Few-Shot Learning for Predicting Anti-VEGF Treatment Response in OCT Images

International audience Diabetic macular edema (DME) is a major cause of vision loss, usually treated with intravitreal injections of anti-VEGF agents such as aflibercept and bevacizumab (ranibizumab is no longer commercialized in Tunisia). However, not all patients respond effectively to this therapy, making it crucial to predict treatment outcomes in advance in order to personalize clinical decisions. One of the main challenges is the limited availability of annotated OCT datasets, which limits the performance of conventional deep learning methods. In this study, we investigate Few-Shot Learning (FSL) approaches, specifically Prototypical Networks, Relation Networks, and Model-Agnostic Meta-Learning (MAML), applied to OCT images for predicting patient response to anti-VEGF treatment. These architectures are designed to generalize from only a few training examples per class, making them suitable for medical scenarios with scarce data. Experimental results obtained on a private OCT dataset (good vs. poor responders) show that the Prototypical Network achieved the best performance with 90% accuracy and 85.5 % F1-score, outperforming both Relation Network (63% accuracy) and MAML (56.5% accuracy). These findings highlight the potential of meta-learning strategies to provide robust and efficient predictive models in ophthalmology, paving the way for improved personalized treatment strategies in patients with DME. The Prototypical Network achieved the best performance with 90% ± 2.5 accuracy and 85.5% ± 3.1 F1-score over five-fold cross-validation, outperforming Relation Network and MAML.