CFR-Net: Collaborative Feature Refinement Network for Medical Image Anomaly Detection

Medical image anomaly detection is difficult because models pretrained on natural images often fail to capture the fine-grained, multi-scale, and orientation-sensitive patterns required for medical diagnosis. The authors seek to overcome this domain discrepancy by improving how teacher-student distillation frameworks learn normal anatomical representations.

The authors propose the Collaborative Feature Refinement Network (CFR-Net). Unlike standard distillation methods that decode features directly, CFR-Net introduces a Multi-Path Feature Refinement Module (MPFRM) that processes both frozen teacher features and trainable student features using shared parameters. This module uses three branches—variance-weighted, multi-scale pyramid, and asymmetric convolution—to capture local dispersion, contextual mismatches, and structural orientation. After decoding, the model employs a cross-space consistency objective, where each stream is constrained by the complementary encoder's feature space, supported by a variance-sensitive loss and a dynamic data reorganization strategy.

CFR-Net demonstrates competitive anomaly classification and strong localization performance across six diverse medical imaging benchmarks, including MRI, CT, and histopathology datasets. The collaborative refinement process effectively mitigates the gap between generic natural-image priors and specific medical-domain requirements, allowing the model to better identify anomalies when trained exclusively on normal data.

This work addresses a critical bottleneck in medical AI: the scarcity of annotated abnormal data. By refining features before decoding and enforcing cross-space consistency, CFR-Net provides a more robust architecture for unsupervised anomaly detection, potentially reducing the reliance on large-scale labeled datasets for clinical diagnostic tools.