Overview
The research introduces CompDiff, a hierarchical compositional diffusion framework designed to address existing imbalances in medical imaging datasets generated for AI. This framework aims to produce high-quality images consistently across diverse demographic groups, specifically targeting difficulties arising from imbalanced training data which often lead to degraded synthesis for rare subgroups and challenges with intersectional demographics not present during training.
Research Context
Generative models are increasingly employed to augment medical imaging datasets to foster fairer AI development. A critical assumption under investigation is whether these generators consistently produce high-quality images across all demographic groups. Models trained on imbalanced data are observed to inherit these imbalances, leading to diminished synthesis quality for less represented subgroups. Furthermore, these models struggle with demographic intersections that are absent from their training data, a phenomenon referred to as the 'imbalanced generator problem.' Current remedies, such as loss reweighting, operate at the optimization level, offering limited efficacy when training signal for specific subgroups is scarce or entirely absent.
The study highlights demographic conditioning as an important, yet underexplored, factor in the development of fair medical image generation techniques.
Approach
CompDiff addresses the imbalanced generator problem by operating at the representation level. It incorporates a dedicated Hierarchical Conditioner Network (HCN), which is responsible for decomposing demographic conditioning information. This decomposition results in single-attribute, pairwise, and composed representations. These representations are then used to produce a demographic token, which is subsequently concatenated with CLIP embeddings. This combined token serves as the cross-attention context for the diffusion model.
The structured factorization facilitated by the HCN is intended to encourage parameter sharing across various subgroups. This design also supports compositional generalization, allowing the model to synthesize images for rare or previously unseen demographic intersections effectively.
Findings
- CompDiff demonstrated favorable performance against standard fine-tuning and FairDiffusion when evaluated on chest X-rays (MIMIC-CXR) and fundus images (FairGenMed) datasets.
- The framework achieved an FID score of 64.3, compared to 75.1 for alternative methods, indicating improved image quality.
- Subgroup equity, measured by ES-FID, also showed improvements.
- Zero-shot intersectional generalization saw up to a 21% FID improvement on held-out intersections, suggesting the model's ability to synthesize for novel demographic combinations.
- Downstream classifiers trained on data generated by CompDiff exhibited improved AUROC (Area Under the Receiver Operating Characteristic curve).
- Training classifiers on CompDiff-generated data resulted in reduced demographic bias.
Why This Matters
The architectural design of demographic conditioning is indicated as a significant factor in achieving fair medical image generation. The findings suggest that by improving the quality and reducing demographic bias in synthetic medical images, AI models developed using these augmented datasets could be more equitable, especially concerning rare or unobserved demographic groups.