AI-driven deep convolutional neural networks for chest X-ray pathology identification

Abstract

BACKGROUND:

Chest X-ray images are widely used to detect many different lung diseases. However, reading chest X-ray images to accurately detect and classify different lung diseases by doctors is often difficult with large inter-reader variability. Thus, there is a huge demand for developing computer-aided automated schemes of chest X-ray images to help doctors more accurately and efficiently detect lung diseases depicting on chest X-ray images.

OBJECTIVE:

To develop convolution neural network (CNN) based deep learning models and compare their feasibility and performance to classify 14 chest diseases or pathology patterns based on chest X-rays.

METHOD:

Several CNN models pre-trained using ImageNet dataset are modified as transfer learning models and applied to classify between 14 different chest pathology and normal chest patterns depicting on chest X-ray images. In this process, a deep convolution generative adversarial network (DC-GAN) is also trained to mitigate the effects of small or imbalanced dataset and generate synthetic images to balance the dataset of different diseases. The classification models are trained and tested using a large dataset involving 91,324 frontal-view chest X-ray images.

RESULTS:

In this study, eight models are trained and compared. Among them, ResNet-152 model achieves an accuracy of 67% and 62% with and without data augmentation, respectively. Inception-V3, NasNetLarge, Xcaption, ResNet-50 and InceptionResNetV2 achieve accuracy of 68%, 62%, 66%, 66% and 54% respectively. Additionally, Resnet-152 with data augmentation achieves an accuracy of 83% but only for six classes.

CONCLUSION:

This study solves the problem of having fewer data by using GAN-based techniques to add synthetic images and demonstrates the feasibility of applying transfer learning CNN method to help classify 14 types of chest diseases depicting on chest X-ray images.

Keywords

Convolution neural network (CNN)deep learning chest diseases chest X-ray images radiographic findings ResNet-152 inception-V3

1 Introduction

Chest-related diseases will become a significant concern in the coming years due to an increase in air pollution, which causes the development of these diseases [1]. These diseases do not affect only a single organ but can cause problems in other organs, as some of them will cause oxygen levels to drop. Most convenient and the first method to detect these diseases is through X-ray images, which are read and interpreted by radiologists. With the increase in the number of patients in the medical system, they can easily become overwhelmed. This burden can be eased by automating the process through the latest image processing and pattern recognition tools. There has been ongoing research that tries to solve this problem by using computer-aided detection or diagnosis schemes using deep learning neural networks aiming to help radiologists in reading and interpreting chest X-ray images [2, 3].

A neural network tries to mimic the inner workings of the human brain toward a specific task. Just as a human brain, it consists of neurons that are modeled based on mathematical equations. These equations have parameters in which the neural network tries to approximate to make the correct prediction. The neural network trains a set of algorithms on existing data to perform clustering or classification tasks [4, 5].

Imbalancing of dataset is a common problem when training a neural network. Data imbalancing causes a model to be biased toward classes with majority data., thus affecting the accuracy of the model. While reducing the data of each class to the size of class with minimum data seems like a sound remedy, this process causes another problem. However, scholars have discovered that CNN models used for classification requires huge amounts of data. CNN models train for large sample sizes and return high levels of accuracy. To solve the problem of fewer data, new data is generated through the deep convolutional generative adversarial network (DCGAN), which synthesizes images. GANs are the networks that generate new images by learning from previously fed images.

This research solves the problem of chest pathology classification via convolutional neural networks (CNNs). Models used were not built from scratch, as the researchers utilized built-in models due to computational restraints. This not only saves time and computational power but also works better when the dataset available is small. Pre-trained models used in this study were NasNetLarge, Xception, InceptionV3, InceptionResNetV2, ResNet50 and ResNet-152. Pre-trained models are trained on greater data, which is specifically related to the problem being solved, i.e., the ResNet152 model was first trained on the ImageNet dataset. To further enhance the models to train for chest x-rays, we modified classifications according to our needs. Thus, the primary contribution of this research includes following:

Fine-tuning pre-trained models by adding more layers for the classification of 14 chest diseases and normal chest x-ray images.

Mitigating the effect of the imbalanced dataset by generating x-ray images using DCGAN.

Comparing different pre-trained models and the different number of classes to get the best model for lungs disease detection using X-rays.

