3D U-Net深度学习模型基于盆腔T2WI自动分割盆腔软组织结构

目的 评估3D U-Net深度学习(DL)模型基于盆腔T2WI自动分割盆腔软组织结构的可行性.方法 回顾性分析147例经病理证实或盆腔MRI随访观察确诊的前列腺癌或良性前列腺增生患者,其中28例接受2次、121例接受1次盆腔MR扫描,共175组T2WI;手动标注T2WI所示软组织结构,包括前列腺、膀胱、直肠、双侧精囊腺...

3D U-Net deep learning model for automatic segmentation of pelvic soft tissue structures based on pelvic T2WI

Objective To explore the feasibility of 3D U-Net deep learning (DL) model for automatic segmentation of pelvic soft tissue structures based on pelvic T2WI. Methods Pelvic MRI of 147 patients with pathologically or pelvic MRI follow-up confirmed prostate cancer or benign prostatic hyperplasia were analyzed retrospectively, including 28 patients underwent 2 times and 121 underwent one time MR scanning, and 175 groups of T2WI were obtained. The pelvic soft tissue structures, i.e. prostate, bladder, rectum, bilateral seminal vesicles, urethra, bilateral obturator muscles and bilateral puborectalis on T2WI were manually labeled. Then the data were divided into training set (n=137), validation set (n=21) and test set (n=17) at the ratio of 8:1:1, and the 3D U-Net segmentation model was trained. Taken manual annotation results as the reference standards, the segmentation performance of 3D U-Net DL model was evaluated according to differences of indexes including Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), precision (PRE), recall (REC), accuracy (ACC) and segmentation volume in test set. Results DSC and JSC of 3D U-Net DL model segmentation of pelvic structures in test set were both >0.90,so were ACC, PRE and REC. The segmentation volumes using 3D U-Net DL model were not statistically different with those of manual annotation (all P>0.05). Conclusion 3D U-Net DL model could be used for automatic segmentation of pelvic soft tissue structures on T2WI.