基于深度学习自动分割模型乳腺癌放疗临床应用与评价

Clinical application and evaluation of automatic segmentation model based on deep learning for breast cancer radiotherapy

目的将深度学习算法与商用计划系统整合,建立乳腺癌靶区和危及器官(OARs)自动分割平台并加以验证。方法入组在中国医学科学院肿瘤医院行保乳术后放疗的左、右乳腺癌患者各400例。基于深度残差卷积神经网络进行训练临床靶区(CTV)和OARs分割模型,建立端到端的基于深度学习的自动分割平台(DLAS)。使用42例左乳腺癌和40例右乳腺癌验证DLAS平台勾画的准确性。分别计算总体戴斯相似性系数(DSC)和平均豪斯多夫距离(AHD)。并计算相对层位置与每层DSC值(DSC_s)的关系,进行逐层分析。结果左/右乳腺癌全乳CTV平均总体DSC和AHD分别为0.87/0.88和9.38/8.71mm,左/右乳腺癌OARs平均总体DSC和AHD范围为0.86~0.97和0.89~9.38mm。对CTV和OARs进行逐层分析,达到0.90以上表示医生只需要较少修改甚至不用修改的层面,左右乳腺癌的CTV勾画占比约44.7%的层面,OARs自动勾画占比范围为50.9%~89.6%。对于DSC_s<0.7,在两侧边界区域(层位置0~0.2和0.8~1.0)CTV和除脊髓以外的感兴趣区域DSC_s值明显下降,且越靠近边缘降低程度越明显。脊髓采用全层勾画,未发现有特殊区域出现DSC_s明显下降。结论建立端到端的DLAS平台整合乳腺癌分割模型取得较好的自动分割效果。在头脚方向的两侧边界区域,勾画的一致性下降较明显,有待进一步提高。

Objective In this study,the deep learning algorithm and the commercial planning system were integrated to establish and validate an automatic segmentation platform for clinical target volume(CTV)and organs at risk(OARs)in breast cancer patients.Methods A total of 400 patients with left and right breast cancer receiving radiotherapy after breast-conserving surgery in Cancer Hospital CAMS were enrolled in this study.A deep residual convolutional neural network was used to train CTV and OARs segmentation models.An end-to-end deep learning-based automatic segmentation platform(DLAS)was established.The accuracy of the DLAS platform delineation was verified using 42 left breast cancer and 40 right breast cancer patients.The overall Dice Similarity Coefficient(DSC)and the average Hausdorff Distance(AHD)were calculated.The relationship between the relative layer position and the DSC value of each layer(DSC_s)was calculated and analyzed layer-by-layer.Results The mean overall DSC and AHD of global CTV in left/right breast cancer patients were 0.87/0.88 and 9.38/8.71 mm.The average overall DSC and AHD range for all OARs in left/right breast cancer patients were ranged from 0.86 to 0.97 and 0.89 to 9.38 mm.The layer-by-layer analysis of CTV and OARs reached 0.90 or above,indicating that the doctors were only required to make slight or no modification,and the DSC_s≥0.9 of CTV automatic delineation accounted for approximately 44.7%of the layers.The automatic delineation range for OARs was 50.9%-89.6%.For DSC_s<0.7,the DSC_s values of CTV and the regions of interest other than the spinal cord were significantly decreased in the boundary regions on both sides(layer positions 0-0.2,and 0.8-1.0),and the level of decrease toward the edge was more pronounced.The spinal cord was delineated in a full-scale manner,and no significant decrease in DSC_s was observed in a particular area.Conclusions The end-to-end automatic segmentation platform based on deep learning can integrate the breast cancer segmentation model and achieve excellent automatic segmentation effect.In the boundary areas on both sides of the superior and inferior directions,the consistency of the delineation decreases more obviously,which needs to be further improved.