The rest of the paper is arranged into five sequential sections. A brief literature review of works in the area of chest disease detection is presented in Section 2. Next, Section 3 explains the dataset used for the current research. In addition to stating the parameter values, Section 4 discusses the models used for the experiments carried out by the investigators. After the fourth section, Section 5 analyzes the study outcomes. Finally, concluding thoughts and points are presented in Section 6.

2 Related work

There is a rich body of literature regarding chest disease detection within the medical field. Maintaining an impressive accuracy of 86.14%, Sivasamy et al. [4] predicted chest diseases using chest X-rays using the Keras framework. Rajpurkar et al. [6] detected pneumonia from frontal-view X-rays using binary classification. Salehinejad et al. [7] used a deep convolutional generative adversarial neural network (DCGAN) to generate artificial chest x-rays. Irvin et al. [8] identified 14 classes of chest diseases using different neural network models. Apostolopoulos et al. [9] used five different pre-trained models to detect COVID-19 from chest X-rays. Pardamean et al. [10] applied transfer learning on mammogram data for breast cancer detection. Bhandary et al. [11] worked on detecting lung abnormalities from chest X-rays and CT scans using pre-trained models. Wang et al. [12] worked with localization of thorax disease using a deep convolutional neural network.

In addition, several researchers also analyzed the segmentation of lungs and other organs for better classification of chest diseases. For example, Wang et al.’s [13] work on segmenting the ribs from chest X-rays helped classify chest diseases. Wang et al. [14] yielded better feature extraction for segmentation by proposing a unified end-to-end trainable deep neural network. Zhao et al. [15] localized organs and then segmented them using the voxel-wise label map. Using this method, Zhao and colleagues concluded segmentation models work good on large organs but mostly fail in classifying small organs.

Overall, previous research obtained state-of-the-art results by using promising techniques with only a few chest diseases or diseases having data present in large quantities. Table 1 provides an overview of the literature by detailing utilized methodology, obtained findings, and identified scholarly gaps for each of the referenced articles. This table also provides insight into the scholarly gap which the current study is attempting to address. As previously stated, the main purpose of this study is to propose a new classification procedure to identify chest diseases based on chest X-rays. Furthermore, this study uses GAN to mitigate the effects of less and imbalanced data, implementing and tuning pre-trained models to get optimal results and best accuracy.

Table 1
A comparison of the chest disease diagnoses

Reference Methodology Findings Gaps identified

Sivasamy, et al. [4] Classifying the 10 chest X-ray diseases using Keras on NIH dataset Training accuracy = 86.14% Validation accuracy = 85.62%. Accuracy increases with an increasing number of epochs

Radford, et al. [16] Unsupervised multi-organ segmentation on pixel-wise annotated DRRs X-ray data Segmented images without ground truth annotation Nodule annotations directly on 2D x-ray images are computationally expensive

Rajpurkar, et al. [6] Detection of pneumonia from frontal-view of X-rays F1 score = 0.435 (95% CI 0.387, 0.481) CheXNet showed higher performance than radiologists

Salehinejad, et al. [7] Generation of chest X-rays using DCGAN 70.87% accuracy for DS1, 58.90% for DS2, and 92.10% for DS3. Balanced, imbalanced data by synthesizing chest X-ray images

Irvin, et al. [8] Classified 14 chest X-ray diseases Obtained the best accuracy with DenseNet121 The models are limited to the CheXpert database and liable to over-fitting

Apostolopoulos and Mpesiana [9] Classified pneumonia, COVID-19, and normal chest conditions Accuracy, sensitivity, and specificity obtained are 96.78%, 98.66%, and 96.46%, respectively. Fast detection of COVID-19 to mitigate the damage to the lungs.

Pardamean, et al. [10] Breast cancer detection using chest X-ray data Maximum accuracy of 98.02 Achieving the best accuracy with less data

Bhandary, et al. [11] Detecting lung abnormalities using CXRs and CT-scan Maximum accuracy using extracted features = 96.8%; Maximum accuracy using handcrafted features = 96.8% Comparative analysis of different techniques to identify the best possible classification

Wang et al. [12] Classification and localization of thorax disease using DCNN ResNet50 achieved the best accuracy Segmentation of diseases in X-ray

