Lightweight Faster R-CNN with Standalone BiFPN for Accurate Detection and Counting of Tuberculosis Bacilli in Ziehl-Neelsen Smear Microscopy

<title>Abstract</title> <p>This study proposes a computationally efficient deep learning framework for automated detection and quantification of Mycobacterium tuberculosis in Ziehl–Neelsen–stained sputum images. The proposed method integrates a Bidirectional Feature Pyramid Network (BiFPN) into the Faster R-CNN pipeline. This approach strengthens multi-scale feature fusion for microscopic bacilli, improving sensitivity to small objects without relying on computationally heavy backbones. The framework is organized into clear stages: (1) dataset preparation and augmentation (1,265 images expanded to 6,375), (2) feature enhancement via BiFPN within the detection architecture, (3) model training and inference for bacilli localization, and (4) quantitative validation by comparing automated counts against manual references. Experiments demonstrate strong detection and counting performance, achieving 93.71% mAP@0.5 alongside high agreement with manual counting (Pearson r = 0.997, MAE = 0.22, RMSE = 0.47), indicating that the proposed architectural refinement delivers reliable diagnostic cues while maintaining practical computational efficiency. In settings where microscopy remains the primary frontline tool for tuberculosis control, this research offers timely benefits for accelerating case detection and reducing diagnostic variability by enhancing routine smear microscopy with accurate, scalable, and effective AI assistance.</p>