Discriminative Features Learning for Fine-Grained Nuclear Cataract Classification in AS-OCT Images with Supervised Contrastive Learning

Currently, most methods for classifying nuclear cataracts (NC) rely on coarse-grained feature extraction techniques. This paper proposes a fine-grained classification model specifically designed to identify pathological regions with subtle differences. The model first utilizes a backbone to extract global features, followed by further processing of these global features to extract discriminative features. These regional features are then fused with the global features, and the combined features are fed into a classification head to generate classification probabilities. Additionally, a supervised contrastive learning strategy is applied to the global features. This strategy significantly improves the similarity between samples of the same class while increasing the differences between samples of different classes, thereby enhancing the representation of the global features and ultimately improving the classification performance of NC. The proposed network was tested on a local AS-OCT dataset for classification, where it demonstrated excellent performance, achieving a high classification accuracy 88.38%.