Research map / Learning theory and algorithms / Fairness

FairSeg: A Large-Scale Medical Image Segmentation Dataset for Fairness Learning Using Segment Anything Model with Fair Error-Bound Scaling

ICLR 2024 · Learning theory and algorithms › Fairness · approach: Post-processing and transport for fairness · 5 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 Harvard-FairSeg, a 10,000-subject SLO fundus optic disc/cup segmentation dataset with six sensitive attributes, and a fair error-bound scaling (FEBS) loss plus equity-scaled metrics (ES-Dice, ES-IoU) benchmarked on SAMed and TransUNet.

Problem

Existing medical fairness datasets target classification and lack sensitive-attribute annotations for segmentation, so fairness of medical segmentation models cannot be studied or compared, and classification-style fairness metrics are hard to interpret for segmentation.

Approach

Reweight the Dice loss per sensitive-attribute group using a tanh-scaled ratio of group losses (fair error-bound scaling) applied to SAMed, and define an equity-scaled metric ESSP = metric / (1 + sum of group-vs-overall disparities) to score fairness alongside accuracy.

Key results

• SAMed+FEBS achieves the best Rim ES-Dice of 0.7529 and ES-IoU of 0.6451 under race attribute, surpassing other SAMed-based methods.
• TransUNet+FEBS obtains the best Rim ES-Dice of 0.7050 (race) versus 0.7034 for TransUNet baseline.
• Under the language attribute, SAMed+ADV gives the best results among SAMed methods, and FEBS yields the top Cup ES-Dice of 0.8291 for SAMed.

Measured effects reported in the paper

Surprising findings

• Adversarial and GroupDRO baselines often perform comparably to FEBS, and FEBS does not consistently win; for many attribute/metric combinations the differences are small (often <0.01).
• TransUNet-based methods underperform SAMed-based methods on Rim segmentation across attributes despite the same training setup.

Limitations

• Fairness gains from FEBS are marginal and not consistently superior to baselines.
• Segmentation labels are derived by registering OCT-derived masks to SLO images, with only about 80% registration success and manual filtering by five graders.
• Only SAMed and TransUNet are benchmarked, with a single disc-cup task on one imaging modality.

Experimental setup

Datasets: Harvard-FairSeg

Benchmarks: Harvard-FairSeg

Metrics: Dice IoU ES-Dice ES-IoU HD95 ASD NSD

Baselines: SAMed SAMed+ADV SAMed+GroupDRO TransUNet TransUNet+ADV TransUNet+GroupDRO

Builds on: Segment Anything Model SAMed TransUNet GroupDRO adversarially fair representations

Closest papers by meaning

• Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective (ICML 2025 · similarity 0.90)
• The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective (NeurIPS 2025 · similarity 0.89)
• FairMedFM: Fairness Benchmarking for Medical Imaging Foundation Models (NeurIPS 2024 · similarity 0.89)
• FairCLIP: Harnessing Fairness in Vision-Language Learning (CVPR 2024 · similarity 0.88)
• FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in Medical Image Analysis (ICLR 2024 · similarity 0.88)
• Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline (CVPR 2025 · similarity 0.85)