基于CT检查的集成深度学习模型对肝门静脉定性与定量分型研究

CT‑based integrated deep learning model for qualitative and quantitative research of hepatic portal vein

目的基于CT检查的集成深度学习模型对肝门静脉定性与定量分型研究。方法采用回顾性研究方法。收集2017年10月至2019年1月清华大学附属北京清华长庚医院收集291例行上腹部增强CT检查患者的影像学数据;男195例,女96例;年龄为(51±12)岁。采用三维重建系统对增强CT检查肝门静脉系统进行重建,并将三维点云输入至编码器模型后,得到三维重建的向量化表示。该向量化表示可以用于预测定性分型和定量表示分型。正态分布的计量资料以x±s表示,组间比较采用配对t检验。计数资料以百分数或绝对数表示,组间比较采用配对χ^(2)检验。结果(1)门静脉三维重建结果及解剖分型。291例患者重建肝门静脉的三维结构,肝门静脉主干分型:Akgul A型211例,Akgul B型29例,Akgul C型16例,Akgul D型10例,无Akgul E型患者,无法分型25例。(2)门静脉主干定性分型的预测。291例患者样本中,排除25例因成像质量差或无法进行Akgul分型的样本,266例样本用于机器模型肝门静脉主干的自动定性分型(肝门静脉分类器),其中Akgul A型211例,Akgul B型29例,Akgul C&D型26例。266例患者Macro‑F1为61.93%±40.50%,准确率为84.99%,Random分类器的Macro‑F1为32.38%±19.81%,准确率为61.65%,两者上述指标比较,差异均有统计学意义(t=7.85,χ^(2)=62.89,P<0.05)。(3)定量表示门静脉分型。定量分型相似样本的Akgul定性分型的概率P@1为73%±45%、P@3为70%±37%、P@5为69%±35%、P@10为67%±32%、MRR为80%±34%;基线模型P@1为57%±50%、P@3为58%±35%、P@5为58%±32%、P@10为58%±30%、MRR为70%±37%;两种模型上述指标比较,差异均有统计学意义(t=5.22,5.11,5.00,4.99,3.47,P<0.05)。结论基于CT检查使用三维重建和深度学习技术建立肝门静脉结构的自动分型模型,可实现自动定性分型并定量描述肝门静脉结构。

Objective To investigate the computed tomography(CT)‑based integrated deep learning model for qualitative and quantitative classification of hepatic portal vein.Methods The retrospective study was conducted.The CT imaging data of 291 patients undergoing upper-abdomen enhanced CT examination in the Beijing Tsinghua Changgung Hospital of Tsinghua University from October 2017 to January 2019 were collected.There were 195 males and 96 females,aged(51±12)years.The hepatic portal vein was reconstructed using the three‑dimensional reconstruction system.Three-dimensional point cloud was input to the encoder model to obtain the three-dimensional reconstructed vectorized representation,which was used for qualitative classification and quantitative representation classification.Measurement data with normal distribution were represented as Mean±SD,and comparison between groups was conducted using the paired t test.Count data were repre-sented as percentages or absolute numbers,and comparison between groups was analyzed using the paired chi‑square test.Results(1)Three-dimensional reconstruction of portal vein and anatomical classification.Three‑dimensional structure was reconstructed in the 291 patients.Classification of main hepatic portal vein showed 211 cases of Akgul type A,29 cases of Akgul type B,16 cases of Akgul type C,10 cases of Akgul type D,and 25 cases of unclassifiable.(2)Prediction of qualitative classification of main hepatic portal vein.Of the 291 patient samples,25 unclassifiable or poor quality samples were excluded,266 samples were used for automated qualitative classification of the main portal vein by machine model.There were 211 cases of Akgul type A,29 cases of Akgul type B,26 cases of Akgul type C&D.The Macro‑F1 of 266 patients was 61.93%±40.50%and the accuracy was 84.99%,versus 32.38%±19.81%and 61.65%of Random classifier,showing significant differ-ences between them(t=7.85,χ^(2)=62.89,P<0.05).(3)Quantitative representation of portal vein classification.The probabilities of quantitative classification for Akgul qualitative classification of similar samples included P@1 as 73%±45%,P@3 as 70%±37%,P@5 as 69%±35%,P@10 as 67%±32%,mean reciprocal rank(MRR)as 80%±34%,versus 57%±50%,58%±35%,58%±32%,58%±30%,70%±37%of the baseline model,showing significant differences between the two analytical methods(t=5.22,5.11,5.00,4.99,3.47,P<0.05).Conclusion The automated classification model for the hepatic portal vein structure was constructed using CT‑based three‑dimensional reconstruction and deep learning technology,which can achieve automatic qualitative classification and quantitatively describe the hepatic portal vein structure.