048
Callbacks Model checkpoint global_average_pooling2d_7: GlobalAveragePooling2D
output: ? , 2048
Reduce LR on plateau
Early stopping input: ? , 2048
dense_16: Dense
output: ? , 1024

input: ? , 1024
3.3 ResNet50 ReLU_1: Activation
output: ? , 1024

The second model used as the backbone for the diagnosis input: ? , 1024
dropout_12: Dropout
task was the ResNet50 [47]. ResNets are an exciting class output: ? , 1024
of models and have served as the state-of-the-art model for
input: ? , 1024
various tasks. A deep neural network architecture tends to dense_17: Dense
output: ? , 512
give a more significant error as compared to a comparatively
shallow neural network. He et al. [47] overcame this problem input: ? , 512
ReLU_2: Activation
by introducing a deep residual learning framework and skip output: ? , 512

or residual connections. Figure 6(b) exposes the architec‑

input: ? , 512
tures of different ResNets. dropout_13: Dropout
output: ? , 512
In this work, ResNet50V2 uses 50 layered deep convo‑
lutional neural networks with multiple residual connec‑ input: ? , 512
dense_18: Dense
tions. The pre-trained model was modified to apply trans‑ output: ? , 256
fer learning for this diagnosis task. For this task, the final
input: ? , 256
fully connected layers of the pre-trained ResNet50V2 model ReLU_3: Activation
output: ? , 256
were replaced by an average pooling, followed by a series
of dense, relu, and dropout layers. A final dense layer with input: ? , 256
dropout_14: Dropout
14 units followed by a sigmoid was utilized for the output. output: ? , 256
The resultant model architecture is shown in Fig. 7. Also,
the training parameters for the ResNet50V2 backbone are
input: ? , 256
dense_19: Dense
depicted in Table 3.
output: ? , 14

input: ? , 14
3.4 EfficientNetB1 Sigmoid_1: Activation
output: ? , 14

The third and last model used as the backbone for this task
was EfficientNetB1 [48]. EfficientNets are a class of effi‑ Fig. 7  Model architecture using ResNet50V2 as the backbone
ciently designed models to optimize the model’s perfor‑
mance while having a considerably low amount of train‑
able parameters. Tan and Le [48] came up with a better way Table 3  Training parameters for ResNet50V2 backbone
of scaling the network, which they call compound scaling,
Training parameter Value
in which they selected an efficient scaling for all − width,
depth, and image resolution. The baseline network architec‑ Input shape (224, 224)
ture of EfficientNetB0 is shown in Fig. 6(c). Output shape (14, )
To use transfer learning on the EfficientNetB1, the final Batch size 16 × 8 (16 batches per TPU core)
Dense layer of the pre-trained model was replaced by a Callbacks Model checkpoint
Dense layer with 14 units, and a sigmoid activation was Reduce LR on plateau
applied to it. The resultant model architecture is shown in Early stopping
Fig. 8. Also, the training parameters for the EfficientNetB1
backbone are presented in Table 4.

13
Interdisciplinary Sciences: Computational Life Sciences

14
input: ? , 224, 224, 3 ∑
input_20: InputLayer J= L(yc , ŷ c )
output: ? , 224, 224, 3
c=1
N
input: ? , 224, 224, 3 1 ∑
efficientnetb1: Functional
L(yc , ŷ c ) = [−wpc × yci × log(̂yci ) − wnc × (1 − yci )
output: ? , 7, 7, 1280
N i=1
× log(1 − ŷ ci )]
input: ? , 7, 7, 1280
global_average_pooling2d_9: GlobalAveragePooling2D
output: ? , 1280 Here, c refers to one of the classes of labels, i refers to the ith
example in the training set, y refers to true label, and ŷ refers
input: ? , 1280 to the predicted label or probability. Similarly, wpc and wnc
dense_21: Dense
output: ? , 14
are defined as follows.
Total negative examples in class c
input: ? , 14
wpc =
Sigmoid_1: Activation
Total examples
output: ? , 14 Total positive examples in class c
wnc =
Total examples
Fig. 8  Model architecture using EfficientNetB1 as a backbone
The model was trained on a TPU V3.8 on Kaggle for 100
epochs. The above-mentioned custom loss, binary accu‑
Table 4  Training parameters for EfficientNetB1 backbone racy, and AUROC score were monitored during training.
Training parameter Value
[49] optimizer was used for training. The learning rate was
reduced by a factor of 10 if no improvements in validation
Input shape (224, 224) loss were seen for two continuous epochs. Early stopping
Output shape (14, ) was used with the patience of 10 to prevent over-fitting of
Batch size 8 × 8 (8 batches per TPU core) the model and prevent wastage of computing time. The end-
Callbacks Model checkpoint to-end open-source deep learning framework Tensor Flow2
Reduce LR on plateau was used to train and evaluate the models.
Early stopping

