Explainable COVID-19 Detection Based on Chest X-rays Using an End-to-End RegNet Architecture

Abstract

COVID-19,which is caused by the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), is one of the worst pandemics in recent history. The identification of patients suspected to be infected with COVID-19 is becoming crucial to reduce its spread. We aimed to validate and test a deep learning model to detect COVID-19 based on chest X-rays. The recent deep convolutional neural network (CNN) RegNetX032 was adapted for detecting COVID-19 from chest X-ray (CXR) images using polymerase chain reaction (RT-PCR) as a reference. The model was customized and trained on five datasets containing more than 15,000 CXR images (including 4148COVID-19-positive cases) and then tested on 321 images (150 COVID-19-positive) from Montfort Hospital. Twenty percent of the data from the five datasets were used as validation data for hyperparameter optimization. Each CXR image was processed by the model to detect COVID-19. Multi-binary classifications were proposed, such as: COVID-19 vs. normal, COVID-19 + pneumonia vs. normal, and pneumonia vs. normal. The performance results were based on the area under the curve (AUC), sensitivity, and specificity. In addition, an explainability model was developed that demonstrated the high performance and high generalization degree of the proposed model in detecting and highlighting the signs of the disease. The fine-tuned RegNetX032 model achieved an overall accuracy score of 96.0%, with an AUC score of 99.1%. The model showed a superior sensitivity of 98.0% in detecting signs from CXR images of COVID-19 patients, and a specificity of 93.0% in detecting healthy CXR images. A second scenario compared COVID-19 + pneumonia vs. normal (healthy X-ray) patients. The model achieved an overall score of 99.1% (AUC) with a sensitivity of 96.0% and specificity of 93.0% on the Montfort dataset. For the validation set, the model achieved an average accuracy of 98.6%, an AUC score of 98.0%, a sensitivity of 98.0%, and a specificity of 96.0% for detection (COVID-19 patients vs. healthy patients). The second scenario compared COVID-19 + pneumonia vs. normal patients. The model achieved an overall score of 98.8% (AUC) with a sensitivity of 97.0% and a specificity of 96.0%. This robust deep learning model demonstrated excellent performance in detecting COVID-19 from chest X-rays. This model could be used to automate the detection of COVID-19 and improve decision making for patient triage and isolation in hospital settings. This could also be used as a complementary aid for radiologists or clinicians when differentiating to make smart decisions.

Reinvestigating the performance of artificial intelligence classification algorithms on COVID-19 X-Ray and CT images

“… 20 0.991 0.960 X-rays Chetoui et al. 21 0.992 0.996 X-rays Ghose et al. 22 1.000 0.996 X-rays Siddhartha et al.…”

“… 17 , 18 However, in the specific context of COVID-19, the comprehensive validation of the bias in AI diagnostic classification algorithms is less explored. Furthermore, we found that most of the datasets used to train AI classifiers were heterologous datasets, and the accuracy of the classifiers was generally high, 19 , 20 , 21 , 22 , 23 , 24 while, the generalization ability of the model trained on such heterogeneous datasets still needs to be validated. We need to look at a series of issues that have been overlooked in existing AI model articles in the field of COVID-19, such as underdiagnosis bias, model generalization ability, and AI universality.…”

Predicting acute pancreatitis severity with enhanced computed tomography scans using convolutional neural networks

“…One of the key advantages of deep learning models is their ability to automatically extract quantitative features from high-throughput images, analyze image data in-depth, and translate microscopic lesion changes into quantitative measures 41 . In this study, referred to some previous CNN model research 42 – 44 , we adopted an end-to-end approach that bypasses ROI extraction and directly employs the entire CT scan for model development.…”