Reference	Methodology	Findings	Gaps identified
Sivasamy, et al. [4]	Classifying the 10 chest X-ray diseases using Keras on NIH dataset	Training accuracy = 86.14% Validation accuracy = 85.62%.	Accuracy increases with an increasing number of epochs
Radford, et al. [16]	Unsupervised multi-organ segmentation on pixel-wise annotated DRRs X-ray data	Segmented images without ground truth annotation	Nodule annotations directly on 2D x-ray images are computationally expensive
Rajpurkar, et al. [6]	Detection of pneumonia from frontal-view of X-rays	F1 score = 0.435 (95% CI 0.387, 0.481)	CheXNet showed higher performance than radiologists
Salehinejad, et al. [7]	Generation of chest X-rays using DCGAN	70.87% accuracy for DS1, 58.90% for DS2, and 92.10% for DS3.	Balanced, imbalanced data by synthesizing chest X-ray images
Irvin, et al. [8]	Classified 14 chest X-ray diseases	Obtained the best accuracy with DenseNet121	The models are limited to the CheXpert database and liable to over-fitting
Apostolopoulos and Mpesiana [9]	Classified pneumonia, COVID-19, and normal chest conditions	Accuracy, sensitivity, and specificity obtained are 96.78%, 98.66%, and 96.46%, respectively.	Fast detection of COVID-19 to mitigate the damage to the lungs.
Pardamean, et al. [10]	Breast cancer detection using chest X-ray data	Maximum accuracy of 98.02	Achieving the best accuracy with less data
Bhandary, et al. [11]	Detecting lung abnormalities using CXRs and CT-scan	Maximum accuracy using extracted features = 96.8%; Maximum accuracy using handcrafted features = 96.8%	Comparative analysis of different techniques to identify the best possible classification
Wang et al. [12]	Classification and localization of thorax disease using DCNN	ResNet50 achieved the best accuracy	Segmentation of diseases in X-ray

3 Materials and methods

A deep learning model requires a vast amount of data for training. The effectiveness of deep learning models enhances with large data sets. Having few data points or imbalanced data can cause the model to under-fit or overfit in specific classes. The data set used in this research was taken from NIH [17], which consisted of 14 underlined diseases related to the chest. As shown in Fig. 1, the NIH dataset contained 91,324 frontal-view CXRs images across 32,717 patients. These images included X-rays with no disease as well as some with disease detection. As presented in Table 2, majority of the data belonged to six classes of chest-related (Thoracic) diseases. A more complete breakdown of the NIH dataset is shown in Table 3, which provides a distribution of the data across all 14 chest-related disease. Additionally, the NIH dataset analyzed in this study included samples across training, testing, and validation in ratios of 70%, 20%, and 10% respectively.

Fig. 1

A bar chart showing 14 underlined diseases related to the chest addressed by this study.

Table 2

Six classes of chest related diseases in the dataset used

Thoracic Diseases	No. of Cases	Training samples	Testing samples	Validation samples
Infiltration	9,547	6,683	1,909	955
Effusion	3,955	2,768	791	396
Atelectasis	4,215	2,950	843	422
Nodule	2,705	1,893	541	271
Mass	2,139	1,497	427	215
Pneumothorax	2,194	1,535	438	221

Table 3

14 classes of chest related diseases in the dataset

Thoracic diseases	No. of Cases	Training samples	Testing samples	Validation samples
Infiltration	9,547	6,683	1,909	955
Effusion	3,955	2,768	791	396
Atelectasis	4,215	2,950	843	422
Nodule	2,705	1,893	541	271
Mass	2,139	1,497	427	215
Pneumothorax	2,194	1,535	438	221
Cardiomegaly	1010	707	202	101
Consolidation	272	190	54	28
Edema	241	169	48	24
Emphysema	142	99	29	14
Fibrosis	138	99	28	13
Hernia	105	73	21	10
Pleural Thickening	165	115	33	16
Pneumonia	140	98	28	14

Images having no disease are the majority (n = 60,361) of the data. Therefore, a significant imbalance in the data for proper training exists, as this might cause the model to over-fit or under-fit for the specific class of “no disease.” This problem was resolved by using additional data from the Kaggle challenge [18] and data generated through image augmentation.

To analyze the data, distributions on the basis of age and gender were done for all chest diseases, which can be seen in Fig. 2. Normal images have more frequency than images having conditions. Among disease case count, infiltration had the maximum frequency of cases (n = 9,547) whereas hernia had the minimum frequency (n = 105).

Fig. 2

Gender and age distribution of chest related disease.

