【作者】 周洁；

【导师】 许峰； 沈亚丽；

【作者基本信息】 四川大学 ， 放射肿瘤学（专业学位）， 2022， 硕士

【摘要】 目的:构建一种基于深度学习的自动勾画方法适用于多种盆腔恶性肿瘤的盆腔放疗淋巴引流区勾画,以期提高盆腔肿瘤放疗靶区勾画的规范性和准确性,并有效提高靶区勾画效率。材料和方法:回顾性收集160例经病理证实为盆腔原发恶性肿瘤(包括直肠癌、前列腺癌和宫颈癌)患者的盆腔放疗定位增强CT图像。勾画对象为盆腔淋巴引流区,包括盆腔骶前区(pP)、髂内淋巴引流区(IIN)、闭孔淋巴引流区(ON)、髂外淋巴引流区(EIN)、腹股沟淋巴引流区(IN)和腹部骶前区(ab P),共六个淋巴引流区(LNR)。采用综合国内外共识、指南及本中心经验制定的勾画标准,由两名有20年以上从业经验的高年资放疗医师进行手动勾画,将该手动勾画作为模型训练、验证以及测试的“标准标注”(Gt)。将数据集划分为训练-验证集(120例)和测试集(40例),用训练-验证集构建基于深度学习的自动勾画模型,用测试集对模型的自动勾画效能进行评估。(1)评估六个淋巴引流区分别的自动勾画结果,并将该模型与既往文献提出的具有较高效能的两个自动勾画模型UNet++和3D UNet在本测试集的自动勾画结果进行比较。(2)为评估自动勾画在模拟临床实际应用中的表现,按照患者实际放疗时的盆腔淋巴引流区临床靶区(CTV＿LN＿actu),从模型产生的六个淋巴引流区的自动勾画中选择相应的LNR,将其融合产生自动勾画淋巴引流区临床靶区(CTV＿LN＿auto),评价CTV＿LN＿auto与CTV＿LN＿actu的相似程度。(3)评价方法方面,一方面采用骰子相似系数(DSC)、平均表面距离(ASD)和95百分位数豪斯多夫距离(HD95)作为自动勾画结果的客观定量评价指标,另一方面,两名高年资放疗医师采用七分量表评分法对六个淋巴引流区分别的自动勾画结果和CTV＿LN＿auto进行主观定性评价,打分范围为1-7分,1分表示不需修改、可直接用于临床,7分表示存在重大错误、有75%以上层面需要修改。(4)一名低年资放射肿瘤科住院医师分别在有和无自动勾画辅助的情况下,对测试集所有淋巴引流区进行手动勾画,对两次勾画的准确度进行对比。(5)对高年资医师手动完成每例Gt勾画和模型产生每例自动勾画所花费的时间均进行了记录。结果:收集了160例盆腔原发恶性肿瘤(直肠癌61例、前列腺癌62例、宫颈癌37例)患者的定位CT图像及相关一般临床信息,构建了一个含盆腔淋巴引流区标准标注的160例盆腔增强放疗定位CT数据集,提出了一个基于深度学习的模型来实现对多种盆腔恶性肿瘤放疗中常包括的六个盆腔淋巴引流区的自动勾画。定量评价中,该模型产生的自动勾画在腹部骶前区、盆腔骶前区、髂内淋巴引流区、髂外淋巴引流区、闭孔淋巴引流区和腹股沟淋巴引流区的平均DSC分别为0.851、0.896、0.894、0.913、0.942和0.875,平均ASD分别为1.0mm、0.4mm、0.6mm、0.5mm、0.6mm和0.7mm,平均HD95分别为3.3mm、2.5mm、3.7mm、2.0mm、2.8mm和2.6mm。亚组分析中,除男性与女性病例之间的DSC以及三个癌种之间的HD95差异存在统计学意义外,其余各定性评价指标在男性与女性病例之间、术后与未手术病例之间、三个瘤种间的差异均无统计学意义。该模型的自动勾画在所有淋巴引流区的平均DSC、ASD和HD95均优于UNet++和3D U-Net。以CTV＿LN＿actu为比较基准,CTV＿LN＿autu的平均DSC、ASD和HD95分别为0.736、2.4mm和14.8mm。定性评价中,医师A和医师B分别对95.9%和96.2%的各个淋巴引流区的自动勾画给出了1-3分,即最多需要少许修改。医师A对大多数CTV＿LN＿auto评为5或6分,而医师B对大多数CTV＿LN＿auto评为5分。低年资医师在自动勾画辅助下的勾画结果的DSC、ASD和HD95均优于无自动勾画辅助时,DSC为0.829 vs 0.690(p<0.0001),ASD为0.9mm vs 2.2mm(p<0.0001),HD95为4.7mm vs 11.1mm(p<0.0001)。高年资放疗医师手动完成每例Gt勾画平均耗时30.4分钟,而该模型产生每例自动勾画平均耗时仅需19.7s。结论:该模型可产生适用于多种盆腔恶性肿瘤放疗的淋巴引流区的自动勾画,具有较高的准确性,大多数自动勾画结果仅需少量修改即可满足临床应用要求,有效提高放疗医师靶区勾画效率、指导低年资医师靶区勾画,具有临床应用价值。更多还原

