基于RNNU-Net深度学习模型的肺癌CT图像心脏自动勾画研究及临床可行性评估

Study and Clinical Feasibility Evaluation of Automatic Heart Segmentation in Lung Cancer CT Images Based on RNNU-Net Deep Learning Model

目的:探讨基于RNNU-Net深度学习模型在肺癌CT图像上自动勾画心脏的准确率和临床可行性。方法:选取75例肺癌患者的CT图,并由相关从业人员勾画心脏标签图,制作成数据集。将数据集随机分为训练集(n=51)、验证集(n=7)和测试集(n=17),并对训练集进行数据扩充,使用戴斯相似度系数(Dice similarity coefficient,DSC)、杰卡德相似度系数(Jaccard similarity coefficient,JSC)、阳性预测率(positive predictive value,PPV)、灵敏度(sensitivity,SE)、豪斯多夫距离(Hausdorf distance,HD)、相对体积差(relative volume difference,RVD)、体积重叠误差(volumetric overlap error,VOE)对模型进行定量评价,并对比U-Net模型和低年资从业人员对心脏的勾画结果。结果:基于RNNU-Net深度学习模型的测试集DSC、JSC、PPV、SE、HD、RVD、VOE评价指标值分别为:91.06%±10.94%、85.09%±15.10%、96.01%±9.35%、88.21%±13.42%、4.66±1.26、12.45%±18.70%、13.48±20.11。RNNU-Net的大多数评价指标与低年资从业人员之间的差异有统计学意义,且RNNU-Net模型表现出更优的效果,同时散点图与箱型图结果显示,相较于低年资从业人员勾画结果,RNNU-Net模型有着更少的0值。定性评价结果显示RNNU-Net深度学习模型准确地划分了下腔静脉与心脏,边界更加平滑,且缓解了漏勾现象,同时,在心脏周围存在肿瘤侵犯的情况下有效缓解了U-Net模型的过分割现象。结论:基于RNNU-Net的深度学习模型在肺癌CT图像上的心脏自动勾画上存在一定的优势,可以缩短临床勾画时间,并有效弥补漏勾画等现象。

Objective:This study aims to assess the accuracy and clinical feasibility of automatic heart segmentation on CT images of lung cancer using the RNNU-Net deep learning model.Methods:CT images from 75 lung cancer patients was collected.Heart label maps were manually delineated by experts to create the dataset.The dataset was randomly classified into a training set(n=51),a validation set(n=7)and a test set(n=17).Data augmentation techniques were applied to expand the training set.Quantitative evaluation metrics including Dice similarity coefficient(DSC),Jaccard similarity coefficient(JSC),positive predictive value(PPV),sensitivity(SE),Hausdorff distance(HD),relative volume difference(RVD),and volume overlap error(VOE)were used to assess the model.The heart segmentation results of the U-Net model were compared to those of junior practitioners.Results:The evaluation indices of the test set using the RNNU-Net deep learning model were as follows:DSC(91.06%±10.94%),JSC(85.09%±15.10%),PPV(96.01%±9.35%),SE(88.21%±13.42%),HD(4.66±1.26),RVD(12.45%±18.70%)and VOE(13.48±20.11).Statistical analysis revealed significant differences between the evaluation indices of the RNNU-Net model and those of junior doctors,with the RNNU-Net model demonstrating superior performance.Scatter plot and box plot results indicated that the RNNU-Net model had fewer zero values compared to the junior doctors.Qualitative evaluation demonstrated that the RNNU-Net deep learning model accurately segmented the inferior vena cava and the heart,resulting in smoother boundaries and decreasing missed segmentation.Furthermore,the RNNU-Net model effectively reduced over-segmentation in cases of invasion around the heart,compared to the U-Net model.Conclusion:The deep learning model based on RNNU-Net exhibits advantages in the automatic delineation of the heart on CT images of lung cancer.It reduces the time required for clinical delineation and effectively compensates for missed delineation.