Beyond Neurofibrillary Tangles: Explainable AI for Microscopic Tauopathy Classification in Immunofluorescence Imaging

The immunodetection of aberrant molecular changes in Tau polypeptide plays a critical role in the diagnosis and pathogenesis of tauopathies such as Alzheimer's Disease (AD), Progressive Supranuclear Palsy (PSP), and Frontotemporal Dementia (FTD). However, histopathological evaluations of immunofluorescence micrographs remain subjective and labor-intensive. This study presents a deep learning framework for multi-class classification of tauopathies using immunofluorescence microscopy. Five Convolutional Neural Networks (CNNs)-ResNet50, ResNet101, DenseNet121, EfficientNet-B0, and Res2Net-are evaluated with explainability techniques, including EigenGradCAM, AblationCAM, FullGrad, LIME, and Deep Feature Factorization. Our findings indicate that high-performing models, particularly EfficientNet-B0 and DenseNet121, frequently emphasize background regions rather than neurofibrillary tangles (NFTs). Ablation studies reveal that removing background features affects classification as significantly as NFT removal, suggesting that secondary microstructural events outside classical tau markers may contribute to model predictions. Lower-performing models relied more strictly on NFTs, potentially reflecting a constrained feature extraction strategy. These results highlight AI's potential in pathology analysis to identify structural biomarkers beyond conventional Tau-based markers. Further validation is required to determine whether background saliency represents meaningful pathology or dataset artifacts. Future work should focus on multi-modal integration and biomarker validation to enhance AI's clinical utility in neuropathology.