Research On Prostate TRUS Image Segmentation Algorithm Based On Improved Hodge Algorithm And U~2-Net

In recent years,with the continuous improvement of the national living standard,the incidence rate of prostate diseases has risen rapidly.Prostate cancer has become the second leading cause of cancer related deaths in men.However,the quality of prostate imaging is low,and doctors may experience misdiagnosis or missed diagnosis during segmentation,and manual segmentation is time-consuming and laborious.Therefore,it is necessary to achieve fast and accurate segmentation of prostate boundaries through computer-aided implementation.Therefore,this article adopts traditional algorithms and deep learning methods to segment prostate TRUS images,and finally proposes an algorithm that combines the two methods to achieve more accurate segmentation.The main work of this article is as follows:(1)Unlike previous segmentation algorithms,this paper proposes a coarse segmentation algorithm that first locates and then uses an average template.Specifically,first train the YOLOv5 model for prostate localization detection,then extract the point set features from the training set through principal component analysis to obtain an average template.Finally,linearly transform the average template to the size of the YOLOv5 localization box and insert the box to complete the initial coarse segmentation.(2)In response to the shortcomings of many false positives and low fault tolerance in the results of Hodge algorithm,this paper proposes a precise segmentation algorithm based on the gray level difference between the inner and outer sector regions to improve the Hodge algorithm.By improving the single line segment subtraction to a sector shaped region centered on the iteration point,the improved algorithm expands the grayscale accumulation area and incorporates nearby pixel information into the segmentation results,greatly reducing false detection points in the segmentation results.(3)In response to the issue of information loss caused by downsampling operations in the U2-Net model and the reduction of useful features due to excessive contextual information obtained by modules,this paper proposes improving downsampling,adding attention modules,and adding multi-scale feature extraction modules in skip connections to improve the model’s feature expression ability.The improved model has varying degrees of improvement in indicators such as Dice,Precision,and Recall.(4)By analyzing the segmentation results of the improved Hodge algorithm and the U2-Net method,this paper proposes a method that combines the segmentation results of the two to achieve as accurate a segmentation algorithm as possible.This algorithm determines the missegmentation point by iteratively calculating the distance between the corresponding segmentation results,and then eliminates the missegmentation point through reconstruction algorithm.This method combines the advantages of traditional algorithms and deep learning,ensuring accurate pixel level segmentation when there are fewer edge noise points,and reducing segmentation errors in areas with severe noise points.The experimental results show that compared to the first two methods,this method has more balanced segmentation results,closer edges to the true contour,and fewer false positives.