5 Results and Performance

4 Model Training
The overall workflow for disease prediction in sample X-rays
The models were individually trained on a TPU using vari‑ is graphically produced, as shown in Fig. 9.
ous batch sizes while maintaining the same hyper-param‑ In this study, the metric used for comparing the models
eters. The models were initialized with parameters from is the AUROC score and curve. In this section, the results
a network that was pre-trained on ImageNet [25] before of the models are discussed and compared. Additionally,
commencing training. The last layers were replaced with all the models are compared to the CheXNet model [3].
a dense layer having 14 units followed by a sigmoid layer AUROC (Area Under the Receiver Operator Curve) is a
for obtaining the predicted probabilities of all 14 pathology performance metric that evaluates classification models for
diseases, as discussed in the previous sections. The images various threshold values. ROC is a probability curve, and
were resized to 224 × 224 pixels before inputting them. Prior AUC represents the degree of separability. The higher the
to feeding the network with the image, each image in the AUC value, the better a model can differentiate between the
training set was subject to random horizontal flips and ran‑ positive and negative classes. An AUROC curve plots the
dom rotations up to 10 degrees. true positive rate against the false-positive rate.
The true positive rate, also referred to as sensitivity,
4.1 Loss Function measures the proportion of positive examples in the data‑
set that the model accurately identified as positive. In other
The data in the dataset were imbalanced. To account for this words, it represents the fraction of total positive examples
imbalance, instead of simply using a binary cross-entropy correctly predicted as positive by the model, i.e.,
loss function, a weighted binary cross-entropy loss is mini‑
mized, as suggested in [3].

2
[Link]

13
 Interdisciplinary Sciences: Computational Life Sciences

Fig. 9  Proposed CXD server, where a Front view of the CXD where given sample using Deep Learning approaches; and d Depict previ‑
boxes are given for sequences and blue button representing submis‑ ous history along with confidence scores generated from embedded
sion option; b Uploading Sample Chest X-ray; c Predicted results of Deep learning approaches

True positive
TPR = Table 5  AUROC scores for different model variants
True positives false negatives
Pathology CheXNet DenseNet121 ResNet50 EfficientNetB1
The false positive rate, also given as 1-sensitivity, is the
fraction of total negative examples in the dataset the model Atelectasis 0.8094 0.8200 0.7630 0.7750
incorrectly predicted as positive, i.e., Cardiomegaly 0.9248 0.9120 0.7410 0.8840
Effusion 0.8638 0.8830 0.8600 0.8460
False positive Infiltration 0.7345 0.7290 0.6660 0.6450
FPR =
True positives false negatives Mass 0.8676 0.8210 0.7480 0.7610
Nodule 0.7802 0.7200 0.5720 0.6240
The AUROC scores for different classes for all three model Pneumonia 0.7680 0.7430 0.6940 0.7130
variants have been listed in Table 5, and the ROC curves Pneumothorax 0.8887 0.8670 0.7790 0.8100
have been illustrated in Fig. 10. The experimental results Consolidation 0.7901 0.8240 0.8030 0.7960
demonstrated that the model constructed using DenseNet121 Edema 0.8878 0.8900 0.8710 0.8380
outperformed the other two models. Therefore, our experi‑ Emphysema 0.9371 0.9120 0.7600 0.8410
mental investigations are in line with several studies [50–52] Fibrosis 0.8047 0.8080 0.6870 0.7310
previously published in the literature. Some possible reasons Pleural Thick‑ 0.8062 0.7470 0.7150 0.7120
for this superior performance could be: ening
Hernia 0.9164 0.9450 0.6970 0.8490

13
Interdisciplinary Sciences: Computational Life Sciences

Fig. 10  a AUROC curve for DenseNet121 backbone; b AUROC curve for ResNet50 backbone; and c AUROC curve for EfficientNetB1 back‑
bone

• The network architecture of DenseNet121 is more com‑ • DenseNet121 incorporates batch normalization and skip
plex, enabling the model to extract more features from connections to enhance convergence and performance.
the data and potentially improve its performance.
• The DenseNet121 model utilizes a more closely Further, we employed a pre-trained DenseNet121
connected pattern between its layers, which can aid in model and modified its fully connected layers as previ‑
decreasing the number of parameters within the model ously described. Subsequently, we assessed the model’s
and avoiding overfitting.