Give the observed examination that the NIH dataset was highly imbalanced, we deduced that:

Infiltration, Effusion, Atelectasis, Nodule, Consolidation, Pleural Thickening, Emphysema, and Hernia affected more males than females.

Pneumothorax, Cardiomegaly, Edema, and Fibrosis affected more females than males.

Cardiomegaly starts affecting people from age ten and is common among people aged 35 and 65, with a median age of 50.

Effusion has the same trend as cardiomegaly but has a median age of 55.

Atelectasis affects people ages 25 years old or older. There is only one case in the data which appeared to be an outlier.

Mass affects people from a younger age, and the number of cases increases with increasing age.

Emphysema primarily affects people ages 40 or older. Cases of individuals under age 40 can be outliers.

The available dataset is highly imbalanced as individuals having no chest disease are at 60,361.

Above deductions make the training of the NN model very difficult, as it can cause over-fitting. To mitigate this effect, DCGAN was used to generate images.

The NIH dataset was divided into training and testing. Each of these two sample types then contained 15 folders regarding types of X-rays, one for each chest-related disease and one for normal images. Validation data was extracted during the model training process. Before using the images, the X-rays were pre-processed by reshaping them into 150, 150, and 3 using nearest neighbor interpolation and then normalizing. To better train the model, data augmentation was applied and images were rotated. Parameters for data augmentation were as follows: rotation range = 40, width shift range = 0.2, height shift range = 40, shear range = 0.2, zoom range = 0.2, horizontal flip = true, and fill mode = nearest. This data augmentation was applied during training on both the already existing and generated images.

3.1 DCGAN

GANs are the neural networks that generate new images by learning from previously given images. They consist of two parts –(a) a convolutional network (discriminator) that extracts the learned features from the images, and (b) a deconvolutional network (generator) that tries to generate images on basis of features learned through the convolutional network. The GAN creates new images by analyzing the error generated when comparing the original images to the images generated by the deconvolutional network. Figure 3 shows the basic working diagram of GANs.

Fig. 3

Basic architecture of Generative Adversarial Network (GAN) model.

DCGAN was proposed by Radford et al. [16]. A summary of the DCGAN architecture used is shown in Fig. 4. Although the same architecture was used as in the original paper, the parameters were tuned using GAN hacks [19, 20]. During the training, Batch size was taken 64 with hyperparameters for the generator (α= 0.0002, β1 = 0.5) and for the discriminator (α=0.00015, β1 = 0.5). Here, α represents the learning rate and β1 symbolizes the exponential decay rate for the 1st-moment estimates. The different numbers of images were generated for different classes to balance the dataset. The number of images used for each class was 5,000. Consequently, the number of images generated was different for some classes. For example, the Infiltration class had 9,547 images; therefore, only 5,000 images were used and extra images were not used. On the other hand, there were 3,955 images present for the Effusion class, which resulted in 1,045 generated images. This was done because we were precautious in using more images generated by DCGAN. Although the images were similar to the original ones, there was still room for error. This dataset was rationally divided further into 3 sets: 70% for training, 20% for testing, and 10% for testing.

Fig. 4

Summary of discriminator model (Up), Summary of generator model (Down).

Tables 2 3 show the distribution of all X-ray images across the six classes and all 14 classes before data augmentation respectively. While Tables 4 5 show the class distribution after data augmentation.

Table 4

Six classes of chest related diseases in the dataset used after data augmentation

Thoracic diseases	Training samples	Testing testing	Validation samples
Infiltration	3,500	1,000	500
Effusion	3,500	1,000	500
Atelectasis	3,500	1,000	500
Nodule	3,500	1,000	500
Mass	3,500	1,000	500
Pneumothorax	3,500	1,000	500
Total	21,000	6,000	3000

Table 5

14 classes of chest related diseases in the dataset used after data augmentation

Thoracic diseases	Training samples	Testing samples	Validation samples
Infiltration	3,500	1,000	500
Effusion	3,500	1,000	500
Atelectasis	3,500	1,000	500
Nodule	3,500	1,000	500
Mass	3,500	1,000	500
Pneumothorax	3,500	1,000	500
Cardiomegaly	3,500	1,000	500
Consolidation	3,500	1,000	500
Edema	3,500	1,000	500
Emphysema	3,500	1,000	500
Fibrosis	3,500	1,000	500
Hernia	3,500	1,000	500
Pleural Thickening	3,500	1,000	500
Pneumonia	3,500	1,000	500
Total	49,000	14,000	7,000

