Uncertainty-Aware Tau Detection in Progressive Supranuclear Palsy Using Object Detection Models

Abstract

Abnormal tau accumulation is a hallmark of neurodegenerative tauopathies such as Progressive Supranuclear Palsy (PSP). Traditional post-mortem assessments rely on manual lesion annotation, which is time-consuming and subjective. Existing machine learning methods typically involve multi-stage, feature-based pipelines, resulting in limited scalability and reliance on handcrafted features. This work evaluates deep learning-based object detection (OD) as a unified, end-to-end alternative for automating tau lesion identification and classification. We fine-tuned YOLO and Faster R-CNN models on a dataset of 29,852 tau lesions from 16 PSP brain slides, achieving mAP@50 of 0.702. To address the limitations of conventional OD models, which assume overconfident spatial predictions, we integrated Monte Carlo Dropout into both architectures to estimate spatial and label uncertainty. We evaluated detection uncertainty using Probability-based Detection Quality. Our results demonstrate that incorporating uncertainty offers a more comprehensive evaluation of detection reliability, complementing traditional metrics like mAP. For clinical application, we integrate a non-stochastic YOLO model into QuPath for whole-slide inference. Downstream clinical validation was conducted, including SHAP-based saliency analysis and correlation between model-predicted tau densities and PSP Rating Scale scores, a clinical disease severity measure. The alignment of detection outputs with pathological features and clinical severity underscores the clinical utility of OD models. Overall, this work highlights the complementary strengths of deep object detectors and uncertainty quantification, offering a path toward scalable, interpretable, and clinically meaningful tau pathology analysis. Code is available at https://github.com/Pidongg/dl_histopathology.