URV-Net: multihead receptive field-based U-network and ViT conjugate medical image segmentation models

Medical image segmentation refers to segmenting pathological regions in medical images. It is an essential step in medical image analysis and an important technical prerequisite for doctors to determine the quantitative extent of pathological changes and treatment plans.In the present paper, we introduce a medical image segmentation method utilizing URV-Net, aiming to address the challenges of ambiguous boundary delineation, limited feature information extraction, and insufficient segmentation precision encountered in current medical image segmentation techniques.The model of this method consists of three major feature extraction modules, namely, U-Net+SE enhanced encoder, multihead sensory field feature extraction module (RFB) with feature residual optimization module (RRM), and ViT encoder with non-local self-attention module (Non-local). Using multipath input, different levels of features extracted from different coding blocks are complemented and optimized with each other by the feature optimization module, which improves the model's ability to capture global dependencies and capture image features from coarse to fine, simultaneously, it enhances the adaptive learning ability of the importance of feature channels. By applying our proposed model, we conducted the training, validation, and testing phases of the model on the GlaS 2015 dataset, and performed comparative experiments with several other deep learning models. Analyzing the experimental result data, the intersection and merger ratio IoU increased by about 10% on average, the Dice coefficient increased by about 5% on average, the accuracy (Acc) increased by about 6.9% on average, and the precision (Prec) increased by about 5.4% on average, and by comparing the visual segmentation results, the performance of the segmentation model proposed in this chapter is superior to other comparative models in all aspects. Furthermore, ablation experiments have also demonstrated the improvements brought by the combination of different modules in the model to the overall model performance. The code for the model proposed in this article is available at https://github.com/manreren/URV-Net.