基于MRI图像的自动勾画技术在宫颈癌放疗中的 应用研究

目的:研究AccuLearning和AccuContour软件在宫颈癌患者MRI图像上进行模型构建和自动勾画的可行性和准确性。方法:随机选取某院放疗科的40例宫颈癌患者,并随机选取其中30例宫颈癌患者的MRI图像,在Accu-Learning中按照推荐参数进行模型训练、验证和测试,统计训练模型和测试病例的平均相似性系数(Dice)。将训练模型导入AccuContour,并输入其余10例患者的MRI图像,以手动勾画为参考,统计自动勾画和手动勾画的平均Dice值、交叉指数(overlap index,OI)以及绝对体积差异。采用SPSS 25.0软件进行统计学分析。结果:训练模型的综合Dice值为0.80,各测试病例的Dice值分别为0.87、0.78和0.75,训练结果满足临床要求。自动勾画与手动勾画膀胱的Dice值和OI值分别为0.91±0.07、0.96±0.03,双侧股骨头的Dice值和OI值分别为左侧0.94±0.02、0.99±0.01,右侧0.91±0.04、0.99±0.01,直肠的Dice值和OI值分别为0.80±0.07、0.97±0.03,乙状结肠的Dice值和OI值分别为0.47±0.14、0.93±0.05。膀胱和乙状结肠的自动勾画绝对体积大于手工勾画,双侧股骨头和直肠的自动勾画绝对体积小于手工勾画。除右侧股骨头外,其余危及器官的自动勾画与手工勾画的绝对体积差异均无统计学意义。结论:基于AccuLearning的小样本训练模型训练效果较好,基于小样本训练模型采用AccuContour进行自动勾画具有临床可行性,可以提高宫颈癌危及器官勾画的质量和效率。

Research on application of automatic contouring based on MRI image in cervical cancer

Objective To investigate the feasibility and accuracy of AccuLearning and AccuContour software for modeling and automatic contouring based on cervical cancer MRI images.Methods Forty cervical cancer patients from the radiotherapy department of some hospital were selected randomly.Then based on the MRI images of 30 selected patients model training,verification and testing with the recommended parameters were carried out in AccuLearning,and the mean Dice similarity coefficient(Dice)of the training model and the selected patients was calculated.The training model and MRI images of the remaining 10 patients were imported into AccuCountour,and the mean Dice values,overlap index(OI)and absolute volume differences between automatic and manual contouring were counted using manual outlining as a reference.SPSS 25.0 software was used for statistical analysis.Results The comprehensive Dice value for the training model was 0.80 and the Dice values for the patients were 0.87,0.78 and 0.75 respectively,and the training results met the clinical requirements.The Dice and OI values of automatic and manual contouring were(0.91±0.07)and(0.96±0.03)respectively for the bladder,(0.94±0.02)and(0.99±0.01)for the left femoral head,(0.91±0.04)and(0.99±0.01)for the right femoral head,(0.80±0.07)and(0.97±0.03)for the rectum and(0.47±0.14)and(0.93±0.05)for the sigmoid colon.The absolute volumes by automatic contouring was bigger than those by manual contouring for the bladder and sigmoid colon,while smaller for the femoral heads and rectum.Except the right femoral head,all the organs at risk(OARs)had no statistically significant differences between the absolute volumes by automatic and manual contouring.Conclusion The AccuLearning-based small-sample training model facilitates automatic contouring by AccuContour for the cervical cancer OARs.