摘要
Abstract
Objective:To construct a deep learning model based on YOLOV8-Seg algorithm to conduct automatic segmentation for the central gland(CG)and peripheral zone(PZ)of prostate,so as to provide a reliable basis for clinical diagnosis and treatment.Methods:The sequence data of T2-weighted imaging(T2WI)of horizontal relaxation time of 158 patients were selected from a public data set of magnetic resonance imaging(MRI)for prostate MRI,which was provided by the Charité University Hospital in Berlin,were selected.The all data were divided into a training set(109 cases),a validation set(16 cases),and a test set(33 cases)as the ratio of 7 to1 to 2.A lightweight asymmetric decoupled head(LADH)structure and the large kernel UniRepLKNetBlock module were integrated into the YOLOV8-Seg algorithm to enhance the capabilities of model's extraction feature,and the new model was named as YOLOV8-URLK.The assessment model with mean Average Precision(mAP),Dice Similarity Coefficient(DSC),95%Hausdorff Distance(HD95),and Average Surface Distance(ASD)was adopted to segment performance of the detection at prostate CG and PZ.Comparative experiments were conducted among that and YOLOV8-Seg,TransU-Net,and U-Net network,so as to validate the effectiveness of YOLOV8-URLK for detection and segmentation at prostate zone.Results:On the test set,the mAP@0.5(box)of YOLOV8-URLK model was 0.878,and the mean Dice coefficients,the mean HD95 values and the ASD values of that at CG and PZ were respectively(0.867,17.123 and 1.461)and(14.902,0.898 and 1.112).On the test set,the mAP@0.5(box)of YOLOV8-Seg model was 0.860,and the mean Dice coefficients of that at CG and PZ were 0.851 and 0.884,the mean HD95 values of that at them were 19.174 and 15.298,and ASD values of that at them were 1.781 and 1.219,respectively.On test set,the mean Dice coefficients of TransU-Net model at CG and PZ were 0.864 and 0.824,and the mean HD95 values of that at them were 18.134 and 19.402,and ASD values of that at them were 1.698 and 1.717,respectively.On the test set,the mean Dice coefficients of the U-Net model at CG and PZ were 0.857 and 0.690,and the mean HD95 values of that at them were 18.976 and 26.934,and ASD values of that at them were 1.753 and 2.135.The YOLOV8-URLK model can better reappear the segmentation trend of manual annotations.Conclusion:The YOLOV8-URLK model demonstrates higher precision in the detection and segmentation of MRI images of prostate,which were superior to YOLOV8-Seg,TransU-Net and U-Net.It can enhance the efficiency of the detection and segmentation.关键词
前列腺/磁共振成像(MRI)/YOLO算法/自动分割Key words
Prostate/Magnetic resonance imaging(MRI)/You Only Look Once(YOLO)algorithm/Automatic segmentation分类
医药卫生
