A Multi-stage framework for COVID-19 detection and severity assessment from chest radiography images using advanced fuzzy ensemble technique

Abstract

Chest radiography (CXR) is a popular imaging modality for screening lung abnormalities. It plays a vital role in monitoring disease progression, allocating limited medical resources effectively, and planning treatment in clinical practice. However, manual CXR image screening is time-consuming and susceptible to human errors. Therefore, computer-aided diagnosis (CAD) systems can be an additional decision-making tool for radiologists. Although prior research had remarkable success in COVID-19 classification tasks, it is still difficult to accurately localize the infection and quantify its severity, which restricts the generalizability and interpretability of CAD-based systems. This work proposes an end-to-end multi-stage architecture for COVID-19 classification, infection region identification, and disease severity assessment. The first stage of the proposed architecture identifies the relevant lung regions by eliminating the irrelevant portions. The second stage discriminates COVID-19 from pneumonia and normal images. In this study, instead of a single pre-trained model, we used three models and a fuzzy rank-based ensemble method to combine the results. Unlike the simple ensemble techniques, the proposed approach generates the final predictions for the test samples by evaluating the confidence in the predictions of the base models. In the final stage of the network, the infection regions are identified, and the amount of infection is calculated when the image is classified as COVID-19. Subsequently, a severity score ranging from 0 to 3 is assigned based on the level of infection severity. Based on the predicted scores, COVID-19 images are categorized into four severity levels: mild, moderate, severe, and critical. The final ensemble segmentation model exceeds all the individual segmentation models with an IoU score of 98.99% and a dice score of 99.49%. Additionally, the infection segmentation model also outperforms all other segmentation models individually, achieving a dice score of 96.75% and an IoU score of 93.71%. Our proposed segmentation-based classification model achieves a maximum accuracy of 98.05% with an F1-score of 97.76% and outperforms five state-of-the-art (SOTA) CNN models and five existing ensemble techniques in the literature. An external dataset also validated the model, and the results showed an accuracy of 97.63% and an F1-score of 97.22%.