13
 Interdisciplinary Sciences: Computational Life Sciences

performance without any further training and found that of the model will be at least that value. These intervals
the ROC values were approximately 0.5, indicating that help to understand the uncertainty surrounding a model’s
its predictions were akin to random guessing. Figure 11(a) performance and to compare the performance of different
illustrates the results graphically. Similarly, we froze the models. Based on the assumption of a normal distribution,
DenseNet121 model and only trained the global aver‑ we have provided tables below that show the minimum and
age pooling (GAP) and final softmax layers, using the maximum estimated prevalence of Atelectasis disease with
DenseNet121 model as the backbone, without modifying a 95% confidence level for ResNet50, EfficientNetB1, and
its parameters. The outcomes of this method are demon‑ DenseNet121 models. These tables, labeled as Tables 6, 7,
strated in Fig. 11(b). Essentially, we maintained the original 8, 9, 10,11, indicate the minimum estimated prevalence of
DenseNet121 model and only updated the final layers that Atelectasis disease with a 95% confidence level based on
were added to it. the assumption of a normal distribution for the respective
Furthermore, DenseNet121 was utilized as an unalter‑ models. In these tables, TPR represents the true positive
able feature extractor, and a more intricate fully connected rate, and FPR represents the false positive rate.
layer was trained on it. Only the fully connected layers were Next, we have thoroughly analyzed the lower and upper
modified during the training process, and the rest of the bounds of the precision-recall (PR) curves for the three
DenseNet121 architecture remained unchanged. Figure 12 models being considered. Nevertheless, it is important to
illustrates the new layers appended to the fully connected note that the ROC curve is also a suitable alternative for
network. The ROC curve achieved from this approach is evaluating a classifier’s performance, particularly for data‑
displayed in Fig. 13. sets that have imbalanced classes. This is because, unlike
We have computed the lower and upper confidence the PR curve, which only considers the true positive rate,
intervals for ResNet50, EfficientNetB1, and DenseNet121 the ROC curve considers both the true positive rate and
to further analyze these models. A confidence interval is the false positive rate. Therefore, we have included both
a range of values that are likely to include the true value the PR and ROC curves to provide a more comprehensive
of a population parameter. The lower and upper confidence evaluation of these models, which are depicted in Figs. 14,
intervals for these models indicate the potential range of 15, and 16.
performance when applied to a specific task or dataset. The aforementioned tables, given as Tables 6, 7,
For instance, if the lower confidence interval for a model’s 8, 9, 10,11, also display the F1 score, also known as
accuracy is 95%, there is a 95% chance that the true accuracy the F-measure or F-score. It is a metric that combines

Fig. 11  a ROC curve for DenseNet121 without training model; and b ROC curve for DenseNet121 model after training only the GAP and final
softmax layers

13
Interdisciplinary Sciences: Computational Life Sciences

Fig. 12  The architecture

of extra added layers to the
fully connected network of
DenseNet121

precision and recall into a single score, commonly used year. India has over 10 million cases of pneumonia each
in classification tasks. The score is calculated as the year. Although chest X-rays are the most effective means of
harmonic mean of precision and recall. Precision is the diagnosing pneumonia [53], medical imaging has constraints
number of true positive predictions divided by the total in terms of access to expertise in some areas [54]. Addition‑
number of positive predictions, and recall is the number ally, chest radiographs may also be utilized to diagnose other
of true positive predictions divided by the total number illnesses.
of actual positive samples. The F1 score is valuable for In addition, even expert radiologists are limited by vari‑
evaluating the performance of classification models ous human factors [38, 55–58]. Therefore, the creation of
because it provides a balance between precision and recall detection systems could greatly benefit humanity. As a result
and allows for the comparison of models with different of the difficulty in training these three models on a CPU,
precision and recall values. this study considers using a TPU. The CXD server has an
improved interface that is more efficient and has been devel‑
oped using a large chest X-ray dataset up until January 2021.
6 Conclusions and Future Scopes This extensive and accurate data is being constantly utilized
to enhance our proposed CXD server, ensuring the quality
Pneumonia is a major cause of human fatalities worldwide. of our work.
According to the Centers for Disease Control and Preven‑ The objective of this study was to automate an essen‑
tion,3 over one million adults in the US are hospitalized due tial stage of the radiology process by utilizing three con‑
to pneumonia, and around 50, 000 die from the disease each volutional neural networks (CNNs), namely DenseNet121,
ResNet50, and EfficientNetB1, to precisely detect 14 types
of thoracic pathology diseases from chest radiography
3
[Link] images. A total of 112, 120 chest X-ray datasets containing

13
 Interdisciplinary Sciences: Computational Life Sciences

Fig. 13  ROC curve for extra

added layers to the fully con‑
nected network of DenseNet121

Table 6  The minimum Threshold TPR FPR Accuracy Precision Recall F1 score
estimated prevalence of
Atelectasis disease with a 95% 0.1 0.805903354 0.015975 0.895843 0.972882 0.805903 0.88118
confidence level based on
0.2 0.841285476 0.029508 0.895843 0.966523 0.841285 0.899306
the assumption of a normal
distribution for ResNet50 0.3 0.865006785 0.047695 0.895843 0.959846 0.865007 0.909793
0.4 0.882387525 0.068971 0.895843 0.954492 0.882388 0.916916
0.5 0.899576808 0.101926 0.895843 0.949523 0.899577 0.923817
0.6 0.918142515 0.163129 0.895843 0.942756 0.918143 0.93027
0.7 0.943963684 0.3045 0.895843 0.928266 0.943964 0.936041
0.8 0.972310777 0.600504 0.895843 0.912073 0.972311 0.941086
0.9 0.999218124 0.990938 0.895843 0.896111 0.999218 0.943991

various thoracic pathology diseases were utilized to evalu‑ clinicians to potentially abnormal findings. The results indi‑
ate the performance of these models based on their abil‑ cated that the DenseNet121 model outperformed the other
ity to predict the likelihood of individual diseases and alert two models in terms of the score values achieved for each

