Multi-Scale Three-Path Network (MSTP-Net): A new architecture for retinal vessel segmentation

Abstract

In the deep learning methods, acquiring the semantic information of retinal fundus vessels relies on the feature extraction methods because of the complexity of the geometric structure. However, segmenting complete vascular structure will face challenges of the feature extraction. First, dealing with both global features and local features at the same level also depends on different methods of feature extraction. Besides, the emphasis and methods for feature extraction must be changed across different stages and levels of feature maps. Therefore, we propose MSTP-Net, the Multi-Scale Three-Path Network. It consists of a backbone and three output paths, with a new architecture. The backbone takes different components to transform the original features into high-level features. Three paths, i.e., Local Path, Global Path, and Fusion Path, take different multi-scale tasks to extract features. Local Path uses fine-grained methods to select the local detailed vascular semantics. Global Path focuses on obtaining the global vascular structure. Fusion Path combines the local features from Local Path with the global features from Global Path to extract fusion features. At last we integrate global features, local features and fusion features to obtain the final output. We evaluated our method on four retinal datasets (DRIVE, STARE, CHASE_DB1, HRF). The experiment indicates that MSTP-Net has achieved competitive performance in retinal vessel segmentation. The source code of proposed MSTP-Net is available at https://github.com/KokoloNaga/MSTP-Net.git.