COVIXNet: A Robust and Explainable Deep Transfer Learning Framework for COVID-19 Detection from Chest X-ray Imagery
Résumé
The COVID-19 pandemic has highlighted the urgent need for fast, accurate, and trustworthy diagnostic tools to complement routine testing procedures. While Chest X-ray (CXR) imaging is a valuable alternative, it requires specialized interpretation that is prone to subjectivity, creating a bottleneck in clinical workflows. This study introduces COVIXNet, a robust and explainable deep transfer learning framework designed to automate the detection of COVID-19 from CXR images with high accuracy and provide clinically interpretable justifications for its predictions. COVIXNet is built upon a DenseNet-121 backbone, enhanced with a spatial attention mechanism to focus on diagnostically significant lung regions. Explainability is achieved using Gradient-weighted Class Activation Mapping (Grad-CAM). The model was trained and evaluated on COVIDXSet, a curated multi-source dataset of 8,591 CXR images. A two-phase transfer learning strategy was employed for effective feature adaptation. COVIXNet demonstrated state-of-the-art performance, achieving an accuracy of 96.8% (95% CI: 95.9% - 97.7%), a precision of 97.2%, and an AUC-ROC of 0.993 (95% CI: 0.989 - 0.997). The explainability module generated clinically meaningful heatmaps, with 92% rated as relevant by expert radiologists and achieving a high Pointing Game Accuracy of 81.5%. The model also showed consistent performance across demographic subgroups and varying image quality. COVIXNet offers a powerful combination of high diagnostic accuracy, robustness, and validated explainability. With a 48 ms inference time and 33.2 MB model size, COVIXNet is a promising and efficient tool for deployment in clinical settings to assist healthcare professionals in the rapid triage and diagnosis of COVID-19.
Citer ce document
Accès au document
Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter
Voir l'article sur le site de la revueAuteur(s)
Statistiques
Consultations : 1
Téléchargements : 0