SensaNet: a lightweight DL model for tuberculosis detection in histopathological images

Abstract

Tuberculosis remains a global health problem, particularly in resource-limited settings in which early and accurate diagnosis is paramount. This research presents SensaNet, an efficient yet light-weight binary tuberculosis (TB) classification for histopathological image patches. SensaNet is evaluated against a well-curated set of 27,987 Kinyoun-stained image patches that were scanned from digitized slides of sputum smears, with wellbalanced bacilli-positive and negative distributions. SensaNet combines current architectural innovations such as Squeeze-and-Excitation (SE) blocks, Swish activation, and single-head self-attention to amplify feature representation in channel and spatial domains with lower memory cost. Experimental results confirm that SensaNet achieves an accuracy of 98.54%, precision of 98.19%, and recall of 98.95%, outperforming several baseline architectures on sensitivity with the compact size of 12.2 MB and only above 1 million parameters. Comparative analysis proves SensaNet's suitability for TB diagnosis, offering a good trade-off between diagnostic performance and computational expense. These results weigh in favor of the model's potential deployment for real-time point-of-care diagnostics, particularly for low-resource environments.