13
Interdisciplinary Sciences: Computational Life Sciences

Table 7  The maximum Threshold TPR FPR Accuracy Precision Recall F1 score
estimated prevalence of
Atelectasis disease with a 95% 0.1 0.82011924 0.020878 0.906718 0.978497 0.820119 0.892725
confidence level based on
0.2 0.854381302 0.035998 0.906718 0.972778 0.854381 0.910013
the assumption of a normal
distribution for ResNet50 0.3 0.877223562 0.05577 0.906718 0.966704 0.877224 0.919967
0.4 0.893879821 0.078499 0.906718 0.961793 0.89388 0.926705
0.5 0.91027128 0.113223 0.906718 0.95721 0.910271 0.933211
0.6 0.927862328 0.176824 0.906718 0.950935 0.927862 0.939273
0.7 0.952058015 0.321407 0.906718 0.937392 0.952058 0.944676
0.8 0.97798659 0.618292 0.906718 0.922126 0.977987 0.949381
0.9 0.999957983 0.994081 0.906718 0.906973 0.999958 0.952083

Table 8  The minimum Threshold TPR FPR Accuracy Precision Recall F1 score
estimated prevalence of
Atelectasis disease with a 95% 0.1 0.69087922 0.00361946 0.89584263 0.98295658 0.69087922 0.81089795
confidence level based on
0.2 0.82653939 0.02196173 0.89584263 0.97044309 0.82653939 0.89240883
the assumption of a normal
distribution for EfficientNetB1 0.3 0.88065577 0.06462379 0.89584263 0.95866499 0.88065577 0.91787436
0.4 0.91498338 0.14903047 0.89584263 0.94340845 0.91498338 0.92895791
0.5 0.93846433 0.27433506 0.89584263 0.93229964 0.93846433 0.9353707
0.6 0.95526986 0.44986595 0.89584263 0.92421408 0.95526986 0.93945243
0.7 0.97303584 0.65328083 0.89584263 0.91259259 0.97303584 0.94169824
0.8 0.98557233 0.84551181 0.89584263 0.90482787 0.98557233 0.94315702
0.9 0.99518978 0.97055598 0.89584263 0.89868163 0.99518978 0.94388506

Table 9  The maximum Threshold TPR FPR Accuracy Precision Recall F1 score
estimated prevalence of
Atelectasis disease with a 95% 0.1 0.70759942 0.00616315 0.90671814 0.98736494 0.70759942 0.8249679
confidence level based on
0.2 0.84012727 0.02763145 0.90671814 0.97631232 0.84012727 0.90344688
the assumption of a normal
distribution for EfficientNetB1 0.3 0.8922237 0.07388037 0.90671814 0.96562332 0.8922237 0.92760923
0.4 0.92487865 0.16224773 0.90671814 0.95154207 0.92487865 0.93804209
0.5 0.94693812 0.29075071 0.90671814 0.94117529 0.94693812 0.94404906
0.6 0.96250887 0.46803446 0.90671814 0.93358409 0.96250887 0.94785958
0.7 0.97863467 0.67052869 0.90671814 0.92261785 0.97863467 0.94995146
0.8 0.98960864 0.85845931 0.90671814 0.9152598 0.98960864 0.95130836
0.9 0.99740433 0.97641372 0.90671814 0.90941992 0.99740433 0.95198497

Table 10  The minimum Threshold TPR FPR Accuracy Precision Recall F1 score
estimated prevalence of
Atelectasis disease with a 95% 0.1 0.84306717 0.03027407 0.89584263 0.96648739 0.84306717 0.90031113
confidence level based on
0.2 0.87086545 0.05159917 0.89584263 0.96172846 0.87086545 0.91387505
the assumption of a normal
distribution for DenseNet121 0.3 0.88815469 0.07713721 0.89584263 0.95646389 0.88815469 0.92093865
0.4 0.90111604 0.10671467 0.89584263 0.95189192 0.90111604 0.92574521
0.5 0.91569139 0.15556943 0.89584263 0.94509502 0.91569139 0.93013811
0.6 0.9282404 0.21364821 0.89584263 0.93823539 0.9282404 0.93320833
0.7 0.93842546 0.28161416 0.89584263 0.93317418 0.93842546 0.93579162
0.8 0.94847708 0.37734796 0.89584263 0.92822659 0.94847708 0.93822927
0.9 0.96335002 0.56196969 0.89584263 0.9196265 0.96335002 0.94091021

13
 Interdisciplinary Sciences: Computational Life Sciences

Table 11  The maximum Threshold TPR FPR Accuracy Precision Recall F1 score
estimated prevalence of
Atelectasis disease with a 95% 0.1 0.85610088 0.03684002 0.90671814 0.97274542 0.85610088 0.91096946
confidence level based on
0.2 0.88284634 0.05996685 0.90671814 0.96842214 0.88284634 0.92383087
the assumption of a normal
distribution for DenseNet121 0.3 0.89938912 0.08714851 0.90671814 0.96360561 0.89938912 0.93049949
0.4 0.91173454 0.11823421 0.90671814 0.95939841 0.91173454 0.93502422
0.5 0.92554771 0.16901291 0.90671814 0.95310858 0.92554771 0.93914936
0.6 0.93736856 0.22878152 0.90671814 0.94672453 0.93736856 0.94202619
0.7 0.94690186 0.29815644 0.90671814 0.94199421 0.94690186 0.94444251
0.8 0.9562432 0.39509976 0.90671814 0.9373556 0.9562432 0.94671882
0.9 0.96989882 0.58001505 0.90671814 0.92926239 0.96989882 0.94921784

