Research map / Label-efficient and robust learning / Recognition and long-tail
Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective
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 distribution-aware Mixture of Experts (dMoE), a control-theory-inspired MoE whose routers are conditioned on demographic/clinical attributes, to improve fairness in medical image segmentation across 2D and 3D benchmarks.
Problem
Fairness in medical image segmentation is harmed by imbalanced clinical data across demographic attributes (age, sex, race) and clinical factors (disease severity such as T-stage), causing underperformance on minority subgroups.
Approach
Reinterprets MoE as a feedback control mechanism and adds attribute-specific routers (shared experts, separate gating per attribute flag), framed as mode-switching optimal control so the network adapts its parameter selection to distributional attributes.
Key results
- On Harvard-FairSeg race attribute, dMoE reaches ES-Dice 0.743 for rim segmentation vs 0.716 for FairDiff and 0.705 for FEBS.
- On 3D radiotherapy target segmentation with T-stage, dMoE achieves Dice 0.650 overall and 0.778 on T4, vs 0.608 for MoE and 0.685 for FEBS on T4.
- On HAM10000 age attribute, dMoE improves ES-Dice to 0.801 vs 0.796 for MoE and 0.757 for FEBS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the default MoE router with a distribution-aware router | TransUNet · Harvard-FairSeg rim | ES-Dice: 0.733 (0.713-0.752) -> 0.743 (0.723-0.763) | Helped |
| replace the default MoE router with a distribution-aware router | TransUNet · Harvard-FairSeg cup | ES-Dice: 0.830 (0.809-0.847) -> 0.832 (0.810-0.853) | Helped |
| replace the default MoE router with a distribution-aware router | TransUNet · HAM10000 | ES-Dice: 0.796 (0.741-0.844) -> 0.801 (0.745-0.847) | Helped |
| add dMoE to the 3D ResUNet | 3D ResUNet · Radiotherapy target (T-stage) | ES-Dice: 0.487 (0.447-0.529) -> 0.499 (0.469-0.531) | Helped |
| place dMoE in the decoder only rather than the encoder only (ablation) | TransUNet dMoE · HAM10000 | ES-Dice: 0.841 -> 0.840 | No clear effect |
| place dMoE in both encoder and decoder rather than the encoder only (ablation) | TransUNet dMoE · HAM10000 | ES-Dice: 0.841 -> 0.842 | No clear effect |
| share dMoE parameters across layers (ablation) | TransUNet dMoE · HAM10000 | ES-Dice: 0.812 -> 0.841 | Helped |
| remove mode-switching while retaining feedback control (ablation) | TransUNet dMoE · HAM10000 | ES-Dice: 0.836 -> 0.841 | Helped |
| add feedback and mode-switching control (ablation) | TransUNet dMoE · HAM10000 | ES-Dice: 0.826 -> 0.841 | Helped |
| place dMoE in the decoder only rather than the encoder only (ablation) | 3D ResUNet dMoE · Radiotherapy target (T-stage) | ES-Dice: 0.546 -> 0.569 | Mixed |
| place dMoE in both encoder and decoder rather than the encoder only (ablation) | 3D ResUNet dMoE · Radiotherapy target (T-stage) | ES-Dice: 0.546 -> 0.543 | No clear effect |
| use layer-wise rather than shared dMoE parameters (ablation) | 3D ResUNet dMoE · Radiotherapy target (T-stage) | ES-Dice: 0.546 -> 0.546 | No clear effect |
Surprising findings
- Naive feedback control gave marginal gains or degradation on some datasets, while the proposed mode-switching control consistently improved results.
- On HAM10000, MoE alone matches or beats dMoE on several age subgroups, with dMoE gains mostly in ES metrics and small margins.
Limitations
- Performance gains vary across tasks and datasets; optimal dMoE configuration differs between Transformer and CNN architectures.
- Only a single attribute per task is considered; combined attribute imbalances are not explored.
- ESSP metrics have limitations in capturing gains for minority subgroups, especially in 3D tasks.
Experimental setup
Datasets: Harvard-FairSeg HAM10000 in-house radiotherapy target dataset (prostate CT)
Benchmarks: Harvard-FairSeg HAM10000 in-house radiotherapy target dataset
Metrics: Dice IoU ES-Dice ES-IoU worst-group accuracy
Baselines: TransUNet ADV DRO FEBS FairDiff MoE 3D ResUNet
Builds on: Mixture of Experts Noisy Top-K Gating TransUNet 3D ResUNet
Closest papers by meaning
- FairSeg: A Large-Scale Medical Image Segmentation Dataset for Fairness Learning Using Segment Anything Model with Fair Error-Bound Scaling (ICLR 2024 · similarity 0.90)
- The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective (NeurIPS 2025 · similarity 0.87)
- Diversified and Personalized Multi-rater Medical Image Segmentation (CVPR 2024 · similarity 0.85)
- FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in Medical Image Analysis (ICLR 2024 · similarity 0.85)
- FairMedFM: Fairness Benchmarking for Medical Imaging Foundation Models (NeurIPS 2024 · similarity 0.85)
- Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation (NeurIPS 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.