GlobalLocalSegNet: A Hybrid Model for Complex Medical Image Segmentation Combining Global and Local Features

Complex medical image segmentation holds critical importance in the medical field. However, renowned segmentation models such as U-Net [1] and TransUNet [2] often suffer from diminished performance when dealing with images that have complex backgrounds or high variability. These models also struggle to capture the dependencies between global and local features effectively, limiting their application in complex medical image segmentation tasks. To address these shortcomings, this paper introduces a novel U-shaped model, GlobalLocalSegNet (GLS-Net), specifically designed for precise segmentation of complex medical images. The GLS-Net model comprises two key sub-modules: a transformer model with positional encoding to extract global information, and a convolutional neural network tailored for local information extraction. Moreover, a U-shaped network structure based on the fusion of global and local features was designed to enhance feature extraction and detail capture. The model was tested using three publicly available complex medical image datasets and one non-medical complex image dataset to assess its scalability. A series of comparative experiments in this study confirm the robustness, scalability, and stability of the GLS-Net model.