MHT-Net: A Matching-Based Hierarchical Transfer Network for Glaucoma Detection From Fundus Images

Glaucoma is a chronic and irreversible eye disease. Early detection and treatment can effectively prevent severe consequences. Deep transfer learning is widely used in fundus imaging analysis to remedy the shortage of training data of glaucoma. The model trained on the source domain may struggle to predict glaucoma in the target domain due to distribution differences. Several limitations cannot be ignored: (1) Image matching: enhancing global and local image consistency through bidirectional matching; (2) Hierarchical transfer: developing a strategy for transferring different hierarchical features. To this end, we propose a novel Matched Hierarchical Transfer Network (MHT-Net) to achieve automatic glaucoma detection. We initially create a fundus structure detector to match global and local images using intermediate layers of a pre-trained diagnostic model with source domain data. Next, a hierarchical transfer network is implemented, sharing parameters for general features and using a domain discriminator for specific features. By integrating adversarial and classification losses, the model acquires domain-invariant features, facilitating precise and seamless transfer of fundus information from source to target domains. Extensive experiments demonstrate the effectiveness of our proposed method, outperforming existing glaucoma detection methods. These advantages endow our algorithm as a promising efficient assisted tool in the glaucoma screening.