1 1
Atelectasis Atelectasis
0.98 0.98
Cardiomegaly Cardiomegaly
0.96 Consolidation 0.96 Consolidation

0.94 Edema 0.94 Edema

Emphysema
Precision

Emphysema 0.92

Precision
0.92
Fibrosis Fibrosis
0.9
0.9 Hernia
Hernia
0.88
0.88 Infiltration
Infiltration
0.86 Mass
0.86 Mass
0.84 Nodule
0.84 Nodule
Pleural_Thickening
Pleural_Thickening 0.82
0.82 Pneumonia
Pneumonia 0.8
0.8 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Pneumothorax
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Pneumothorax
Recall Effusion
Recall Effusion
(a) (b)

Fig. 14  PR curve for a lower bound, and b upper bound at 95% confidence interval for ResNet50 model

1 1
Atelectasis Atelectasis
0.98 Cardiomegaly 0.98 Cardiomegaly

0.96 Consolidation 0.96 Consolidation

Edema Edema
0.94 0.94
Emphysema Emphysema
0.92 Fibrosis 0.92
Precision

Fibrosis
Precision

0.9 Hernia 0.9 Hernia

Infiltration Infiltration
0.88 0.88
Mass Mass
0.86 Nodule 0.86 Nodule

0.84 Pleural_Thickening 0.84 Pleural_Thickening

Pneumonia Pneumonia
0.82 0.82
Pneumothorax Pneumothorax
0.8 Effusion 0.8 Effusion
0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1
Recall Recall
(a) (b)

Fig. 15  PR curve for a lower bound, and b upper bound at 95% confidence interval for EfficientNetB1 model

1
1 Atelectasis
Atelectasis
0.98 Cardiomegaly
0.98 Cardiomegaly
0.96 Consolidation
0.96 Consolidation
Edema Edema
0.94 0.94
Emphysema Emphysema
0.92 0.92 Fibrosis
Fibrosis
Precision

Precision

0.9 Hernia 0.9 Infiltration

Infiltration Mass
0.88 0.88
Mass Nodule
0.86 0.86 Pleural_Thickening
Nodule
0.84 Pleural_Thickening 0.84 Pneumonia

Pneumonia Pneumothorax
0.82 0.82
Pneumothorax Effusion
0.8 0.8
Hernia
Effusion
0.8 0.85 0.9 0.95 1 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1
Recall Recall
(a) (b)

Fig. 16  PR curve for a lower bound, and b upper bound at 95% confidence interval for DenseNet121 model

