Abstract
BACKGROUND: With optical coherence tomography (OCT), doctors are able to see cross-sections of the retinal layers and diagnose retinal diseases. Computer-aided diagnosis algorithms such as convolutional neural networks (CNNs) and vision Transformers (ViTs) enhance diagnostic efficiency by automatically analyzing these OCT images. However, CNNs are less effective in extracting global features and ViTs lack the local inductive bias and typically require large amounts of training data. METHODS: In this paper, we presented a hybrid retinal diseases classification and diagnosis network named HyReti-Net which incorporated two branches. One branch extracted local features by leveraging the spatial hierarchy learning capabilities of ResNet-50, while the other branch was established based on Swin Transformer to consider the global information. In addition, we proposed a feature fusion module (FFM) consisting of a concatenation and residual block and the improved channel attention block to retain local and global features more effectively. The multi-level features fusion mechanism was used to further enhance the ability of global feature extraction. RESULTS: Evaluation and comparison were used to show the advantage of the proposed architecture. Five metrics were applied to compare the performance of existing methods. Moreover, ablation studies were carried out to evaluate their effects on the foundational model. For each public dataset, heatmaps were also generated to enhance the interpretability of OCT image classification. The results underscored the effectiveness and advantage of the proposed method which achieved the highest classification accuracy. CONCLUSION: In this article, a hybrid multi-scale network model integrating dual-branches and a features fusion module was proposed to diagnose retinal diseases. The performance of the proposed method produced promising classification results. On the OCT-2014, OCT-2017 and OCT-C8, experimental results indicated that HyReti-Net achieved better performance than the state-of-the-art networks. This study can provide a reference for clinical diagnosis of ophthalmologists through artificial intelligence technology.