Harnessing semi-supervised graph-based learning to advance automated bacilli detection in digital tuberculosis microscopy with limited expert annotations.

Automated acid-fast bacilli (AFB) detection with digital microscopy is important for the improvement of tuberculosis diagnosis in communities or populations where expert annotations are not available. In this paper, we propose a semi-supervised graph-based learning approach with label spreading and effectively make use of both the small number of labeled images and the huge collection of unlabeled high-resolution smear microscopy images. The approaches diffuse label evidence over graph-derived representations of the images and use smooth diffusion-over-graph algorithms with soft constraints to alleviate error propagated by standard methods. The experimental results show that this framework can robustly and precisely localize bacilli, which is highly competitive overall accuracy and F1-score even with limited expert annotation. These findings demonstrate that the method is applicable to rapid, scalable AI-powered tuberculosis diagnosis in low-resource settings and suggest a role for semi-supervised learning in reducing manual annotation requirements while enabling digital pathology workflows.