Source-linked AI summary
Can AI help in screening Viral and COVID-19 pneumonia?
Muhammad E. H. Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandaker Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al-Emadi, Mamun Bin Ibne Reaz, T. I. Islam
TL;DR
Rapid, accurate COVID-19 screening could help address diagnostic challenges using accessible chest X-rays. This paper evaluates transfer-learned CNNs for classifying normal, viral-pneumonia, and COVID-19 X-rays, achieving high reported performance across classification schemes.
Problem
Transfer-learning approaches for COVID-19 X-ray detection had promising results but lacked rigorous evaluation on large databases containing COVID-19 and non-COVID classes.
Method
The study trains, validates, and tests eight pre-trained CNNs using transfer learning and image augmentation on chest X-ray classification tasks.
Results
99.7% accuracy was reported for normal-versus-COVID-19 classification, while three-class normal, viral-pneumonia, and COVID-19 classification achieved 97.9% accuracy.
Takeaways & Limitations
The authors conclude that deep CNN transfer learning can automatically detect COVID-19 pneumonia from chest X-rays, with DenseNet201 performing best when augmentation was used.
Abstract
from arXiv · showhide
Coronavirus disease (COVID-19) is a pandemic disease, which has already caused thousands of causalities and infected several millions of people worldwide. Any technological tool enabling rapid screening of the COVID-19 infection with high accuracy can be crucially helpful to healthcare professionals. The main clinical tool currently in use for the diagnosis of COVID-19 is the Reverse transcription polymerase chain reaction (RT-PCR), which is expensive, less-sensitive and requires specialized medical personnel. X-ray imaging is an easily accessible tool that can be an excellent alternative in the COVID-19 diagnosis. This research was taken to investigate the utility of artificial intelligence (AI) in the rapid and accurate detection of COVID-19 from chest X-ray images. The aim of this paper is to propose a robust technique for automatic detection of COVID-19 pneumonia from digital chest X-ray images applying pre-trained deep-learning algorithms while maximizing the detection accuracy. A public database was created by the authors combining several public databases and also by collecting images from recently published articles. The database contains a mixture of 423 COVID-19, 1485 viral pneumonia, and 1579 normal chest X-ray images. Transfer learning technique was used with the help of image augmentation to train and validate several pre-trained deep Convolutional Neural Networks (CNNs). The networks were trained to classify two different schemes: i) normal and COVID-19 pneumonia; ii) normal, viral and COVID-19 pneumonia with and without image augmentation. The classification accuracy, precision, sensitivity, and specificity for both the schemes were 99.7%, 99.7%, 99.7% and 99.55% and 97.9%, 97.95%, 97.9%, and 98.8%, respectively.
I. INTRODUCTION
COVID-19’s similarity to viral pneumonia creates diagnostic risks, while existing X-ray deep-learning studies often rely on small datasets. The section motivates rigorous validation of transfer-learning approaches on larger COVID and non-COVID databases.
- Epidemiological context: COVID-19 was declared a pandemic by WHO on 11th March 2020 because of its worldwide spread and severity.The section describes COVID-19 as extremely contagious and the first recorded coronavirus-caused pandemic.
- Clinical motivation: RT-PCR is the most used clinical screening method, but similar COVID-19 and viral-pneumonia symptoms can cause incorrect diagnoses.Mislabeling non-COVID viral pneumonia may delay treatment and increase exposure risks in overloaded hospitals.
- AI-based screening: Deep-learning methods, particularly CNNs, can extract image features that are not apparent in original images.CNNs have been widely adopted for feature extraction and learning in biomedical imaging problems.
- Limitations of prior work: Recent X-ray COVID-19 studies commonly used only a few COVID-19 samples, limiting result generalizability to larger datasets.The section notes that reported performance may not be retained when models are tested on larger datasets.
- Research gap: A rigorous transfer-learning experiment using a large database of COVID and non-COVID classes remains scarce.Prior studies reported promising results, but large-database verification was still needed.
II. METHODOLOGY · A. DATABASE DESCRIPTION
The study assembled a publicly available chest X-ray database by combining six sub-databases and author-collected COVID-19 images. It used AP/PA radiographs and included normal, viral pneumonia, and COVID-19 categories for subsequent analysis.
- II. METHODOLOGY: Transfer learning was selected to train deep CNNs with a comparatively small dataset, using pre-trained weights for faster application-specific training.The pre-trained models originated from large datasets such as ImageNet, with modified Softmax and classification layers.
- II. METHODOLOGY: The authors created a comparatively large, publicly available COVID-19 chest X-ray dataset because online images were limited and scattered.A Kaggle database was created to make the dataset available to researchers worldwide.
- A. DATABASE DESCRIPTION: The database combined six sub-databases, with COVID-19 images developed from collected and public sources and normal and viral pneumonia images obtained from public Kaggle databases.The study used posterior-to-anterior or anterior-to-posterior chest X-ray views because radiologists widely use these views clinically.
- A. DATABASE DESCRIPTION: The COVID-19 sub-database comprised 423 AP/PA images collected from four major data sources.One source was the SIRM COVID-19 database, which contained 384 radiographic images, including 94 chest X-ray images and 290 lung CT images.
