基于3D ResSE-Unet的智能靶区勾画在乳腺癌术后放疗中的应用研究

Application of 3D ResSE-Unet-based intelligent delineation of clinical target volume in postoperative radiotherapy for breast cancer

目的评估基于3D ResSE-Unet的智能靶区勾画在乳腺癌术后辅助放疗靶区勾画中应用的有效性及可行性。方法选取2018年9月至2022年6月在广西医科大学第四附属医院肿瘤诊疗中心治疗的乳腺癌术后辅助放疗病例974例,其中全乳切除术后614例,保乳术后360例。分别设置:训练集874例,用于建立基于3D ResSE-Unet智能靶区勾画模型;验证集40例,用于配比评估人工智能乳腺癌放疗靶区设计临床应用的可行性及有效性;测试集60例,用于测试智能靶区的准确性。比较智能靶区勾画模型的戴斯相似系数(DSC),95%豪斯多夫距离(HD95),平均表面距离(ASD)。结果该智能勾画模型精确度较高,全乳切除术后的临床靶区CTVcw的DSC均>0.80,保乳术后的临床靶区CTVb的DSC均>0.88。保乳术后临床靶区CTVb与全乳切除术后临床靶区CTVcw相比,保乳术后CTVb具有更高的DSC(0.91±0.03 vs.0.83±0.05,t=7.11,P<0.05);而两种靶区的HD95值[(7.56±3.42)vs.(8.77±5.89)mm]及ASD值[(1.85±0.71)vs.(1.86±0.83)mm]均较低,差异无统计学意义(P>0.05)。双侧锁骨上下区靶区(CTV2)的DSC均>0.8;双侧CTV2的HD95、ASD均值均较小,差异无统计学意义(P>0.05)。结论基于3D ResSE-Unet的智能靶区勾画在乳腺癌术后辅助放疗中有较好的勾画一致性及可行性,尤其是保乳术后的靶区。

Objective To evaluate the effectiveness and feasibility of 3D ResSE-Unet-based intelligent delineation of clinical target volume(CTV)in postoperative adjuvant radiotherapy for breast cancer.Methods A total of 974 cases of breast cancer treated in the Cancer Diagnosis and Treatment Center of the Fourth Affiliated Hospital of Guangxi Medical University from September 2018 to June 2022 were enrolled in this study,including 614 cases receiving total mastectomy and 360 cases treated with breast-conserving surgery.They were divided into a training set,a validation set,and a testing set.The training set consisted of 874 cases and was used to build a model of 3D ResSE-Unet-based intelligent CTV delineation.The validation set comprised 40 cases and was used to evaluate the feasibility and effectiveness of the clinical application of AI-based CTV design in the radiotherapy for breast cancer.The testing set was composed of 60 cases and was used to test the accuracy of intelligent CTV.The Wilcoxon rank test was used to compare the Dice similarity coefficient(DSC),95%Hausdorff distance(HD95),and average surface distance(ASD)obtained using the intelligent delineation model.Results The intelligent delineation model showed high precision.The CTV of cases treated with total mastectomy(CTVcw)and the CTV of cases treated with breast-conserving surgery(CTVb)had DSCs greater than 0.80 and greater than 0.88,respectively.Therefore,compared with CTVcw,CTVb had a higher DSC(0.91±0.03 vs.0.83±0.05,t=7.11,P<0.05).Both CTVcw and CTVb had lower HD 95[(7.56±3.42)mm vs.(8.77±5.89)mm]and ASD[(1.85±0.71)mm vs.(1.86±0.83)mm],without statistically significant difference(P>0.05).The left/right supraclavicular and infraclavicular CTV(CTV2)had DSCs greater than 0.8.CTV2 also had low average HD95 and ASD,without statistically significant difference(P>0.05).Conclusions The 3D ResSE-Unet-based intelligent CTV delineation has better consistency and feasibility in postoperative adjuvant radiotherapy for breast cancer,especially the CTVs after breast-conserving surgery.