Image-level multi-label retinal disease classification with a novel classification head

Automatic recognition of retinal diseases is an important step that serves to halt the disease’s progression. Considering that people can suffer from more than one disease at the same time and the need to know which eye(s) is diseased, this study focused on detecting retinal diseases from multi-label fundus images separately. In the proposed model, the image is first passed through the data augmentation step and then given to the Swin Transformer V2 backbone, which focuses on capturing the global context. The powerful features that were obtained are given to the newly developed Shunted Cross-Attention (SCA) classification head, which strengthens classification ability by preventing information loss and detecting features at different scales. The proposed model incorporates the Adaptive Sharpness-Aware Minimization (ASAM) optimizer to improve convergence ability and the Scalable Neighbor Discriminative Loss (SNDL) to effectively capture inter-label dependencies on multi-label datasets. The performance evaluations have been conducted on the publicly available Ocular Disease Intelligent Recognition dataset. Considering the final score, which is the average of Kappa, F1, and Area Under Curve scores, 87.60% and 85.11% are achieved for off-site and on-site test scenarios, respectively, which is the best in the literature. When each metric is evaluated separately, it is seen at the top in almost all of them. To further emphasize the proposed SCA classification head’s robustness, it is compared with different popular classification heads and tested with different backbones and datasets, and superior results are obtained for all scenarios.