Abstract:
Tuberculosis remains a formidable global health challenge, requiring advanced and interpretable methodologies. Conventional diagnostic approaches for tuberculosis often suffer from outdated techniques and unnecessary features, impairing their reliability, especially in developing countries like Sri Lanka. This study introduces an explainable Hybrid Convolutional Neural Network (CNN) Swin-Transformer (Swin-T) tailored for Tuberculosis detection in chest X-ray images by leveraging the power of pre-trained CNN and Swin-T. A dataset comprising 270 ethically sourced chest X-ray images, including 171 healthy sub- jects and 99 Pulmonary Tuberculosis subjects, was meticulously Curated from Trincomalee General Hospital, Sri Lanka. The proposed network showcased exceptional performance, yielding a precision of 82.14%, a recall of 92.00%, a specificity of 82.76%, and an accuracy of 92.22%. Notably, employing the gradient- based class activation map (Grad-CAM) technique, the model elucidated Tuberculosis-indicative regions in the chest X-ray images, offering transparency in its diagnostic decisions. These
findings underscore the potential of the explainable hybrid CNN Swin-T Network as a powerful and interpretable tool for early Tuberculosis diagnosis. By highlighting crucial regions indicative of the disease, this model aids clinicians in expedited and accurate
diagnosis, contributing to improved disease management and better patient outcomes.
