基于深度学习和图谱库方法自动勾画肿瘤放疗中危及器官的比较

【摘要】目的:评估比较基于深度学习（DL）和图谱库（Atlas）方法自动勾画不同部位肿瘤放疗中危及器官（OARs）轮廓的几何学精度，为临床应用提供依据。方法:选择40例肿瘤患者的CT图像（头颈部、胸部、腹部和盆腔肿瘤患者各10例），由资深放射治疗医师手动勾画OARs，然后再分别使用基于DL和Atlas方法的自动勾画软件勾画OARs。采用形状相似性指数（DC）、Jaccard系数（JC）、Hausdorff距离（HD）、体积差异（VD）等多个指标评价基于DL和Atlas自动勾画与手动勾画OARs的几何学一致性。结果:除直肠外，采用DL方法勾画的多数OARs的DC指标高于0.7，优于Atlas方法，差异有统计学意义（P<0.05）。此外，DL方法的JC值除晶体、直肠、脊髓外也都大于0.7。HD中最大的是脊髓，两种方法均超过20 mm。DL方法中VD绝对值较大的是直肠。结论:基于DL方法自动勾画的OARs几何精确度总体上高于Atlas方法。下一步，通过继续增大训练集的数据量可进一步提高基于DL方法模型的鲁棒性，从而更好地辅助放射肿瘤医师，使肿瘤患者获益。

Comparative study of deep learning- versus Atlas-based auto-segmentation of organs-at-risk in tumor radiotherapy

Abstract: Objective To evaluate and compare the geometric accuracy between deep learning (DL)- and Altas-based auto-segmentation technologies for contouring organs-at-risk (OARs) in radiotherapy for tumors locating in different sites so as to provide a basis for the clinical application. Methods The OARs in CT images of 40 patients with tumors in different sites (head and neck, thorax, abdomen, and pelvic cavity) were manually segmented by senior physicians, and then automatically segmented by DL- and Atlas-based auto-segmentation methods. Several evaluation indicators such as Dice coefficient (DC), Jaccard coefficient (JC), Hausdorff distance (HD) and volume difference (VD) were used to evaluate the geometric accuracy between DL- or Atlas-based auto-segmentations and manual segmentation. Results The DC values of OARs except for rectum segmented by DL-based method were higher than 0.7, higher than the results obtained by Atlas-based method, with statistical significance (P<0.05). In addition, the JC values obtained by DL-based method were also higher than 0.7, except for the JC values of lens, rectum and spinal cord. Spinal cord had the highest HD value, exceeding 20 mm in both methods. The rectum segmented by DL method had relatively high absolute VD. Conclusion The geometric accuracy of DL-based auto-segmentation is generally superior to that of Atlas-based auto-segmentation. In the further study, the robustness of DL model will be increased by expanding the training dataset, thereby better assisting radiation oncologists in routine clinical work and bringing benefits to tumor patients.