FairEnc: A Fair Vision-Language Model for Glaucoma Detection

Researchers have introduced FairEnc, an equitable pretraining approach for vision-language models (VLMs) designed to facilitate automated glaucoma detection. This technique concurrently mitigates biases related to sensitive characteristics such as race, gender, ethnicity, and language across both text and images. In the textual encoder, a substantial language model creates synthetic clinical narratives that incorporate diverse sensitive attributes while maintaining the essence of the disease, utilizing contrastive alignment to promote demographic-invariant representations. For the visual encoder, a dual-level fairness strategy that integrates mutual information regularization is applied. This initiative tackles fairness issues in AI-based healthcare, particularly in efforts to avert irreversible vision loss due to glaucoma.