Comparative analysis of CNN, Transformer, and Mamba-based models for automated eye disease classification from fundus images

We present a comparative analysis of six models belonging to Convolutional Neural Network (CNN), Transformer, and Mamba families for automated eye disease classification from fundus images under identical experimental conditions. To the best of our knowledge, our work is the first to explore Mamba-based models for automated multi-class eye disease classification from color fundus images. Our analysis suggests that, when evaluated fairly, modern CNN, Transformer, and Mamba-based models provide similar performance for eye disease classification on the Eye Disease Image Dataset (EDID). Along with analyzing individual model performance, we further evaluate performance using a logit-level ensemble of the top three models. The ensemble results imply that these model families capture partially complementary visual features, resulting in a slight performance gain when combined. In both cases, class-wise performance examination indicates that diseases with distinctive structural characteristics are recognized more consistently, while categories with subtle or overlapping features remain challenging. Our implementation, including the training and evaluation pipelines as well as the data splits, is publicly available at https://github.com/ahanaf019/EyeDisease2026.