Local Detail Enhancement Network for CNV Typing in OCT Images

Choroidal neovascularization (CNV) is one of the severe eye disease. The severe results will cause of loss of acuity, scotomata, and distortion of vision. Automatic and accurate classification of CNV with optical coherence tomography (OCT) images can assist doctors in treatment. However, the existing methods ignore the fact that semantic feature maps, used for classification, lose much feature detail information. Therefore, we proposed a local detail enhancement network for CNV classification, which includes both progressive training mode and local detail enhancement (LDE) module. With the progressive training mode, the learned features fuse shallow and stable fine-grained information with high-level semantic information, which promote the diversity of the learned features. In LDE module, the detail feature learn (DFL) module is introduced to learn the underlying detail information and embed it into the semantic feature map. The semantic feature map with detail information is propitious to capture the subtle discrepancy between different CNV types and promote the classification performance. Sufficient experiments are performed on our self-build CNV dataset. Our method excelled existing methods and in evaluation indicators ACC, AUC, SEN, and SPE are 92.3%, 87.1%, 91.5%, and 90.9%.