Loss-Conditioned Utility-Fairness Boundary Modeling for Medical Imaging
| Type of publication: | Conference paper |
| Citation: | Ozbulak_MICCAI_2026 |
| Publication status: | Accepted |
| Booktitle: | Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention |
| Year: | 2026 |
| Month: | September |
| Abstract: | Medical image classifiers can achieve strong global performance while exhibiting different error profiles across sensitive groups. We propose a loss-conditioned framework for efficient utility-fairness boundary modeling in medical image classification. The method combines binary cross-entropy with a differentiable Equalized Odds loss and uses a two-dimensional preference vector to control their relative weights. Independently trained models define reference frontiers, while a single loss-conditioned backbone receives the same preference vector and is queried at both training and unseen preferences. We evaluate the framework on Harvard-GF-3300 glaucoma detection with gender as the sensitive attribute using ResNet-18, EfficientNet-B1, DenseNet-121, and ViT-Small. Conditioned models are trained with nine preference vectors and evaluated on denser aligned grids to assess interpolation. Across architectures, conditioned models approximate useful regions of independently trained ACC-EOD frontiers and enable dense aligned preference queries without additional optimization. These results support loss-conditioned learning as an efficient frontier-exploration tool rather than a single-point accuracy maximization method. |
| Main Research Program: | AI for Everyone |
| Additional Research Programs: |
AI for Life |
| Keywords: | |
| Projects: |
FAIRMI |
| Authors: | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
