Explainable topology-aware frequency-semantic transformer for robust multi-class retinal OCT diagnosis

The diagnosis of retinal disease from optical coherence tomography (OCT) images has recently received a lot of attention because it is possible to detect abnormalities which can threaten the vision at an early stage. The major hurdle in this task, however, is the subtle difference among intra-classes, the similarities between retinal diseases, and complex anatomical deformations, which make the task of accurate multi-class retinal OCT classification difficult. This research suggests a frequency-semantic transformer framework for robust multi-class retinal OCT diagnosis, which is topology-aware. The designed solution combines spatial semantic learning, frequency-aware representation extraction and topology-preserving structural modeling in order to incorporate the local retinal violations as well as the global anatomic continuity. To improve semantic coordination between fine-grained retinal textures and large-scale structural representations, a cross-scale transformer interaction module is proposed. In addition, an 'explainability guided attention' module is also incorporated to enhance the localization of lesions and interpretability of the model. Experiments were conducted on the Retinal OCT-C8 dataset comprising 24,000 OCT images across eight retinal disease categories. The proposed framework achieved a classification accuracy of 97.86%, precision of 97.53%, recall of 97.21%, F1-score of 97.36%, and an AUC of 99.02%, outperforming state-of-the-art CNN- and Transformer-based models. Specifically, the proposed method improved classification accuracy by 1.48% over the strongest CNN–Transformer hybrid baseline and by 5.02% over ResNet50. These results demonstrate that topology-aware frequency-semantic learning substantially enhances retinal disease discrimination, robustness, cross-domain generalization, and explainability for complex multi-class retinal OCT diagnosis.