13
Interdisciplinary Sciences: Computational Life Sciences

class on the dataset. Furthermore, the performance of these learning. arXiv preprint arXiv:​1711.​05225. [Link]
models was compared to that of the ChexNet model. 48550/​arXiv.​1711.​05225
4. Rajpurkar P, Irvin J, Ball RL, Zhu K, Yang B, Mehta H, Duan T,
Our future plans involve expanding our CNN training by Ding D, Bagul A, Langlotz CP et al (2018) Deep learning for chest
incorporating extra data and assessing different architectures radiograph diagnosis: a retrospective comparison of the chexnext
for diagnosing other thoracic pathology diseases. We are algorithm to practicing radiologists. PLoS Med 15(11):1002686.
convinced that a computer-aided diagnostic tool of this kind [Link]
5. Cicero M, Bilbily A, Colak E, Dowdell T, Gray B, Perampaladas
could enhance the effectiveness and precision of diagnosing K, Barfett J (2017) Training and validating a deep convolutional
thoracic pathology diseases, including pandemics like neural network for computer-aided detection and classification
COVID-19 and Swine Flu, significantly. This tool could prove of abnormalities on frontal chest radiographs. Investig Radiol
to be especially useful during a pandemic when the demand 52(5):281–287. [Link]
6. Bar Y, Diamant I, Wolf L, Lieberman S, Konen E, Greenspan H
for prevention and treatment often surpasses the available (2015) Chest pathology detection using deep learning with non-
resources. medical training. In: 2015 IEEE 12th International Symposium on
Biomedical Imaging, pp. 294–297. [Link]
2015.​71638​71. IEEE
7. Guendel S, Grbic S, Georgescu B, Liu S, Maier A, Comaniciu D
Appendix (2018) Learning to recognize abnormalities in chest x-rays with
location-aware dense networks. In: Iberoamerican Congress on
The proposed server “CXD” is accessible at: [Link] Pattern Recognition, pp. 757–765. [Link]
google.​com/​file/d/​1gKJN​FfJc2​FQoDo​4lGz1​0wdTe​nbcAh​ 3-​030-​13469-3_​88. Springer
8. Yan C, Yao J, Li R, Xu Z, Huang J (2018) Weakly supervised
C73/​view?​usp=​shari​ng. deep learning for thoracic disease classification and localiza‑
NIH dataset (tfrecords) can be accessible at: [Link] tion on chest x-rays. In: Proceedings of the 2018 ACM Interna‑
www.​kaggle.​com/​harsh​soni/​nih-​chest-​xray-​tfrec​ords. tional Conference on Bioinformatics, Computational Biology,
Kaggle TPU documentations can be accessible at: and Health Informatics, pp. 103–110. [Link]
32335​47.​32335​73
[Link] 9. Wang H, Xia Y (2018) Chestnet: a deep neural network for
The source code and sample data can be accessible at: classification of thoracic diseases on chest radiography. arXiv
[Link] preprint arXiv:​1 807.​0 3058 . [Link] oi.​o rg/​1 0.​4 8550/​a rXiv.​
1807.​03058
10. Lakhani P, Sundaram B (2017) Deep learning at chest radiog‑
Supplementary Information The online version contains supplemen‑ raphy: automated classification of pulmonary tuberculosis by
tary material available at [Link] oi.o​ rg/1​ 0.1​ 007/s​ 12539-​023-​00562-2. using convolutional neural networks. Radiology 284(2):574–582.
[Link]
Acknowledgements Dong-Qing Wei is funded by the National Science 11. Stirenko S, Kochura Y, Alienin O, Rokovyi O, Gordienko Y, Gang
Foundation of China (Grant numbers 32070662, 61832019, 32030063), P, Zeng W (2018) Chest x-ray analysis of tuberculosis by deep
the Science and Technology Commission of Shanghai Municipality learning with segmentation and augmentation. In: 2018 IEEE
(Grant number 19430750600), the SJTU JiRLMDS Joint Research 38th International Conference on Electronics and Nanotechnol‑
Fund, the Joint Research Funds for Medical and Engineering and Sci‑ ogy (ELNANO), pp. 422–428. [Link] oi.o​ rg/1​ 0.1​ 109/E ​ LNANO.​
entific Research at Shanghai Jiao Tong University (YG2021ZD02), and 2018.​84775​64. IEEE
the PCL Major Key Project (PCL2021A13). The computations were 12. Maduskar P, Muyoyeta M, Ayles H, Hogeweg L, Peters-Bax L,
partially performed at the Pengcheng Lab and the Center for High- vanGinneken B (2013) Detection of tuberculosis using digital
Performance Computing at Shanghai Jiao Tong University. chest radiography: automated reading vs. interpretation by clini‑
cal officers. Int J Tuberc Lung Dis 17(12):1613–1620. [Link]
org/​10.​5588/​ijtld.​13.​0325
Declarations 13. Hwang S, Kim H-E, Jeong J, Kim H-J (2016) A novel approach
for tuberculosis screening based on deep convolutional neural net‑
Conflict of interest The corresponding author states, on behalf of all works. In: Medical Imaging 2016: Computer-aided Diagnosis, vol.
authors, that there is no conflict of interest. 9785, pp. 750–757. [Link] SPIE
14. Rahman T, Chowdhury ME, Khandakar A, Islam KR, Islam KF,
Mahbub ZB, Kadir MA, Kashem S (2020) Transfer learning with
deep convolutional neural network (CNN) for pneumonia detec‑
References tion using chest x-ray. Appl Sci 10(9):3233. [Link]
3390/​app10​093233
1. Krizhevsky A, Sutskever I, Hinton GE (2017) Imagenet classifi‑ 15. Dey N, Zhang Y-D, Rajinikanth V, Pugalenthi R, Raja NSM
cation with deep convolutional neural networks. Commun ACM (2021) Customized vgg19 architecture for pneumonia detection
60(6):84–90. [Link] in chest x-rays. Pattern Recognit Lett 143:67–74. [Link]
2. Raoof S, Feigin D, Sung A, Raoof S, Irugulpati L, Rosenow EC 10.​1016/j.​patrec.​2020.​12.​010
III (2012) Interpretation of plain chest roentgenogram. Chest 16. Wang L, Lin ZQ, Wong A (2020) Covid-net: a tailored deep con‑
141(2):545–558. [Link] volutional neural network design for detection of covid-19 cases
3. Rajpurkar P, Irvin J, Zhu K, Yang B, Mehta H, Duan T, Ding from chest x-ray images. Sci Rep 10(1):1–12. [Link]
D, Bagul A, Langlotz C, Shpanskaya K et al. (2017) Chexnet: 1038/​s41598-​020-​76550-z
radiologist-level pneumonia detection on chest x-rays with deep

