Generative AI to augment the fairness of foundation models in cancer pathology diagnosis.

• DOI: 10.1200/jco.2025.43.16_suppl.e23230
• Publication Date: 2025-05-28T16:26:46Z
• Authors: Shih-Yen Lin; Pei-Chen Tsai; Yixuan Dong; Po-Chih Lin; Fang-Yi Su; Yuk Yeung Ho; Ting-Wan Kao; Tsung-Lu Michael Lee; Po-Jen Lin; Jung-Hsien Chiang; Kun-Hsing Yu
• Keywords: Augment; Foundation (evidence)

e23230 Background: Pathology foundation models, a type of state-of-the-art deep learning models trained on diverse and large-scale datasets, have shown the ability to extract useful pathology patterns for cancer diagnosis. However, their reliability across different demographic groups is hindered by limited training samples from minority populations. To address this challenge, we developed generative AI-based approach, Fairness Denoising Diffusion Probabilistic Models (DDPM), enhance fairness models. Methods: We obtained 30,664 whole-slide images The Cancer Genome Atlas (TCGA) database covering 33 types. Self-reported race, sex, age, were collected alongside images. DDPM mitigates biases augmenting data populations via diffusion model. evaluated three (Gigapath, UNI, CHIEF) tumor detection genetic mutation tasks. further incorporated DPPM in process model performance patient assessed using accuracy difference (AccDiff), area under receiver operating characteristic curve (AUCDiff), equal opportunity (EOpp), balanced (EBAcc). Results: significantly reduced AI bias datasets tasks (Table 1). Specifically, completely eliminated racial AUCDiff EBAcc CHIEF AUCDiff, AccDiff, UNI In addition, 80.0% gender Gigapath Across diagnostic tasks, all sensitive attributes (race, age). Conclusions: Our study shows that effectively By incorporating DDPM, algorithms achieved greater equity populations, representing pivotal step toward global adoption fair reliable diagnoses. Comparison conventional (a) Evaluation Metrics Baseline % Mitigated EOpp 21/41 41/230 51.22% 7/17 17/115 41.18% AccDiff 8/14 14/115 57.14% 5/18 18/115 27.78% (b) 23/34 34/230 67.65% 7/12 12/115 58.33% 7/13 13/115 53.85% 6/16 16/115 37.50% (c) GigaPath 8/21 21/230 38.10% 8/10 10/115 80.00% 4/10 40.00% 7/14 50.00% *Cancer types covered study: BRCA, LUAD, UCEC, COAD, READ, LUSC, HNSC, KIRC, LGG, SKCM, STAD, BLCA, LIHC, SARC, THYM, CESC, PAAD, KICH, CHOL, OV, THCA.