Research map / Label-efficient and robust learning / OOD and anomaly detection

OpenMIBOOD: Open Medical Imaging Benchmarks for Out-Of-Distribution Detection

CVPR 2025 · Label-efficient and robust learning › OOD and anomaly detection · approach: Spurious-aware OOD detection · 10 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces OpenMIBOOD, a medical imaging OOD detection benchmark with three MIB settings (MIDOG, PhaKIR, OASIS-3) spanning 14 datasets, and evaluates 24 post-hoc OOD methods, finding natural-image rankings do not transfer.

Problem

Existing OOD detection benchmarks (e.g., OpenOOD) are built on natural images and lack medical imaging coverage; they also merge covariate-shifted ID data with ID, conflating OOD detection with generalization.

Approach

Build medical OOD benchmarks with separate cs-ID, near-OOD and far-OOD categories around three ID classification tasks, and evaluate post-hoc OOD detectors to show that natural-image rankings do not carry over to medical data.

Key results

• MDSEns achieves the best average near-OOD AUROC across the three MIBs (ranked first in Table 1), while ViM reaches 80.74 average AUROC on the MIBs versus 72.08 on ImageNet1k.
• Classification-based post-hoc methods perform poorly on medical data; e.g., the average AUROC of classification-based approaches on EndoSeg18 is 33.71.
• Feature-space methods consistently outperform probability/logit-based methods on the medical benchmarks, reversing the trend seen on natural images.

Measured effects reported in the paper

Surprising findings

• Classification-based OOD methods sometimes assign higher confidence to OOD inputs than ID inputs, e.g., mean softmax for Grasper rises from 79.07% (ID) to 94.04% (EndoSeg18 OOD).
• Top-ranked methods on natural-image benchmarks do not transfer; ranking on ImageNet1k differs substantially from medical rankings.

Limitations

• Focuses only on classification tasks; segmentation settings are not covered.
• Evaluates only post-hoc methods without training-based OOD approaches.
• Some datasets are curated via crops and heuristics, which may introduce unintended shifts.

Experimental setup

Datasets: MIDOG CCAgT FNAC 2019 PhaKIR Cholec80 EndoSeg15 EndoSeg18 Kvasir-SEG CATARACTS OASIS-3 ATLAS BraTS

Benchmarks: OpenMIBOOD (MIDOG, PhaKIR, OASIS-3 settings) ImageNet1k (OpenOOD)

Metrics: AUROC FPR@95 AUPR-IN AUPR-OUT F1-score

Baselines: MDS Residual ViM KNN ReAct ODIN EBO MSP

Builds on: OpenOOD OpenOOD v1.5 Full-Spectrum OOD Detection

Closest papers by meaning

• MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.88)
• DIsoN: Decentralized Isolation Networks for Out-of-Distribution Detection in Medical Imaging (NeurIPS 2025 · similarity 0.88)
• Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.87)
• Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · similarity 0.87)
• Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation (NeurIPS 2025 · similarity 0.86)
• ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.86)