From CNNs to transformers: an in-depth study of mycobacteria detection models, augmentation techniques, and ensemble methods

This study presents the first large-scale head-to-head evaluation of over 20 convolutional neural networks (CNNs) and vision transformer (ViT) architectures for automated detection of Mycobacterium tuberculosis in Ziehl–Neelsen (ZN) stained histopathology patches. Our findings reveal that RegNetX-8GF achieves the best performance among individual models, with an accuracy of 95.99%. Contrary to conventional expectations, we show that standard color and geometric augmentation techniques consistently reduce performance, with models trained without augmentation achieving superior results. Finally, a lightweight ensemble of DenseNet-169 and ResNet-34 yields the highest overall accuracy of 96.04%, outperforming all single models across multiple metrics including AUC (0.987), F1 score (0.961), precision (0.9577), and recall (0.964). Unlike prior work that has focused almost exclusively on CNN back-bones, this is the first systematic benchmark of Vision-Transformer architectures (ViT, DeiT, SwinV2) on Ziehl–Neelsen-stained WSI patches for bacilli detection. These results provide a strong, reproducible baseline for patch-level TB detection in low-resource settings.