13
 Interdisciplinary Sciences: Computational Life Sciences

17. Alakus TB, Turkoglu I (2020) Comparison of deep learning covid-19 and their usability for detecting omicron. J Exp Theor
approaches to predict covid-19 infection. Chaos Solit Fractals Artif Intell. [Link]
140:110120. [Link] 33. Khan SH, Sohail A, Khan A, Hassan M, Lee YS, Alam J, Basit A,
18. Basu S, Mitra S, Saha N (2020) Deep learning for screening Zubair S (2021) Covid-19 detection in chest x-ray images using
covid-19 using chest x-ray images. In: 2020 IEEE Symposium deep boosted hybrid learning. Comput Biol Med 137:104816.
Series on Computational Intelligence (SSCI), pp. 2521–2527. [Link]
[Link] IEEE 34. Khan SH (2022) Covid-19 detection and analysis from lung ct
19. Jamshidi M, Lalbakhsh A, Talla J, Peroutka Z, Hadjilooei F, Lal‑ images using novel channel boosted cnns. arXiv preprint arXiv:​
bakhsh P, Jamshidi M, La Spada L, Mirmozafari M, Dehghani 2209.​10963. [Link]
M et al (2020) Artificial intelligence and covid-19: deep learning 35. Khan SH, Sohail A, Khan A, Lee Y-S (2022) Covid-19 detec‑
approaches for diagnosis and treatment. IEEE Access 8:109581– tion in chest x-ray images using a new channel boosted cnn.
109595. [Link] Diagnostics 12(2):267. [Link]
20. Zhang J, Xie Y, Li Y, Shen C, Xia Y (2020) Covid-19 screen‑ 0267
ing on chest x-ray images using deep learning based anomaly 36. Kumar P, Grewal M, Srivastava MM (2018) Boosted cascaded
detection. arXiv preprint arXiv:​2003.​1233827. [Link] convnets for multilabel classification of thoracic diseases in
48550/​arXiv.​2003.​12338 chest radiographs. In: International Conference Image Analysis
21. Wang S, Zha Y, Li W, Wu Q, Li X, Niu M, Wang M, Qiu X, Li and Recognition, pp. 546–552. [Link]
H, Yu H et al (2020) A fully automatic deep learning system for 319-​93000-8_​62. Springer
covid-19 diagnostic and prognostic analysis. Eur Respir J. [Link] 37. Yao L, Poblenz E, Dagunts D, Covington B, Bernard D, Lyman
doi.​org/​10.​1183/​13993​003.​00775-​2020 K (2017) Learning to diagnose from scratch by exploiting
22. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, dependencies among labels. arXiv preprint arXiv:​1710.​10501.
Thrun S (2017) Dermatologist-level classification of skin cancer [Link]
with deep neural networks. Nature 542(7639):115–118. [Link] 38. Pesce E, Withey SJ, Ypsilantis P-P, Bakewell R, Goh V, Mon‑
doi.​org/​10.​1038/​natur​e21056 tana G (2019) Learning to detect chest radiographs containing
23. Rezvantalab A, Safigholi H, Karimijeshni S (2018) Dermatologist pulmonary lesions using visual attention networks. Med Image
level dermoscopy skin cancer classification using different deep Anal 53:26–38. [Link]
learning convolutional neural networks algorithms. arXiv preprint 39. Wang X, Peng Y, Lu L, Lu Z, Bagheri M, Summers RM (2017)
arXiv:​1810.​10348. [Link] Chestx-ray8: Hospital-scale chest x-ray database and bench‑
24. Hosny KM, Kassem MA, Foaud MM (2018) Skin cancer clas‑ marks on weakly-supervised classification and localization of
sification using deep learning and transfer learning. In: 2018 9th common thorax diseases. In: 2017 IEEE Conference on Com‑
Cairo International Biomedical Engineering Conference (CIBEC), puter Vision and Pattern Recognition, pp. 3462–3471. [Link]
pp. 90–93. [Link] IEEE doi.​org/​10.​1109/​CVPR.​2017.​369
25. Manning DJ, Ethell S, Donovan T (2004) Detection or decision 40. Summers R (2019) NIH chest X-ray dataset of 14 common tho‑
errors? Missed lung cancer from the posteroanterior chest radio‑ rax disease categories. figshare [Link]
graph. Br J Radiol 77(915):231–235. [Link] Chest​Xray-​NIHCC/​file/​22066​07896​10
28883​951 41. Chen Y, Zhang Y, Huang Z, Luo Z, Chen J (2021) Celebhair: A
26. Ciompi F, Chung K, Van Riel SJ, Setio AAA, Gerke PK, Jacobs new large-scale dataset for hairstyle recommendation based on
C, Scholten ET, Schaefer-Prokop C, Wille MM, Marchiano A et al celeba. In: Knowledge Science, Engineering and Management:
(2017) Towards automatic pulmonary nodule management in lung 14th International Conference, KSEM 2021, Tokyo, Japan,
cancer screening with deep learning. Sci Rep 7(1):1–11. [Link] August 14–16, 2021, Proceedings, Part III, pp. 323–336. [Link]
doi.​org/​10.​1038/​srep4​6479 doi.​org/​10.​1007/​978-3-​030-​82153-1_​27. Springer
27. Ardila D, Kiraly AP, Bharadwaj S, Choi B, Reicher JJ, Peng L, 42. Shrimali S (2021) Plantifyai: a novel convolutional neural net‑
Tse D, Etemadi M, Ye W, Corrado G et al (2019) End-to-end work based mobile application for efficient crop disease detec‑
lung cancer screening with three-dimensional deep learning on tion and treatment. Procedia Comput Sci 191:469–474. [Link]
low-dose chest computed tomography. Nat Med 25(6):954–961. doi.​org/​10.​1016/j.​procs.​2021.​07.​059
[Link] 43. Kenter T, Jones L, Hewlett D (2018) Byte-level machine reading
28. Baltruschat IM, Nickisch H, Grass M, Knopp T, Saalbach A across morphologically varied languages. In: Proceedings of the
(2019) Comparison of deep learning approaches for multi-label AAAI Conference on Artificial Intelligence, vol. 32. [Link]
chest x-ray classification. Sci Rep 9(1):1–10. [Link] org/​10.​1609/​aaai.​v32i1.​12050
1038/​s41598-​019-​42294-8 44. Mariani S, Rendu Q, Urbani M, Sbarufatti C (2021) Causal
29. Ayan E, Ünver HM (2019) Diagnosis of pneumonia from chest dilated convolutional neural networks for automatic inspection
x-ray images using deep learning. In: 2019 Scientific Meeting on of ultrasonic signals in non-destructive evaluation and struc‑
Electrical-Electronics & Biomedical Engineering and Computer tural health monitoring. Mech Syst Signal Process 157:107748.
Science (EBBT), pp. 1–5. [Link] [Link]
87415​82. IEEE 45. Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L (2009) Ima‑
30. Khan SH, Sohail A, Khan A, Lee YS (2020) Classification and genet: A large-scale hierarchical image database. In: 2009 IEEE
region analysis of covid-19 infection using lung ct images and Conference on Computer Vision and Pattern Recognition, pp.
deep convolutional neural networks. arXiv preprint arXiv:​2009.​ 248–255. [Link] IEEE
08864. [Link] 46. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ (2017)
31. Khan SH, Sohail A, Zafar MM, Khan A (2021) Coronavirus dis‑ Densely connected convolutional networks. In: 2017 IEEE
ease analysis using chest x-ray images and a novel deep convolu‑ Conference on Computer Vision and Pattern Recognition, pp.
tional neural network. Photodiagnosis Photodyn Ther 35:102473. 2261–2269 . [Link] oi.​o rg/​1 0.​1 109/​C VPR.​2 017.​2 43. IEEE
[Link] Computer Society
32. Khan A, Khan SH, Saif M, Batool A, Sohail A, Waleed Khan M 47. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for
(2023) A survey of deep learning techniques for the analysis of image recognition. In: Proceedings of the IEEE Conference on

