Deep Neovascularization Segmentation by Learning Graphical Connectivity in Optical Coherence Tomography

Segmentation of neovascularization (NV) within plaques from medical images can assist physicians in locating, evaluating, and diagnosing vessel lesions and plaque abnormalities. It has been demonstrated that NV serves as a crucial indicator of plaque vulnerability. Current convolutional neural network based (CNN-based) methods primarily rely on learning local appearance on a regular image grid without considering the structural characteristics of NV. To solve this problem, graph convolutional networks (GCN) are integrated into a unified convolutional architecture to effectively combine local image information with graph connection information. This integration enables us to achieve NV segmentation in Optical Coherence Tomography (OCT) images by leveraging various types of features. GCN learns vertex features on this irregular pattern so that when a vertex is surrounded by NV nodes, it becomes more likely to be labeled as NV. Simultaneously, the spatial attention mechanism is used in the coding-decoding structure to emphasize the effective feature information in the feature mapping and enhance the representation of NV-related features in the image. The proposed method achieves accurate end-to-end segmentation of NV in OCT images.