Uncertainty-aware Cycle Diffusion Model for Fair Glaucoma Diagnosis

Ziheng Wang, Shuran Yang, Yan Lin, zang wenrui, Yanda Meng

Keywords: Fairness Learning, Image Synthesis, Diffusion Models, ControlNet

TL;DR: Shape-controlled diffusion augments underrepresented data to improve fairness and diagnosis in healthcare AI.

Abstract: Fairness has become a critical ethical concern, particularly in AI-based healthcare applications. The unbalanced and insufficient data size leads to relatively lower diagnosis performance. Conversely, this harms the fairness of AI when applied to real-world scenarios. Generative models, like diffusion models, offer a promising solution by generating diverse synthetic data to support underrepresented groups. This improves fairness and performance while mitigating privacy risks. We propose a shape-controlled framework that incorporates demographic information into an end-to-end diffusion model, along with an automatic selection strategy to identify overconfidently misclassified samples. These challenging samples are then augmented via the generative model to enhance its classification performance. The strategy also removes potentially misleading "lower-quality" synthetic samples. Two ophthalmic experts validated the clinical relevance and plausibility of our synthetic images through random external examination. Our method outperforms state-of-the-art methods on the Harvard-FairVLMed dataset in both fairness and diagnosis accuracy. Our code is available at https://github.com/WANG-ZIHENG/CCG.

Primary Subject Area: Fairness and Bias

Secondary Subject Area: Generative Models

Registration Requirement: Yes

Reproducibility: https://github.com/WANG-ZIHENG/CCG

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 81