- A. DATABASE DESCRIPTION: The Novel Corona Virus 2019 Dataset contributed COVID-19 radiographic images collected from published articles and online resources, including chest X-ray and CT images.The database contained 319 radiographic images covering COVID-19, MERS, SARS, and ARDS, including 250 COVID-19-positive chest X-ray images.
- A. DATABASE DESCRIPTION: Additional COVID-19 chest X-ray images were collected from published articles after the authors found that existing SIRM and GitHub databases contained relatively few images.The authors reviewed literature published during a period when more than 1200 articles appeared in less than two months.
- A. DATABASE DESCRIPTION: The normal and viral pneumonia sub-databases contained 1579 and 1485 X-ray images, respectively, using RSNA and Kaggle chest X-ray resources.The RSNA database supplied normal and non-COVID pneumonia images, while the Kaggle database included normal, viral, and bacterial pneumonia cases.
B. CNN Model Selection
The study selected and evaluated eight pre-trained CNN models for COVID-19 chest X-ray classification. Experiments covered two-class and three-class classification, each with and without image augmentation.
- CNN Model Selection: Eight different pre-trained CNN models were trained, validated, and tested in the study.The models included MobileNetv2, SqueezeNet, ResNet18, ResNet101, DenseNet201, CheXNet, Inceptionv3, and VGG19.
- CNN Model Selection: Five CNN models were implemented in MATLAB 2020a, while CheXNet, Inceptionv3, and VGG19 were implemented using PyTorch.The MATLAB experiments used an Intel i7-core processor, 16GB RAM, and an 8-GB NVIDIA GeForce GTX 1080 GPU on 64-bit Windows 10.
- CNN Model Selection: The study evaluated two-class and three-class image classification with and without image augmentation.The three-class task addressed normal, viral, and COVID-19 pneumonia images.
B. PREPROCESSING
Preprocessing resized chest X-ray images to meet each CNN’s input dimensions and normalized them according to pre-trained model standards. The two studies differed in database balancing and augmentation, with both evaluated using stratified 5-fold cross-validation.
- Image preparation: Images were resized to CNN-specific dimensions and normalized according to each pre-trained model’s standards.SqueezeNet used 227×227 pixels; mobilenetv2, ResNet18, ResNet101, VGG19, and DenseNet201 used 224×224; Inceptionv3 used 299×299.
- Dataset preparation: Study1 used 423 images each from COVID-19, normal, and viral pneumonia classes without augmentation, while Study2 used the entire database.Study2 included 423 COVID-19, 1579 normal, and 1485 viral pneumonia images.
- Validation: Both experiments used stratified 5-fold cross-validation.The cross-validation scheme was applied to both study designs.
- Image augmentation: COVID-19 images were augmented six times, whereas normal and viral pneumonia images were augmented once.The augmentation imbalance applied to the image classes during preprocessing.
C. IMAGE AUGMENTATION
The study used rotation and translation to augment COVID-19 training images, while translation was also applied to viral and normal X-ray training images.
- C. IMAGE AUGMENTATION: Rotation augmented COVID-19 training images at 5°, 10°, and 15° clockwise and counterclockwise angles.Translation augmented images horizontally and vertically from −5% to 5%; it was also applied to viral and normal X-ray training images.
- C. IMAGE AUGMENTATION: Image translation shifted images horizontally and vertically by −5% to 5%.The passage specifies translation as the augmentation applied to viral and normal X-ray training images.
D. INVESTIGATION OF THE DEEP LAYER FEATURES · E. PERFORMANCE EVALUATION MATRIX
The study examined deep-layer activation patterns in chest X-ray models and evaluated their performance across two- and three-class classification schemes using five-fold cross-validation.
- D. INVESTIGATION OF THE DEEP LAYER FEATURES: Deep-layer features were investigated by comparing convolutional activation areas with corresponding regions in the original X-ray images.Activation maps were normalized between 0 and 1 before comparison.
- D. INVESTIGATION OF THE DEEP LAYER FEATURES: The strongest activation channels were identified for COVID-19, normal, and viral pneumonia X-ray images and compared with the originals.The strongest channel was observed to activate on edges with positive activation.
- D. INVESTIGATION OF THE DEEP LAYER FEATURES: Early convolutional layers detect features such as color and edges, whereas deeper layers combine earlier features into more complex representations.Figure 4 presents activation maps from early and deep convolutional layers and their strongest channels for each model.
- E. PERFORMANCE EVALUATION MATRIX: The trained deep-learning algorithms were evaluated for X-ray classification under two different classification schemes.The evaluation compared algorithm performance across the study’s two-class and three-class tasks.
- E. PERFORMANCE EVALUATION MATRIX: 5-fold cross-validation was used, with accuracy, sensitivity or recall, specificity, precision (PPV), and F1 score as evaluation metrics.Per-class values were computed from an overall confusion matrix accumulating all test-fold results.
