| Medical image segmentation is an important step in medical image analysis,which involves segmenting lesion areas in medical images.It is a crucial technique for doctors to determine the degree of lesion quantification and treatment plans.Deep learning-based methods are currently among the most efficient algorithms for medical image segmentation,among which the U-Net network model has become the mainstream network model in the field of medical image segmentation due to its symmetric structure and innovative skip-connection design.However,traditional U-Net network models still have problems such as insensitivity to fine feature attention and parameter redundancy,and their segmentation accuracy still needs to be improved.This article focuses on improving the accuracy and applicability of the U-Net network model for medical image segmentation.The main research contents and innovative work are as follows:1.A PS-UNet model is proposed to address problems such as lack of continuity in medical image segmentation results,confusion in segmentation,blurred boundary segmentation,and incomplete segmentation.PS-UNet can achieve both coarse segmentation of organ contours and fine segmentation of retinal vessels and cells.PS-UNet adds residual blocks to the U-Net to speed up network convergence and avoid overfitting.PS-UNet also includes an enhanced specific semantic segmentation module that locates and captures extensive deep information,reducing the feature blur caused by continuous convolution operations.The enhanced specific semantic segmentation module consists of a position-channel attention module and a spatial pyramid pooling module.PS-UNet is tested on publicly available datasets for segmentation of lungs,retinal vessels,and cells,among other parts.Experimental results show that PS-UNet outperforms other advanced network models,with improvements in accuracy and sensitivity for lung segmentation relative to the U-Net network model of 2.03% and 2.24%,respectively,and a Dice similarity coefficient of97.16%.2.A FM-UNet model is proposed to address problems such as parameter explosion and insufficient multi-scale feature-extraction in medical image segmentation network models that pursue accuracy.FM-UNet’s encoder consists of a feature-extraction module and a multi-scale attention module to capture richer multi-scale features and reduce model complexity.The feature-extraction module consists of depth-separable convolution,ACON-C activation function,and residual connection,while the multi-scale attention module consists of a multi-scale channel spatial attention module.The decoder uses ACON-C activation function instead of Re LU activation function,which allows each neuron to adaptively select whether to be activated,which helps to improve model transfer performance and generalization.Experimental results on publicly available datasets for retinal vessels,lungs,and cells show that FM-UNet can better capture multi-scale features,with segmentation performance superior to other advanced models,with accuracy reaching 96.63%,98.98%,and 92.42%,respectively.In summary,the proposed PS-UNet and FM-UNet network models have good performance in medical image segmentation,providing theoretical support and technical assistance for disease prevention and diagnosis,and providing feasible reference value for medical researchers. |
