危及器官自动勾画在鼻咽癌、乳腺癌与直肠癌中应用研究

Clinical evaluation of automatic commercial software for intensity modulation radiotherapy in nasopharyngeal carcinoma, breast cancer and rectum cancer

目的 评价自动勾画与手动勾画软件在危及器官(OAR)勾画方面差异,研究OAR剂量学参数的变化。方法 选择在放射治疗科行调强放射治疗患者33例,其中男性11例,女性22例;年龄38~81岁,平均年龄63.4岁;鼻咽癌13例,乳腺癌10例,直肠癌10例。在同一定位CT序列图像分别进行手动勾画与自动勾画。首先使用AccuContour软件评估自动勾画结构与手动勾画的相似性指数,然后将自动勾画与手动勾画的OAR结果回传至Pinnacle 9.8计划系统,将以手动勾画为参考的剂量分布复制到自动勾画靶区上,评价手动勾画与自动勾画OAR的体积剂量和剂量体积等参数的变化。结果 鼻咽癌、乳腺癌和直肠癌平均手动勾画的时间分别为(56.50±9.00) min、(23.12±4.23) min和(45.23±2.39) min;AccuContour自动勾画的平均时间为(1.5±0.23) min、(1.45±0.78) min和(1.80±0.56) min。鼻咽癌中,眼球勾画获得最佳Dice相似系数(DSC)为0.907±0.020,脊髓获得最差的DSC为0.459±0.112;乳腺癌中,所有OAR包括肺、心脏与脊髓均获得很好勾画效果(DSC> 0.7),其中肺的勾画最佳DSC为0.944±0.030,最差勾画脊髓DSC为0.709±0.100;直肠癌中,OAR膀胱勾画获得最佳DSC为0.91±0.04,而股骨头的勾画最差,DSC为0.43±0.10。尽管大部分器官都得出很好的自动勾画效果,然而自动勾画的剂量学参数却有重要的差别,勾画效果较差的小体积器官,如晶状体、视神经等剂量学参数差异无统计学意义(P> 0.05),而勾画效果较好的大体积器官,如剂量学参数差异有统计学意义(P <0.05)。结论 基于深度学习的自动勾画方法具有很高临床研究价值,然而DSC的值不能完全反映剂量分布。膀胱、肺等边界清楚的器官可以采用自动勾画,而对于股骨头、脊髓和脑干等一些依赖于人工经验及勾画习惯的可以采用半自动勾画的方式,自动勾画不仅为临床应用节省了大部分的勾画时间,也对临床手动勾画起到监督的作用。

Objective To compare the differences of manual contour and automatic contour system AccuContour in automatic contour for organ at risk(OAR) and probe the parameters of dose volume histogram(DVH) in radiotherapy. Methods A total of33 patients performed intensity-modulated radiotherapy were enrolled, which included 11 males and 22 females, aged 38-81years old with mean age of 63.4 years old. They were 13 cases of nasopharyngeal carcinoma, 10 of breast cancer and 10 of rectal cancer. The manual contour and automatic contour were performed on the CT sequence images at the same location.First, the AccuContour software was used to evaluate similarity index between automatic and manual contour, and then OAR results of automatic and manual contour were transferred to Pinnacle 9.8 planning system. As reference, the dose distribution of manual delineation was copied to automatic contour target volume, and changes of OAR volume dose and dose-volume parameter were evaluated between 2 methods. Results The mean manual contour time for nasopharyngeal carcinoma, breast cancer and rectal cancer was(56.50 ± 9.00) minutes,(23.12 ± 4.23) minutes and(45.23 ± 2.39) minutes;the mean times for AccuContour automatic contour were(1.5 ± 0.23) minutes,(1.45 ± 0.78) minutes and(1.80 ± 0.56) minutes. In nasopharyngeal carcinoma, the best Dice similarity coefficient(DSC) for eyeball contour was 0.907 ± 0.020, and the worst DSC for spinal cord was 0.459 ± 0.112. In breast cancer, all OAR including lung, heart and spinal cord were well contoured(DSC > 0.7), of which the best DSC of lung was 0.944 ± 0.030, and the worst DSC of spinal cord was 0.709 ± 0.100. In rectal cancer, OAR bladder contoured the best DSC of 0.91 ± 0.04, while the femoral head contoured the worst DSC of 0.43 ± 0.10. Though most organs showed good automatic contoured effect, the dosimetry parameters were significantly different. There was no statistically significant difference in dosimetry parameters such as lens and optic nerve for small-volume organs with poor contoured effect(P > 0.05),while large-volume organs with better contoured had statistically significant differences in dosimetry parameters(P < 0.05). Conclusion It is demonstrated that the automatic contour approach based deep learning method displays sufficient accuracy for research purposes, but the value of DSC cannot fully reflect the dose distribution. The organs with clear boundaries, such as bladder and lungs could be delineated automatically, while those rely on manual experience and delineation habits, such as femoral head,spinal cord and brain stem could be delineated semi-automatically. The automatic contour not only saves most contour time for clinic, but also supervises manual contour.