- E. PERFORMANCE EVALUATION MATRIX: The two-class problem comprised COVID-19 and Normal, while the three-class problem comprised COVID-19, Normal, and Viral Pneumonia.These class definitions determine the classification schemes used for performance evaluation.
III. RESULTS AND DISCUSSION
The study evaluated two classification schemes using eight pre-trained CNN models, with training conducted both with and without image augmentation. The schemes distinguished COVID-19 from normal images, and classified COVID-19, normal, and viral pneumonia images.
- Two classification schemes were studied using eight different pre-trained CNN models.Both schemes were trained with and without image augmentation.
- The first scheme classified COVID-19 and normal chest X-ray images.Eight pre-trained CNN models were evaluated with and without image augmentation.
- The second scheme classified COVID-19, normal, and viral pneumonia images.The same eight pre-trained CNN models were trained with and without image augmentation.
A. EXPERIMENTAL RESULTS – TWO CLASS PROBLEM
Pre-trained CNNs performed very well in distinguishing COVID-19 from normal chest X-ray images in the two-class classification problem. Performance was comparable across models, with only a small improvement from image augmentation.
- A. EXPERIMENTAL RESULTS – TWO CLASS PROBLEM: All evaluated pre-trained CNNs performed very well in classifying COVID-19 and normal chest X-ray images.The comparison considered two-class classification with and without image augmentation.
- A. EXPERIMENTAL RESULTS – TWO CLASS PROBLEM: The weighted average performance metrics of eight networks were very similar, with a small gain from image augmentation.The passage attributes this gain to training with augmented images.
- A. EXPERIMENTAL RESULTS – TWO CLASS PROBLEM: ROC curves showed comparable performance across all tested networks with and without image augmentation.This comparable performance was observed for the two-class problem involving normal and COVID-19 pneumonia images.
B. EXPERIMENTAL RESULTS – THREE CLASS
In the three-class task, pre-trained CNNs performed similarly overall, but performance declined relative to two-class classification as models distinguished COVID-19 from viral pneumonia. DenseNet201 with augmentation showed strong discrimination, while errors arose when COVID-19 lacked distinctive radiographic features.
- Three-class performance: Pre-trained CNNs showed similar overall performance, with three-class results generally lower than two-class results because COVID-19 and viral pneumonia were harder to distinguish.CheXNet was an exception during training without image augmentation and continued to perform well on the small dataset.
- Three-class performance: ResNet18 outperformed the deeper ResNet101, whereas DenseNet201 outperformed CheXNet despite CheXNet’s original X-ray training.The results indicate that deeper networks trained on larger image sets can learn and perform better on new image collections.
- Three-class performance: DenseNet201 with image augmentation significantly increased overall network performance in normal, COVID-19, and viral pneumonia classification.Figure 6 compares ROC curves without and with augmentation for the three-class task.
- Confusion-matrix results: Only four viral pneumonia images were misclassified as COVID-19 and 33 as normal out of 1485 images in the DenseNet201 three-class confusion matrix.Several normal images were misclassified as viral pneumonia, while COVID-19 misclassification to normal was identified as more consequential.
- Feature analysis and limitations: DenseNet201’s 14th convolutional layer detected features distinguishing all three classes, but COVID-19 images without evident light-focus edges could be misclassified as normal.The reported failure cases had low opacity in upper-lobe and suprahilar regions resembling normal X-rays.
IV. CONCLUSION
The work presents a deep CNN transfer-learning approach for automatically detecting COVID-19 pneumonia from chest X-ray images. It highlights DenseNet201’s performance with image augmentation and the potential of computer-aided diagnosis for early screening.
- Approach: The study presents deep CNN-based transfer learning for automatic COVID-19 pneumonia detection using chest X-ray images.Eight previously reported CNN algorithms were trained, validated, and tested to classify normal and pneumonia patients.
- Model performance: DenseNet201 outperformed the other deep CNN networks when image augmentation was used during training.The conclusion identifies DenseNet201 as the strongest-performing network among the evaluated models.
- Clinical motivation: Computer-aided diagnosis may support early screening and proper care when limited clinician time affects patient management.The motivation is the large number of patients attending outpatient or emergency services during the COVID-19 threat.
- Clinical motivation: Variability in chest X-ray images from different machines and radiologist expertise underscores the need for computer-aided diagnostic support.The passage notes substantial variation in the input images attributable to X-ray equipment and radiologist expertise.
AUTHORS CONTRIBUTION
The authors contributed across conceptualization, data, methodology, analysis, software, writing, supervision, and project administration. Publication funding came from the Qatar National Library and Qatar National Research Fund support.
- Muhammad E. H. Chowdhury led conceptualization, review and editing, supervision, and project administration.
- Tawsifur Rahman contributed data curation, methodology, software, validation, formal analysis, and writing-review and editing.
- Amith Khandakar contributed data curation, investigation, resources, and original-draft, review, and editing work.
- Rashid Mazhar contributed original-draft writing, review, and editing, while Muhammad Abdul Kadir contributed methodology, visualization, and editing.
- The Qatar National Library funded publication, and NPRP12S-0227-190164 from the Qatar National Research Fund supported the work.The statements made herein are solely the responsibility of the authors.