Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective

Yujin Oh, Pengfei Jin, Sangjoon Park, Sekeun Kim, Siyeop Yoon, Kyungsang Kim, Jin Sung Kim, Xiang Li, Quanzheng Li

TL;DR

The paper tackles fairness in medical image segmentation under distributional biases across demographic and clinical factors. It proposes Distribution-aware Mixture of Experts ($dMoE$), a control-inspired MoE that incorporates distribution information into gating to enable adaptive, attribute-aware parameter selection across Transformer and CNN backbones for 2D and 3D tasks. The approach is theoretically framed within optimal-control and mode-switching concepts and validated on Harvard-FairSeg, HAM10000, and a radiotherapy target dataset, showing state-of-the-art fairness metrics and robust generalization, with extensive ablations. The work advances fairness in clinical AI by bridging control theory and distribution-aware learning, providing a practical, transferable framework and releasing source code for reproducibility and further deployment in diverse hospital settings.

Abstract

Ensuring fairness in medical image segmentation is critical due to biases in imbalanced clinical data acquisition caused by demographic attributes (e.g., age, sex, race) and clinical factors (e.g., disease severity). To address these challenges, we introduce Distribution-aware Mixture of Experts (dMoE), inspired by optimal control theory. We provide a comprehensive analysis of its underlying mechanisms and clarify dMoE's role in adapting to heterogeneous distributions in medical image segmentation. Furthermore, we integrate dMoE into multiple network architectures, demonstrating its broad applicability across diverse medical image analysis tasks. By incorporating demographic and clinical factors, dMoE achieves state-of-the-art performance on two 2D benchmark datasets and a 3D in-house dataset. Our results highlight the effectiveness of dMoE in mitigating biases from imbalanced distributions, offering a promising approach to bridging control theory and medical image segmentation within fairness learning paradigms. The source code will be made available. The source code is available at https://github.com/tvseg/dMoE.

Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective

TL;DR

The paper tackles fairness in medical image segmentation under distributional biases across demographic and clinical factors. It proposes Distribution-aware Mixture of Experts (), a control-inspired MoE that incorporates distribution information into gating to enable adaptive, attribute-aware parameter selection across Transformer and CNN backbones for 2D and 3D tasks. The approach is theoretically framed within optimal-control and mode-switching concepts and validated on Harvard-FairSeg, HAM10000, and a radiotherapy target dataset, showing state-of-the-art fairness metrics and robust generalization, with extensive ablations. The work advances fairness in clinical AI by bridging control theory and distribution-aware learning, providing a practical, transferable framework and releasing source code for reproducibility and further deployment in diverse hospital settings.

Abstract

Ensuring fairness in medical image segmentation is critical due to biases in imbalanced clinical data acquisition caused by demographic attributes (e.g., age, sex, race) and clinical factors (e.g., disease severity). To address these challenges, we introduce Distribution-aware Mixture of Experts (dMoE), inspired by optimal control theory. We provide a comprehensive analysis of its underlying mechanisms and clarify dMoE's role in adapting to heterogeneous distributions in medical image segmentation. Furthermore, we integrate dMoE into multiple network architectures, demonstrating its broad applicability across diverse medical image analysis tasks. By incorporating demographic and clinical factors, dMoE achieves state-of-the-art performance on two 2D benchmark datasets and a 3D in-house dataset. Our results highlight the effectiveness of dMoE in mitigating biases from imbalanced distributions, offering a promising approach to bridging control theory and medical image segmentation within fairness learning paradigms. The source code will be made available. The source code is available at https://github.com/tvseg/dMoE.

Paper Structure

This paper contains 28 sections, 20 equations, 3 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fairness Learning in Medical Image Segmentation
  4. Mixture of Expert in Multi-distribution Learning
  5. Training Neural Networks as Optimal Control
  6. Method
  7. Revisiting MoE
  8. Distribution-aware MoE
  9. Interpreting dMoE Through Optimal Control
  10. Experimental Results
  11. Datasets
  12. Harvard-FairSeg
  13. HAM10000
  14. Radiotherapy Target Dataset
  15. Implementation Details
  16. ...and 13 more sections

Figures (3)

  • Figure 1: The influence of clinical data distribution on medical image segmentation and the role of dMoE as a distribution-aware control mechanism to address inequity challenges. Transparent blue lines within the violin plots connect the most densely populated regions of each attribute, visually representing overall equity.
  • Figure 2: (a) Schematic of the dMoE segmentation network for fairness learning, and (b) its interpretation through a control system.
  • Figure 3: (a) Violin plots depicts attribute-wise segmentation performance. Transparent blue lines within the plots connect the most densely populated regions for each attribute, visually representing overall equity. (b) Qualitative comparison across different subgroups.