基于CT图像利用深度学习方法自动定位盆腔淋巴结分区的初步研究

A preliminary study for segmentation of the areas of pelvic lymph node on CT images based on deep lear-ning algorithms

目的:探索基于CT图像的深度学习方法自动定位盆腔淋巴结区域的可行性。方法:回顾性连续搜集在本院就诊的符合研究要求的盆腔恶性肿瘤患者的腹盆部CT图像,共将178个腹盆部薄层门静脉期扫描序列的图像数据纳入研究,并将其按疾病类型分为两个数据集:数据集1包括2018年8月-2021年4月共131例前列腺癌患者的131个序列的图像,用于模型训练;数据集2包括2021年1月-2021年6月本院47例盆腔肿瘤(卵巢癌、宫颈癌和直肠癌)患者的47个序列的图像,用于外部验证。在数据集1中每个序列的CT图像上,由两位影像科医师标注盆腔淋巴结的区域定位(共划分为13个分区,包括主动脉旁、双侧髂总动静脉、双侧髂内外动静脉、双侧闭孔、双侧腹股沟、骶前和直肠旁区域)。将131个序列的图像数据随机分为训练集(train set,n=99)、调优集(validation set,n=17)和测试集(test set,n=15)。通过训练U-net 3D深度学习网络,建立淋巴结自动定位分区模型,对模型在数据集1的测试集中的定位能力进行定量评价,评价指标包括交并比(IOU)、体积相似度(VS)和关键点正确估计比例(PCK)。对模型在数据集2中自动定位淋巴结分区的能力进行定性评价,评价指标包括模型定位的淋巴结区域的覆盖程度(0~2分)、超出程度(0~1分)及超出范围(0~2分)分级,总分值为0~5分(不满意~满意)。结果:在数据集1的测试集中,盆腔淋巴结自动定位分区模型预测各组淋巴结的交并比(IOU)为0.28~0.77(P<0.001),体积相似度(VS)为0.62~0.99(P<0.001),关键点正确估计比例(PCK)-10 mm为53.85%~100%(P=0.446)。数据集2的评价结果显示,模型预测各区域盆腔淋巴结主观评价各项指标得分之和的中位数:双侧髂总动静脉、双侧腹股沟、双侧髂内动静脉、双侧闭孔、主动脉、骶前和直肠旁为5分,左侧髂外动静脉为4分,右侧髂外动静脉为3分。以总评分≥4分为达到临床满意的标准,模型对84.59%淋巴结的自动定位分区结果准确,其中以双侧腹股沟区域的满意率最高,达100%。在13个淋巴结分区中,11个分区的满意率超过80%,其中4个分区在90%以上;以双侧髂外动静脉淋巴结区域的定位满意率较差(左侧为60%,右侧为51%)。结论:通过深度学习方法在CT图像上自动定位盆腔淋巴结区域是可行的。

Objective:The purpose of this study was to evaluate the feasibility of automatic segmentation of pelvic lymph node areas on CT images based on deep learning algorithms.Methods:Two datasets of the consecutive abdominal and pelvic CT images in patients suspected of pelvic malignant tumor were retrospectively collected.178 thin-layer portal vein phase CT sequences were incorporated into this study and divided into 2 datasets by primary disease.In dataset 1,131 imaging series of 131 patients with prostate cancer from August 2018 to April 2021 were recruited for the training of a deep learning model.In dataset 2,47 imaging series of 47 patients with ovarian cancer,cervical cancer or rectal cancer from Jan 2021 to Jun 2021 were recruited for the external validation model.All the images of dataset 1 were reviewed and annotated the located areas of pelvic lymph nodes by two radiologists.A total of 13 areas were divided including the lymph node groups of inferior para-aortic,left common iliac arteries and veins,left external iliac arteries and veins,left inguinal,left internal iliac arteries and veins,left obturator,right common iliac arteries and veins,right external iliac arteries and veins,right inguinal,right internal iliac arteries and veins,right obturator,pre-sacral,and peri-rectum.All images of 13 series of dataset 1 were randomly divided into train set(n=99),validation set(n=17)and test set(n=15).A U-net 3D network was trained to establish a model to locate the pelvic lymph node areas automatically.In dataset 1,the intersection over union(IOU),volume similarity(VS)and percentage of correct key points(PCK)were used to evaluate the model efficiency in localization of the pelvic node groups.In dataset 2,the qualitative evaluation metrics were used to evaluate the perfor-mance of the model(on a scale of 0~5 indicated unsatisfying to satisfying),including degree of model coverage(on a scale of 0~2),degree of model excess(on a scale of 0~1),and extra-nodal region(on a scale of 0~2).Results:In dataset 1,the intersection over union(IOU),volume similarity(VS)and percentage of correct keypoints(PCK)in the test set were 0.28~0.77(P<0.001),0.62~0.99(P<0.001)and 53.85%~100%(P=0.446).In dataset 2,the medians of sum of all qualitative evaluation metrics score of pelvic lymph node groups:5 at areas of inferior para-aortic,left common iliac arteries and veins,left inguinal,left internal iliac arteries and veins,left obturator,right common iliac arteries and veins,right inguinal,right internal iliac arteries and veins,right obturator,pre-sacral and peri-rectum,4 at areas of left external iliac arteries areas,and 3 at right external iliac arteries and veins.With a standard of sum of all qualitative evaluation metrics score≥4,84.59%of the predicted outcomes of automatic location of lymph nodes areas were satisfying,and the bilateral inguinal groups were the highest,with satisfaction rate of both 100%.In a total of 13 lymph node groups,the satisfaction rate of 11 groups was more than 80%,among which that of the 4 groups was more than 90%.The satisfaction rate of location of bilateral external iliac lymph node group was the lowest(left:60%;right:51%).Conclusion:It is feasible to use a deep learning model to segment the areas of the pelvis lymph nodes on CT images.