残差U-net与先验知识协同模型自动勾画宫颈癌术后临床靶区和危及器官北大核心CSCD

目的搭建残差U-net(RU)网络与先验知识协同(RPKC)自动勾画模型,评估其自动勾画宫颈癌术后患者临床靶区(CTV)和危及器官(OAR)的准确性。方法基于48例(训练集)宫颈癌术后定位CT训练RPKC模型。以临床医师勾画的CTV及OAR为标准,采用戴斯相似系数(DSC)和第95百分位豪斯多夫距离(HD95)评估RPKC模型与RU模型勾画另20例宫颈癌术后患者(测试集)CTV及OAR(包括肠袋、直肠、膀胱、骨盆及双侧股骨头)的准确性。结果RPKC模型自动勾画上述结构的DSC均高于RU模型,其中CTV及肠袋勾画效果差异有统计学意义(P均<0.05);除直肠外,RPKC模型自动勾画的HD95均低于RU模型,二者勾画CTV效果差异差异有统计学意义(P<0.05)。结论RPKC模型能更准确地勾画宫颈癌术后CTV和OAR,有助于提高深度学习自动勾画的临床实用性。

Residual U-net and prior knowledge collaboration model for automatic segmentation of clinical target volume and organ at risk after resection of cervical cancer

Objective To build a residual U-net (RU) and prior knowledge collaboration (RPKC) automatic segmentation model, and to evaluate its accuracy of automatic segmentation of clinical target volume (CTV) and organ at risk (OAR) for patients with postoperative uterine cervical neoplasms. Methods Localization CT images of 48 cases with postoperative uterine cervical neoplasms (training set) were trained to build RPKC model. Based on CTV and OAR delineated by a physician, the accuracy of CTV and OAR (including bowel bag, rectum, bladder, pelvis and bilateral femoral head) delineated with RPKC and RU of another 20 cases with postoperative uterine cervical neoplasms (test set) were evaluated according to Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). Results For the above structures, the DSC of RPKC was higher than that of RU, especially DSC of RPKC for CTV and bowel bag were significantly improved (both P<0.05). Except for the rectum, the HD95 of RPKC of the above structures were all lower than that of RU, and the difference of CTV was significant (P<0.05). Conclusion RPKC could automatically delineate CTV and OAR accurately for patients with postoperative uterine cervical neoplasms, hence improving the clinical practicality of automatic segmentation based on deep learning.