13
Interdisciplinary Sciences: Computational Life Sciences

Computer Vision and Pattern Recognition, pp. 770–778. [Link] (2011) White paper report of the 2010 rad-aid conference on inter‑
doi.​org/​10.​1109/​CVPR.​2016.​90 national radiology for developing countries: identifying sustain‑
48. Tan M, Le Q (2019) Efficientnet: Rethinking model scaling for able strategies for imaging services in the developing world. J Am
convolutional neural networks. In: International Conference on Coll Radiol 8(8):556–562. [Link]
Machine Learning, pp. 6105–6114. [Link] 011
arXiv.​1905.​11946. PMLR 55. Fitzgerald R (2001) Error in radiology. Clin Radiol 56(12):938–
49. Kingma DP, Ba J (2014) Adam: a method for stochastic optimiza‑ 946. [Link]
tion. arXiv preprint arXiv:​1412.​6980. [Link] 56. Donovan T, Litchfield D (2013) Looking for cancer: expertise
arXiv.​1412.​6980 related differences in searching and decision making. Appl Cogn
50. Shazia A, Xuan TZ, Chuah JH, Usman J, Qian P, Lai KW (2021) Psychol 27(1):43–49. [Link]
A comparative study of multiple neural network for detection 57. Bass JC, Chiles C (1990) Visual skill. Correlation with detection
of covid-19 on chest x-ray. EURASIP J Adv Signal Process of solitary pulmonary nodules. Investig Radiol 25(9):994–998.
2021(1):1–16. [Link] [Link]
51. Hasan N, Bao Y, Shawon A, Huang Y (2021) Densenet convolu‑ 58. Carmody DP, Nodine CF, Kundel HL (1980) An analysis of per‑
tional neural networks application for predicting covid-19 using ceptual and cognitive factors in radiographic interpretation. Per‑
CT image. SN Comput Sci 2(5):389. [Link] ception 9(3):339–344. [Link]
s42979-​021-​00782-7
52. Ogundokun RO, Maskeliūnas R, Misra S, Damasevicius R (2022) Springer Nature or its licensor (e.g. a society or other partner) holds
A novel deep transfer learning approach based on depth-wise sep‑ exclusive rights to this article under a publishing agreement with the
arable CNN for human posture detection. Information 13(11):520. author(s) or other rightsholder(s); author self-archiving of the accepted
[Link] manuscript version of this article is solely governed by the terms of
53. Organization WH et al. (2001) Standardization of interpretation such publishing agreement and applicable law.
of chest radiographs for the diagnosis of pneumonia in children.
Technical report, World Health Organization. [Link] pps.w​ ho.i​ nt/​
iris/​bitst​ream/​handle/​10665/​66956/​WHO_V_​and_B_​01.​35.​pdf
54. Welling RD, Azene EM, Kalia V, Pongpirul K, Starikovsky A,
Sydnor R, Lungren MP, Johnson B, Kimble C, Wiktorek S et al

13