3.2 Pre-trained models

Pre-trained models were previously trained on other dataset, which gave the advantage of less training time when the same model is trained on some other dataset. Pre-trained models used by us were NasNetLarge, Xception, InceptionV3, InceptionResNetV2, ResNet50 and ResNet-152. These models were pre-trained on the ImageNet dataset. This dataset classifies 1,000 classes. These models were used because mostly it is said that the deeper model works better. This was true for AlexNet and VGG, however, the deeper networks were difficult to optimize due to a gradient exploding/vanishing problem. The models used in this research were based on architectural engineering. In the case of ResNet, the model t employed residual learning and input from previous layers that were used in preceding layers. InceptionNet applied filters of different sizes on output from previous layers and feeds it to next layers, thus making the network deeper and wider. In the end, NasNet worked as an AutoML network that found the best CNN architecture for the problem at hand.

These models were modified and fine-tuned by adding additional layers to meet the need of new NIH dataset. Table 6 shows the additional layers added to pre-existing layers in the pre-trained neural networks.

Table 6
Six classes of chest related diseases in the dataset used

Type Output Kernel

Average Pooling 2048 2×2

Flatten 8192 –

Dense 1024 –

Dropout (0.5) 1024 –

Dense 1024 –

Dropout (0.5) 1024 –

Dense 15 –

Type	Output	Kernel
Average Pooling	2048	2×2
Flatten	8192	–
Dense	1024	–
Dropout (0.5)	1024	–
Dense	1024	–
Dropout (0.5)	1024	–
Dense	15	–

4 Experiments and results

For the sake of classification, models used were ResNet-152, Inception-V3, Inception ResNet V2, Xception, NasNetLarge and ResNet-50. Pre-trained modules of these models have shown promising results. We trained these models on the collected and generated data for the purpose of CRXs classification. In all, each model was trained for 50 epochs. Table 6 shows the additional layers added to the already existing, and original models’ different values of hyper-parameters were tested for each model. The best parameter values were chosen, which produced the best accuracy. The best hyper-parameter values of each model used are shown in Table 7.

Table 7
The proposed models and comparisons of their results

Models Augmentation? Accuracy Epochs = 50 Activation Total trainable parameters Optimizer

ResNet-152 Yes 0.67 ReLU 3,479,631 Adam

ResNet-152 No 0.62 ReLU 3,657,221 Adam

Transfer learning (Inception-V3) – 0.68 ReLU 38,551,567 RMSprop

NasNetLarge Yes 0.62 ReLU 17,580,046 Adam

Xception Yes 0.66 ReLU 9,453,582 Adam

ResNet-50 Yes 0.66 ReLU 9,453,582 Adam

InceptionResNetV2 Yes 0.54 ReLU 7,356,430 Adam

ResNet-152 with 6 targeted classes Yes 0.83 ReLU 38,543,367 RMSprop

Models	Augmentation?	Accuracy Epochs = 50	Activation	Total trainable parameters	Optimizer
ResNet-152	Yes	0.67	ReLU	3,479,631	Adam
ResNet-152	No	0.62	ReLU	3,657,221	Adam
Transfer learning (Inception-V3)	–	0.68	ReLU	38,551,567	RMSprop
NasNetLarge	Yes	0.62	ReLU	17,580,046	Adam
Xception	Yes	0.66	ReLU	9,453,582	Adam
ResNet-50	Yes	0.66	ReLU	9,453,582	Adam
InceptionResNetV2	Yes	0.54	ReLU	7,356,430	Adam
ResNet-152 with 6 targeted classes	Yes	0.83	ReLU	38,543,367	RMSprop

In the first model, the Adam optimizer was used with leaky-ReLU as the activation function. The model was trained for 14 classes of chest-related diseases and achieved a validation accuracy of 67%. Using the same ResNet-152 model, the next model did not apply time augmentation to the data. Instead, a 512 layered model was used with ReLU as activation function and Softmax as the activation function of the output layer with 14 classes. For fine-tuning, an extra batch normalization layer was added after every convolution layer and dense layer. This model achieved an accuracy of 62%, which is 5% less than the previous model.The following five models were transfer learning models: Inception V3, Inception ResNet-V2, Xception, NasNet large, and ResNet-50. Inception-V3 model had 1,024 fully connected layers and the ReLU activation function. RMSprop was used as an activation function and achieved 68% accuracy. The next four models were trained using ReLU as an activation function along with Adam optimizer. For the purpose of fine-tuning, an additional layer of average pooling was added, followed by a flattening layer. After that, two dense layers were added, each of which was followed by one layer of dropout (0.5). In the end, a dense layer was used as an output layer. The accuracy achieved by each of these five models is shown in Table 7.

