Medical Image Understanding and Analysis (MIUA'2025)

In this work, we present a preliminary study that investigates the robustness of vision transformers (ViTs) compared to convolutional neural networks (CNNs) for the classification of referable diabetic retinopathy, under synthetic noise perturbations. We initially observed that ViTs outperformed CNNs on uncorrupted data, aligning with their known capabilities in capturing complex patterns. However, our preliminary results indicate that the robustness of ViTs diminished as the levels of synthetic noise within the images increased, and we did not observe improved robustness when comparing the ViTs to the baseline CNNs.