[論文レビュー] PDD: Manifold-Prior Diverse Distillation for Medical Anomaly Detection

Xijun Lu, Hongying Liu+3 more|arXiv (Cornell University)|Mar 7, 2026

Anomaly Detection Techniques and Applications被引用数 0

ひとこと要約

PDD unifies dual teachers into a high-dimensional manifold and distills to dual students with diverse learning paths, achieving state-of-the-art medical anomaly detection across multiple datasets.

ABSTRACT

Medical image anomaly detection faces unique challenges due to subtle, heterogeneous anomalies embedded in complex anatomical structures. Through systematic Grad-CAM analysis, we reveal that discriminative activation maps fail on medical data, unlike their success on industrial datasets, motivating the need for manifold-level modeling. We propose PDD (Manifold-Prior Diverse Distillation), a framework that unifies dual-teacher priors into a shared high-dimensional manifold and distills this knowledge into dual students with complementary behaviors. Specifically, frozen VMamba-Tiny and wide-ResNet50 encoders provide global contextual and local structural priors, respectively. Their features are unified through a Manifold Matching and Unification (MMU) module, while an Inter-Level Feature Adaption (InA) module enriches intermediate representations. The unified manifold is distilled into two students: one performs layer-wise distillation via InA for local consistency, while the other receives skip-projected representations through a Manifold Prior Affine (MPA) module to capture cross-layer dependencies. A diversity loss prevents representation collapse while maintaining detection sensitivity. Extensive experiments on multiple medical datasets demonstrate that PDD significantly outperforms existing state-of-the-art methods, achieving improvements of up to 11.8%, 5.1%, and 8.5% in AUROC on HeadCT, BrainMRI, and ZhangLab datasets, respectively, and 3.4% in F1 max on the Uni-Medical dataset, establishing new state-of-the-art performance in medical image anomaly detection. The implementation will be released at https://github.com/OxygenLu/PDD

研究の動機と目的

• Motivate the need for manifold-level modeling in medical anomaly detection due to diffuse, heterogeneous anomalies in medical images.
• Introduce a dual-teacher framework to fuse global and local priors from heterogeneous backbones.
• Develop a dual-student distillation scheme with manifold-aware modules and a diversity loss to prevent representation collapse.
• Demonstrate strong results across multiple medical datasets and analyze components via ablations.

提案手法

• Two frozen teachers (VMamba-Tiny and wide-ResNet50) provide global contextual and local structural priors.
• Inter-Level Feature Adaption (InA) fuses shallow features from both backbones.
• Manifold Matching and Unification (MMU) aligns heterogeneous manifolds into a unified space.
• Dual students learn with: (1) layer-wise distillation via InA (local consistency); (2) skip-projected manifold features via MPA (cross-layer dependencies).
• Manifold Prior Affine (MPA) injects prior knowledge through MLP-based affine transforms with skip connections.
• Diversity loss prevents collapse by encouraging different representations at low-dimensional layers while maintaining similarity at high-dimensional layers.
• Overall objective combines knowledge distillation loss, prior-guided reconstruction loss, and a diversity loss with learnable weights.

Figure 1 : Grad-CAM visualization of frozen Vmamba and ResNet across medical and industrial images. Within each group, feature maps progress from low-dimensional to high-dimensional feature representations (top to bottom). At the same feature dimension, the dispersed and aggregated activation patter

実験結果

リサーチクエスチョン

• RQ1 Can dual heterogeneous backbones be unified into a common manifold to better model normal anatomy for medical anomaly detection?
• RQ2 Does dual-student, diversity-aware distillation improve robustness and detection of subtle medical anomalies compared to single-teacher approaches?
• RQ3 How do intra-backbone fusion (InA) and manifold-unification (MMU) contribute to performance on diverse medical datasets?
• RQ4 What is the effect of prior-affine injection (MPA) and diversity regularization on anomaly localization and false positives?

主な発見

Method	HeadCT AUROC	Zhanglab AUROC	BrainMRI AUROC	CheXpert AUROC	Notes
f-AnoGAN	82.6	75.5	77.1	65.8	Baseline methods comparison
CutPaste	73.0	73.3	67.0	65.5	Industrial/medical baseline
RD4AD	74.3	87.5	80.9	71.9	Knowledge-distillation baseline
SQUID	75.4	87.6	74.7	78.1	Medical UAD baseline
SIMSID	74.9	91.1	81.5	79.7	Self-supervised baseline
Skip-TS	85.7	79.2	88.2	68.7	Skip/teacher-student baseline
Ours	97.5	94.0	96.7	79.1	PDD with dual-teacher and dual-student distillation

• PDD achieves state-of-the-art AUROC on HeadCT (97.5), Zhanglab (94.0), and BrainMRI (96.7).
• PDD reaches competitive CheXpert performance (79.1% AUROC) versus the top method.
• On Uni-Medical, PDD attains the best F1 max across brain, liver, and retinal categories and the top AP in retinal.
• Ablations show the value of the dual-teacher MMU/InA, the addition of MPA, and the dual-student design, with the full model achieving 94.0% AUROC on ZhangLab and 96.6% F1 on BrainMRI in ablations.
• Diversity loss plus teacher-student alignment yields the strongest anomaly detection and localization performance.
• PDD demonstrates fewer false positives on normal samples compared to baselines in qualitative analyses.

Figure 2 : Overview of the proposed PDD framework. The framework employs a dual-teacher and dual-student architecture. The teachers consist of frozen VMamba-Tiny and frozen wide-ResNet50 encoders, whose intermediate features are fused via the InA module (shown in (b)) to obtain $f_{b}^{i}$ . The two