Research on Image Segmentation Method Based on Multi-Scale Feature Fusion and Dual Attention

Research on Image Segmentation Method Based on Multi-Scale Feature Fusion and Dual Attention

This paper proposes a multi-scale attention fusion mechanism for automatic gland segmentation based on the colorectal adenocarcinoma cell dataset, aiming to address the issue of unclear cell adhesion and confusion with the background in colorectal adenocarcinoma cell instance segmentation. Deep learning methods are employed for top-down gland cell instance segmentation. The neural network features a novel spatial-pyramid dual-path attention module that not only integrates multi-dimensional feature map spatial information but also enriches feature space through cross-dimensional feature information interaction. With the assistance of the new fusion module, it can perceive higher resolution features effectively fuse multi-scale features, leading to higher segmentation accuracy, stronger robustness, and generalization. It demonstrates excellent performance on the GlaS and CRAG datasets.