ABSTRACT
Respiratory diseases are one of the leading causes of death and disability in the world. Integration of AI with existing Chest X-Ray (CXR) diagnostics is currently a hot research topic. On similar lines, we propose a technique termed "Swasta-shwasa" for multi-class classification that associates CXR with one among Tuberculosis, COVID-19, Viral pneumonia, Bacteria Pneumonia, Normal and Lung Opacity ailments based on Deep Learning. The proposed technique which has accomplished an overall 98% test accuracy, 0.9991 AUROC, average Specificity of 99.82% and average Sensitivity of 98.51% involves four stages: Pre-processing, Segmentation, Classification and Saliency map visualization. Further, the trained model is used to predict on unseen real life data of COVID-19 cases from India and a cross-population generalization accuracy of 85% is witnessed. XAI is augmented for model interpretability. We also explore why CLAHE may not be suitable choice for pre-processing of CXRs.