Reproduce “Multi-layer Aggregation as a key to feature-based OOD detection” on OpenResearch

About this paper

Benjamin Lambert, Florence Forbes, Senan Doyle, Michel Dojat

Multi-layer feature aggregation improves feature-based out-of-distribution detection for 3D brain MRI segmentation compared with using a single network layer. Across a benchmark of nearly 7,800 shifted volumes, multi-layer Spectral Signatures reached 0.93 AUROC and FRODO reached 0.95 with an Attention UNet, although performance varied by shift type and segmentation architecture.

Key insights

• Representations from multiple network depths provide more robust OOD evidence than a single selected layer: multi-layer Spectral Signatures, Prototypes, and Mahalanobis scoring outperform their single-layer variants overall.
• Detection difficulty depends strongly on the shift: far-OOD and modality changes are generally easy to identify, while unseen diagnoses and some artifacts remain challenging.
• OOD performance depends on the segmentation architecture itself and is not predicted solely by in-distribution Dice; feature collapse can make anomalous inputs resemble ID examples.

Summary by alphaXiv.