The Inception-V3 model achieved maximum accuracy among all these models, although it was not a significant improvement. The last model, Also the ResNet-152 model, the last model only utilized six classes. This model also used image augmentation with the parameters mentioned above. This model was trained to classify only six classes, so it had an advantage over other models. In this instance, theResNet-152 achieved an accuracy of 83%, which was a significant improvement compared to other models. This showed that reducing the class variable drastically increases the model accuracy. Therefore, using a multistep model would be a better solution for classification of chest-related diseases in X-ray images and should be considered doing in the future.

5 Discussion

Eight models were trained for automatic detection of chest diseases using X-ray images. After obtaining the results, we deduced that ResNet-152 trained on six classes with image augmentation had significant data available and produced promising accuracy of 0.83, then models trained on 14 classes, including classes having less available data with and without augmentation. The accuracy achieved in these cases was 67% and 62%, respectively. The ResNet-152 produced a training accuracy of 99% with a low validation accuracy of only 62%. These results indicate that the model was over-fitting. ResNet-152 with six classes showed promising results, indicating that balanced training data produces better results. The model trained with augmented images produced better accuracy than models trained on non-augmented data. The results showed low accuracy because the dataset used was still not adequate as 5,000 images per class was used. We hypothesize that increasing data points will increase the accuracy of model classification. It was also observed that pre-trained models helped reduce the computational cost and gave better accuracy.

It is important to acknowledge that we conducted a similar experiment to Sivasamy [4]. However, whereas the difference Sivasamy [4] used 10 classes, our study utilized 14 classes. Despite using 10 classes, Sivasamy [4] yielded lower results than our findings. Sivasamy [4] achieved a maximum AUC value of 0.74 for edema, while a minimum AUC of 0.52 was achieved with Edema. In addition to producing high results, we also showed that we can increase model accuracy by decreasing the number of classes to train on original images.

In summary, this research attempted to solve the problem of early chest disease detection using chest X-rays using advanced techniques in image processing and pattern recognition. This medical approach will not only increase the speed of detection but will also reduce the overall effort (i.e., time and energy) of radiologists. Using deep learning as the detection tool, this research explored the efficiency of different models in chest disease detection and under which conditions a model can work efficiently.

After experimentation, it is deduced that well-balanced data can produce better results. In our case, this is elaborated when ResNet-152 gave better accuracy with a smaller number of classes having an almost equal number of classes. Thus, it is also observed that data augmentation increases the accuracy of a model. For example, ResNet152 produces better results with data augmentation. As a result, the availability of fewer data negatively effects the model accuracy in a severe manner.

Finally, this study also indicates that a lot of research work can be done to detect and localize diseases in X-rays in the future. Models can be observed over time as more data becomes available. More classes can be added, and models can be fine-tuned accordingly to get exceptional performance.

Footnotes

Acknowledgments

The researchers would like to thank the Deanship of Scientific Research, Qassim University for funding the publication of this project.

Conflicts of interest

The authors declare that they have no conflicts of interest to report regarding the present study.

Availability of Data and Materials

The experiment uses a public dataset shared by 2019 Kaggle: Digital pathology classification challenge, 2019., and the Covid-19 images from the paper: Joseph Paul Cohen and Paul Morrison and Lan Dao, COVID-19 image data collection, arXiv:2003.11597, 2020 ().

References

Goodfellow

, Bengio

and Courville

, Deep learning. MIT Press, USA, 2016. [Online] https://mitpress.mit.edu/books/deep-learning.

Nijiati

, Zhang

, Abulizi

, et al., Deep learning assistance for tuberculosis diagnosis with chest radiography in low-resource settings, Journal of X-ray Science and Technology 29(5) (2021), 785–796.

Albahli

, Ayub

and Shiraz

, Coronavirus disease (COVID-19) detection using X-ray images and enhanced DenseNet, Applied Soft Computing 110(10) (2021), 210–221.

