基于深度学习的多模态影像脑胶质母细胞瘤放疗靶区的自动勾画研究

目的 通过建立深度学习模型,探索多模态影像对脑胶质母细胞瘤放疗靶区自动勾画效果的影响。方法 收集30例脑胶质母细胞瘤患者的电子计算机断层扫描(CT)序列和磁共振成像(MRI)的对比增强T1加权序列(T1C)以及T2液体衰减反转恢复序列(T2-FLAIR),对每例病例的原发肿瘤靶区(GTV)及其对应的临床靶区1(CTV1)和临床靶区2(CTV2)根据RTOG标准进行人工勾画,并设计4种不同的数据集:CT数据集(仅含30例CT序列的单模态数据)、CT-T1C数据集(包含30例CT序列和T1C序列的双模态数据)、CT-T2-FLAIR数据集(包含30例CT序列和T2-FLAIR序列的双模态数据)和CT-MRIs数据集(包含30例CT序列、T1C序列和T2-FLAIR序列的三模态数据)。使用每种数据集中的25例对改进后的3D U-Net进行训练,并用剩余5例进行测试。评价测试样本中GTV、CTV1和CTV2的自动勾画效果,定量评估指标包括Dice相似系数(DSC),95% Hausdorff距离(HD95)和相对体积误差(RVE)。结果 该3D U-Net模型在多模态影像CT-MRIs上获得最好的GTV自动分割结果,与在单模态影像CT的自动分割结果相比(DSC: 0.94 vs. 0.79, HD95: 2.09 mm vs. 12.33 mm and RVE: 1.16% vs. 20.14%),DSC(t=3.78,P<0.05)和HD95 (t=4.07, P<0.05)的差异有统计学意义;在多模态影像CT-MRIs的CTV1和CTV2自动分割结果(DSC: 0.90 vs. 0.91, HD95: 3.78 mm vs. 2.41 mm, RVE: 3.61% vs. 5.35%)也均有较好的一致性,但与单模态影像CT的自动分割结果相比,两个靶区的DSC和HD95的差异均无统计学意义(P>0.05)。该模型对于GTV的上下界和CTV2临近的重要器官(如脑干和眼球)的自动勾画有一定的局限性。结论 基于改进后的3D U-Net在多模态影像数据集CT-MRIs上对脑胶质母细胞瘤放疗靶区具有更好的分割效果,显示出较好的临床应用价值。

Application of deep learning-based multimodal imaging to the automatic segmentation of glioblastoma targets for radiotherapy

Objective To explore the effects of multimodal imaging on the performance of automatic segmentation of glioblastoma targets for radiotherapy based on a deep learning approach.Methods The computed tomography (CT) images and the contrast-enhanced T1 weighted (T1C) sequence and the T2 fluid attenuated inversion recovery (T2-FLAIR) sequence of magnetic resonance imaging (MRI) of 30 patients with glioblastoma were collected.The gross tumor volumes (GTV) and their corresponding clinical target volumes CTV1 and CTV2 of the 30 patients were manually delineated according to the criteria of the Radiation Therapy Oncology Group (RTOG).Moreover,four different datasets were designed,namely a unimodal CT dataset (only containing the CT sequences of 30 cases),a multimodal CT-T1C dataset (containing the CT and T1C sequences of 30 cases),a multimodal CT-T2-FLAIR dataset (containing the CT and T2-FLAIR sequences of the 30 cases),and a trimodal CT-MRI dataset (containing the CT,T1C,and T2-FLAIR sequences of 30 cases).For each dataset,the data of 25 cases were used for training the modified 3D U-Net model,while the data of the rest five cases were used for testing.Furthermore,this study evaluated the segmentation performance of the GTV,CTV1,and CTV2 of the testing cases obtained using the 3D U-Net model according to the indices including Dice similarity coefficient (DSC),95% Hausdorff distance (HD95),and relative volume error (RVE).Results The best automatic segmentation result of GTV were achieved using the CT-MRI dataset.Compared with the segmentation result using the CT dataset (DSC:0.94 vs. 0.79,HD95:2.09 mm vs. 12.33 mm,and RVE:1.16%vs. 20.14%),there were statistically significant differences in DSC (t=3.78,P<0.05) and HD95(t=4.07,P<0.05) obtained using the CT-MRI dataset.Highly consistent automatic segmentation result of CTV1 and CTV2 were also achieved using the CT-MRI dataset (DSC:0.90 vs. 0.91,HD95:3.78 mm vs. 2.41 mm,RVE:3.61%vs. 5.35%).However,compared to the CT dataset,there were no statistically significant differences in DSC and HD95 of CTV1 and CTV2(P>0.05).Additionally,the 3D U-Net model yielded some errors in predicting the upper and lower bounds of GTV and the adjacent organs (e.g.,the brainstem and eyeball) of CTV2.Conclusions The modified 3D U-Net model based on the multimodal CT-MRI dataset can achieve better segmentation result of glioblastoma targets and its application potentially benefits clinical practice.