Can Demographic Information Be Reduced in Retinal Fundus Images While Preserving Glaucoma-Relevant Features?

The study investigated whether demographic information (race, sex, ethnicity) encoded in color fundus photographs can be reduced using adversarial perturbations while preserving glaucoma-relevant features. The research included 4,271 patients with 13,959 images from Massachusetts Eye and Ear. Initially, images could predict glaucoma with AUC 0.958 to 0.963 and demographic information with AUC 0.955 to 0.992. Standard adversarial attacks reduced demographic predictions but often degraded glaucoma detection. Disease-aware optimization (DA-PGD and DA-Diffusion) successfully reduced demographic information to 30% or less of baseline while preserving at least 90% glaucoma detection accuracy. Results suggest that demographic and glaucoma-relevant information are partially separable, although reduced demographic recoverability did not fully transfer across different architectures.