【Abstract】 Objective:To construct a deep learning-based automatic segmentation method to autodelineate pelvic lymph node regions(LNRs)in patients with various pelvic primary malignancies,expecting to improve the standardability,accuracy and efficiency of target volume contouring in pelvic radiotherapy.Materials and Methods:Contrast-enhanced planning CT studies from 160 patients with pathologically proven pelvic primary malignancies,including rectal cancer,prostate cancer and cervical cancer,were retrospectively collected.The object of delineation is pelvic LNR.Six LNRs i.e.,abdominal presacral region(ab P),pelvic presacral region(p P),internal iliac nodes region(IIN),external iliac nodes region(EIN),obturator nodes region(ON)and inguinal nodes region(IN)were delineated for 160 patients by two senior radiation oncologists(both with over 20 years of experience in radiotherapy of patients with pelvic malignancies).The delineation was based on several widely used consensus and guidelines at home and aboard and experience of our department.The manual contour by the two senior radiation oncologists was set as the gold standard(Gt)for the training,validation and testing of the deep learning-based model.The entire dataset was divided into training-validation cohort(n=120)and testing cohort(n=40).The training-validation cohort was exploited to construct the model,which was subsequently tested in the testing cohort.Firstly,the automatic delineation of six LNRs was evaluated,and the results were compared with those of the UNet++ and3 D U-Net,two automatic segmentation models showing good performance in previous literature.Secondly,to evaluate the performance of automatic delineation in simulated clinical application,the clinically relevant ones were selected from the artificial intelligience(AI)-generated contours of all LNRs and were merged together(CTV＿LN＿auto).The clinically relevant LNRs depended on tumor stage and location,and were the same as the ones that were included in the lymph node clinical target volume approved for actual clinical application before(CTV＿LN＿actu).In terms of evaluation metrics,on one hand,Dice similarity coefficient(DSC),Average surface distance(ASD)and 95 th percentile Hausdorff distance(HD95)were calculated as quantitative evaluation metrics.On the other hand,a 7-point scale score was given,through visual examination by the same two senior radiation oncologists,as qualitative evaluation to evaluate the AI-generated contours of all LNRs and CTV＿LN＿auto,with score 1 meaning good agreement,acceptable to treat’as is’ and score 7 meaning gross error,no resemblance to the clinical structure or >75% slices needing edit.Besides,a medical resident delineated the testing cohort twice,once without AI assistance and again with AI assistance.Meantime,the time needed to complete the Gt contour for each patient by the senior radiation oncologists and the model’s computing time for each patient was recorded.Results:We retrospectively collected pelvic contrast-enhanced planning CT studies and related information of 160 consecutive patients with pelvic primary malignancies(including 61 cases of rectal cancer,62 cases of prostate cancer and 37 cases of cervical cancer),and constructed a dataset of 160 patients’ contrast-enhanced pelvic planning CT studies with standard and consistent annotations of six LNRs,and proposed a deep-learning based model to automatically segment six LNRs frequently contoured in radiation treatment planning of different pelvic primary malignancies.In terms of quantitative evaluation results,the mean DSC of AI-generated contours is0.851,0.896,0.894,0.913,0.942 and 0.875 for ab P,p P,IIN,EIN,ON and IN,respectively.The mean ASD of AI-generated contours is 1.0mm,0.4mm,0.6mm,0.5m,0.6mm and 0.7mm for these six LNRs,respectively.The mean HD95 of AIgenerated contours is 3.3mm,2.5mm,3.7mm,2.0mm,2.8mm and 2.6m,respectively.In the subgroup analysis,except for the difference in DSC between male and female cases,and HD95 among the three cancer types,there is no statistically significant difference in other qualitative evaluation metrics among the three tumor types,between male and female cases,and postoperative and postoperative cases.Our model showed significantly higher accuracy in terms of DSC,ASD and HD95 in comparison with the 3D U-Net and UNet++.When it comes to the accuracy of CTV＿LN＿auto contours compared with CTV＿LN＿actu contours,the mean values of DSC,ASD and HD95 are 0.736,2.4mm,and 14.8mm,respectively.In terms of qualitative evaluation,95.9% and 96.2% automatic contour of six LNRs got a score of 1-3,meaning requiring only minor edits at most,by radiation oncologist A and radiation oncologist B,respectively.Most CTV＿LN＿auto got a score 5 or 6 by senior radiation oncologist A and a score of 5 by senior radiation oncologist B.When with AI-generated contours as assistance,the medical resident’s contours achieved better DSC(0.829 vs 0.690,p<0.0001),ASD(0.9mm vs 2.2mm,p<0.0001),and HD95(4.7mm vs 11.1mm,p<0.0001).The average time needed to complete the Gt contours for each patient was30.4 min,while the runtime by the model to automatically delineate all six LNRs separately was only on average 19.7 seconds.Conclusion:The Cascade Multi-head U-net(CMU-net)constructed for automated delineation of the pelvic lymph node regions in variable pelvic malignancies was shown to produce high accuracy,with most automatic contour needing only minor modifications to meet clinical requirments and demonstrated its role in improving contouring efficiency,as well as in guiding radiation oncology residents,which shows its application value in clinical practice.更多还原