DuetFair: Coupling Inter- and Intra-Subgroup Robustness for Fair Medical Image Segmentation
Researchers have introduced DuetFair, a novel dual-axis fairness framework designed to address performance disparities in medical image segmentation models across different subgroups. Existing methods often overlook internal heterogeneity within subgroups, leading to 'intra-group hidden failures' where difficult cases are masked by average performance metrics. To resolve this, the team developed FairDRO, which integrates distribution-aware mixture-of-experts (dMoE) with subgroup-conditioned distributionally robust optimization (DRO). This approach simultaneously enhances inter-subgroup adaptation and intra-subgroup robustness. Evaluations on three benchmarks, including Harvard-FairSeg, HAM10000, and a 3D radiotherapy cohort, demonstrate significant improvements. FairDRO achieved superior equity-scaled performance and notably improved worst-case subgroup outcomes. Specifically, it increased worst-group Dice scores by up to 7.4% under institution-based grouping compared to strong baselines. This advancement promises more equitable and reliable AI-driven diagnostic tools, ensuring that minority or complex cases within demographic groups receive accurate segmentation, thereby reducing potential health disparities in automated medical imaging analysis.
Wire timeline
DuetFair: Coupling Inter- and Intra-Subgroup Robustness for Fair Medical Image Segmentation
Researchers have introduced DuetFair, a novel dual-axis fairness framework designed to address performance disparities in medical image segmentation models across different subgroups. Existing methods often overlook internal heterogeneity within subgroups, leading to 'intra-group hidden failures' where difficult cases are masked by average performance metrics. To resolve this, the team developed FairDRO, which integrates distribution-aware mixture-of-experts (dMoE) with subgroup-conditioned distributionally robust optimization (DRO). This approach simultaneously enhances inter-subgroup adaptation and intra-subgroup robustness. Evaluations on three benchmarks, including Harvard-FairSeg, HAM10000, and a 3D radiotherapy cohort, demonstrate significant improvements. FairDRO achieved superior equity-scaled performance and notably improved worst-case subgroup outcomes. Specifically, it increased worst-group Dice scores by up to 7.4% under institution-based grouping compared to strong baselines. This advancement promises more equitable and reliable AI-driven diagnostic tools, ensuring that minority or complex cases within demographic groups receive accurate segmentation, thereby reducing potential health disparities in automated medical imaging analysis.
cs.AI updates on arXiv.org