Interpretable COVID-19 Chest X-Ray Classification via Orthogonality Constraint

Deep neural networks have increasingly been used as an auxiliary tool in healthcare applications, due to their ability to improve the performance of several diagnosis tasks. However, these methods are not widely adopted in clinical settings due to practical limitations in the reliability, generalizability, and interpretability of deep learning-based systems. As a result, recent methods have highlighted the importance of additional constraints with the cross-entropy loss function to overcome some of the challenges as well as facilitate their acceptance in the healthcare community. In this work, we investigate the benefit of using Orthogonal Spheres (OS) constraints for the classification of COVID-19 cases from chest X-ray images. The OS constraint can be written as a simple orthonormality loss that is used in conjunction with the standard cross-entropy loss during network training. Previous studies have demonstrated significant benefits in applying such constraints to deep learning models. Our findings corroborate these observations, indicating that the orthonormality loss effectively produces improved semantic localization in Grad-CAM visualizations, better classification performance, reduced model calibration error, and improved model generalizability and reliability on the COVID-19 dataset. Our approach achieves an increase in accuracy of 1.6% and 4.8% for two- and three-class classification compared to the baseline models, respectively.