Research map / Label-efficient and robust learning / OOD and anomaly detection
DIsoN: Decentralized Isolation Networks for Out-of-Distribution Detection in Medical Imaging
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.
Proposes Decentralized Isolation Networks (DIsoN) and class-conditional CC-DIsoN, which detect OOD medical images by measuring how many federated-style communication rounds a binary isolation classifier needs to separate a test sample from training data, exchanging only model parameters.
Problem
OOD detection in medical imaging typically cannot use training data at deployment because the data are private, proprietary, or too large to ship, so direct comparison of test samples with training data is infeasible.
Approach
Train a binary classifier to isolate a single target sample from source training data and use the convergence time (number of steps or communication rounds) as the OOD score; decentralize this via federated-style parameter exchange between source and target nodes, and condition sampling on the predicted class (CC-DIsoN).
Key results
- CC-DIsoN achieves average AUROC 80.00 and FPR95 59.20 across Chest X-Ray, Dermatology and Breast Ultrasound, versus 70.16 AUROC and 72.77 FPR95 for the best baseline CIDER.
- On MIDOG near-OOD, CC-DIsoN averages 69.6 AUROC and 81.0 FPR95; on far-OOD it reaches 98.3 AUROC and 4.4 FPR95.
- Class-conditioning reduces average FPR95 by 8.2% across the three artifact datasets, and augmentations improve Dermatology AUROC by 13.69%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CC-DIsoN instead of CIDER | OOD detection on Chest X-Ray, Dermatology, and Breast Ultrasound · Average (Chest X-Ray, Dermatology, Breast Ultrasound) | AUROC: 70.16{ $\pm$3.4} -> 80.00{ $\pm$0.7} | Helped |
| Use CC-DIsoN instead of CIDER | OOD detection on Chest X-Ray, Dermatology, and Breast Ultrasound · Average (Chest X-Ray, Dermatology, Breast Ultrasound) | FPR95: 72.77{ $\pm$2.5} -> 59.20{ $\pm$3.1} | Helped |
| Use CC-DIsoN instead of ViM | OOD detection on MIDOG near-OOD domains · MIDOG near-OOD average | AUROC: 62.4 -> 69.6 | Helped |
| Use CC-DIsoN instead of PALM | OOD detection on MIDOG far-OOD · MIDOG far-OOD average | AUROC: 98.3 -> 98.3 | No clear effect |
| Add class-conditional source sampling by predicted class (ablation) | DIsoN on Chest X-Ray, Dermatology, and Breast Ultrasound · Average (Chest X-Ray, Dermatology, Breast Ultrasound) | AUROC: 78.3 -> 80.0 | Mixed |
| Add class-conditional source sampling by predicted class (ablation) | DIsoN on Chest X-Ray, Dermatology, and Breast Ultrasound · Average (Chest X-Ray, Dermatology, Breast Ultrasound) | FPR95: 67.4 -> 59.2 | Helped |
| Assign incorrect classes to all ID and OOD targets (ablation) | CC-DIsoN on Breast Ultrasound · Breast Ultrasound | AUROC: 65.6 -> 61.5 | Hurt |
| Assign incorrect classes to all ID and OOD targets (ablation) | CC-DIsoN on Dermatology · Dermatology | AUROC: 89.5 -> 83.1 | Hurt |
| Assign incorrect classes to all ID and OOD targets (ablation) | CC-DIsoN on Chest X-Ray · Chest X-Ray | AUROC: 84.9 -> 77.9 | Hurt |
| Assign incorrect classes only to ID targets (ablation) | CC-DIsoN on Breast Ultrasound · Breast Ultrasound | AUROC: 65.6 -> 34.6 | Hurt |
| Assign incorrect classes only to OOD targets (ablation) | CC-DIsoN on Breast Ultrasound · Breast Ultrasound | AUROC: 65.6 -> 89.6 | Helped |
| Add random image augmentations during training (ablation) | DIsoN · Breast Ultrasound | AUROC: +4.43% | Helped |
Surprising findings
- Mislabeling only OOD targets slightly improves detection, since OOD samples are compared to unrelated classes and isolate faster.
- Deeper ResNet34 and slimmer ResNet18 variants did not outperform the standard ResNet18.
Limitations
- Substantial per-sample inference compute (40 s to 4 min), requiring one isolation training run per target sample.
- Performance drops when predicted classes for ID targets are wrong, so class-conditioning relies on a reasonably accurate primary model.
- Only single-target-sample processing is supported; multiple targets are not handled simultaneously.
Experimental setup
Datasets: Dermatology (D7P) BreastMNIST CheXpert MIDOG OpenMIBOOD CCAgT FNAC 2019
Benchmarks: Dermatology artifact OOD Breast Ultrasound artifact OOD Chest X-ray pacemaker OOD MIDOG near-OOD MIDOG far-OOD (CCAgT, FNAC 2019)
Metrics: AUROC FPR95
Baselines: MSP MDS fDBD ViM iForest CIDER PALM
Builds on: Isolation Forest Deep Isolation Forest FedAvg
Closest papers by meaning
- OpenMIBOOD: Open Medical Imaging Benchmarks for Out-Of-Distribution Detection (CVPR 2025 · similarity 0.88)
- Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.87)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.85)
- Enhancing the Power of OOD Detection via Sample-Aware Model Selection (CVPR 2024 · similarity 0.85)
- Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection (ICML 2025 · similarity 0.85)
- DPU: Dynamic Prototype Updating for Multimodal Out-of-Distribution Detection (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.