Reproduce “PDD: Manifold-Prior Diverse Distillation for Medical Anomaly Detection” on OpenResearch

About this paper

Xijun Lu, Hongying Liu, Fanhua Shang, Yanming Hui, Liang Wan

The PDD framework, developed by researchers at Tianjin University, introduces a dual-teacher, dual-student deep learning architecture for unsupervised medical anomaly detection. It unifies complementary global and local features from heterogeneous backbones to learn a robust normal manifold, achieving state-of-the-art AUROC scores on several medical imaging datasets like HeadCT (97.5%) and BrainMRI (96.7%).

Key insights

• Integrating complementary global contextual priors (Mamba) and local structural priors (ResNet) from heterogeneous backbones is essential for forming a comprehensive and anatomically coherent normal manifold in medical images.
• Unifying diverse features from heterogeneous backbones into a shared high-dimensional manifold effectively leverages their collective knowledge through specialized modules like InA and MMU.
• Employing a dual-student architecture with distinct distillation strategies and a diversity loss prevents representation collapse, enabling students to capture varied aspects of normal patterns for enhanced anomaly sensitivity without sacrificing consistency.

Summary by alphaXiv.