Domain Generalization for Multi-disease Detection in Fundus Photographs

Domain generalization (DG) is a paradigm ensuring machine learning algorithms predict well on unseen domains. Recent computer vision research in DG highlighted how inconsistencies in datasets, architectures, and model criteria challenge fair comparisons. In the medical domain, the application of DG algorithms assumes an even more challenging task as medical data often exhibit significant variability due to diverse imaging modalities, patient demographics, and disease characteristics. In light of this, DG algorithms need to generalize effectively across different medical settings and patient populations for ensuring robustness and fairness in healthcare applications. In this paper, we evaluate various DG algorithms and strategies for the application of multi-disease detection in fundus photographs. We conducted extensive experiments using four heterogeneous datasets: OPHDIAT (France, diabetic population), OphtaMaine (France, general population), RIADD (India, general population) and ODIR (China, general population). The following diseases were targeted: diabetes, glaucoma, cataract, age-related macular degeneration, hypertension, myopia and other diseases/abnormalities.