Sivasamy

and Subashini

T.S.

, Classification and predictions of lung diseases from chest X-rays using MobileNet, The International Journal of Analytical and Experimental Modal Analysis 12(3) (2020), 665–672.

Mastouri

, Khlifa

, Neji

and Hantous-Zannad

, Deep learning-based CAD schemes for the detection and classification of lung nodules from CT images: A survey, Journal of X-ray Science and Technology 28(4) (2020), 591–617.

Rajpurkar

, Irvin

, Zhu

, et al., Chexnet: Radiologist-levelpneumonia detection on chest X-rays with deep learning, ArXivpreprint arXiv 17(11) (2017), 522–535.

Salehinejad

, Valaee

, Dowdell

, et al., Generalization of deep neural networks for chest pathology classification in x-rays using generative adversarial networks, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Calgary, AB, Canada, (2018), 990–994.

Irvin

, Rajpurkar

, Ko

, et al., A large chest radiograph dataset with uncertainty labels and expert comparison, The AAAI Conference on Artificial Intelligence, Honolulu, Hawaii, USA 33(2) (2019), 590–597.

Apostolopoulos

and Mpesiana

, Covid-19: automatic detection from X-ray images utilizing transfer learning with convolutional neural networks, Physical and Engineering Sciences in Medicine 43(2) (2020), 635–640.

10.

Pardamean

, Cenggoro

, Rahutomo

, et al., Transfer learning from chest X-ray pre-trained convolutional neural network for learning mammogram data, Procedia Computer Science 135(3) (2018), 400–407.

11.

Bhandary

, Prabhu

, Rajinikanth

, et al., Deep-learning framework to detect lung abnormality–A study with chest X-Ray and lung CT scan images, Pattern Recognition Letters 129(1) (2020), 271–278.

12.

Wang

, Peng

, Lu

, et al., Chest X-ray 8: Hospital-scale chest X-ray database and benchmarks on weakly supervised classification and localization of common thorax diseases, in Proc. The IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, (2017), 2097–2106.

13.

Wang

, Zhang

, Ding

, et al., Rib segmentation algorithm for X-ray image based on unpaired sample augmentation and multi-scale network, Neural Computing and Applications, (2021), 1–15.

14.

Wang

, Zhang

, Fang

, et al., A unified two-parallel-branch deep neural network for joint gland contour and segmentation learning, Future Generation Computer Systems 110(10) (2019), 316–324.

15.

Zhao

, Li

, Wan

, et al., Knowledge-aided convolutional neural network for small organ segmentation, IEEE Journal of Biomedical and Health Informatics 23(4) (2019), 1363–1373.

16.

Radford

, Metz

and Chintala

, Unsupervised representation learning with deep convolutional generative adversarial networks, in Proc. The 4th International Conference on Learning Representations, San Juan, Puerto Rico, (2017), 1–16.

17.

NIH Chest X-ray Dataset, [Online: https://www.kaggle.com/nih-chest-xrays/data.

18.

Kaggle, Digital Pathology Classification Challenge, [Online]. Available: https://www.kaggle.com/c/digitalpathology/data.

19.

Soumith

M. A. M. M.

and Den-ton

, How to train a GAN? Tips and tricks to make GANs work, on GitHub, 2017. [Online]. https://github.com/soumith/ganhacks.

20.

Albahli

, Efficient GAN-based Chest Radiographs (CXR) augmentation to diagnose coronavirus disease pneumonia, International Journal of Medical Sciences 17(10) (2020), 1439–1448.

AI-driven deep convolutional neural networks for chest X-ray pathology identification

Abstract

BACKGROUND:

OBJECTIVE:

METHOD:

RESULTS:

CONCLUSION:

Keywords

1 Introduction

2 Related work

Table 6 Six classes of chest related diseases in the dataset used Type Output Kernel Average Pooling 2048 2×2 Flatten 8192 – Dense 1024 – Dropout (0.5) 1024 – Dense 1024 – Dropout (0.5) 1024 – Dense 15 –

Footnotes

Acknowledgments

Conflicts of interest

Availability of Data and Materials

References

Table 6
Six classes of chest related diseases in the dataset used

Type Output Kernel

Average Pooling 2048 2×2

Flatten 8192 –

Dense 1024 –

Dropout (0.5) 1024 –

Dense 1024 –

Dropout (0.5) 1024 –

Dense 15 –