ARC-Inspired CNN Architecture for Enhanced Tuberculosis Detection from Chest Radiographs

Tuberculosis (TB) remains a significant global health challenge, particularly in resourceconstrained regions. Computer-aided diagnosis systems in deep learning offer promising solutions for rapid TB screening from chest radiographs (CXRs). This paper introduces a novel Adaptively Refined Convolutional (ARC) neural network architecture specifically optimized for TB detection. Unlike conventional approaches that focus primarily on deeper architectures, our method employs strategic convolution patterns and dynamic feature integration to enhance diagnostic accuracy while maintaining computational efficiency. Experimental results demonstrate that our approach achieves 99.67 % training accuracy and 93.47 % test accuracy on a standard TB chest radiograph dataset, outperforming several state-of-the-art architectures. The proposed model exhibits superior convergence properties with significant accuracy improvements observed from the sixth epoch onward. This work bridges a crucial research gap by addressing the need for specialized architectures tailored to TB's unique radiographic signatures rather than relying on general-purpose CNN designs.