机器学习优化能谱CT预测胃腺癌的浸润性

Machine Learning Optimization of Spectral CT for Predicting Invasiveness in Gastric Adenocarcinoma

目的探讨机器学习(ML)算法结合能谱CT定量参数和临床模型在预测胃腺癌(GAC)患者淋巴血管浸润(LVI)和神经周围浸润(PNI)的潜在价值。方法收集2017年12月—2022年5月经病理证实的GAC患者114例。研究参数涉及血清肿瘤标志物、CT-TN分期、CT评估壁外血管浸润(CT-EMVI)以及能谱CT定量参数。通过WEKA软件的Best-First算法进行特征筛选,并运用贝叶斯网络(BN)及支持向量机(SVM)算法建立模型。结果相较于LVI/PNI阴性组,LVI/PNI阳性组中CT-T_(3~4)期、CT-N阳性、CT-EMVI阳性、血清肿瘤标志物[糖类抗原(CA)72-4和CA19-9]更为常见,能谱CT参数也更高,差别均有统计学意义(P<0.05)。经特征选择,关键变量包括CT-T分期、CT-EMVI、VP-NIC和EP-70 keV CT值。基于这些变量,分别使用BN和SVM构建临床参数模型、能谱CT参数模型和混合模型,共6个模型。6个模型均展现了高预测性能,无过拟合现象。BN的混合模型预测性能最佳,AUC值为0.890~0.933,Delong检验显示其在统计学上具有显著优势(P<0.05);而SVM的混合模型与另外2种模型间的差别无统计学意义(P>0.05)。结论结合临床和能谱CT参数的ML模型能够高效能评估GAC患者的LVI和PNI状态,其中BN混合模型的准确性最高。

Objective It is to investigate the potential value of machine learning(ML)algorithms combined with quantitative parameters of spectral CT and clinical models in predicting lymphovascular invasion(LVI)and perineural invasion(PNI)in patients with gastric adenocarcinoma(GAC).Methods A total of 114 patients with GAC confirmed by pathology from December 2017 to May 2022 were collected.The study parameters involved serum tumor markers,CT-TN staging,extramural venous invasion assessed by CT(CT-EMVI),and quantitative parameters of spectral CT.Feature screening was conducted using the Best-First algorithm in WEKA software,and models were built using Bayesian network(BN)and support vector machine(SVM)algorithms.Results Compared with the LVI/PNI(-)group,CT-T_(3-4)stage,CT-N positive,CT-EMVI positive,and serum tumor markers(CA72-4,CA19-9)were more common in LVI/PNI(+)group,and the spectral CT parameters were higher in LVI/PNI(+)group,with statistically significant differences observed(P<0.05).After feature screening,key variables included CT-T staging,CT-EMVI,VP-NIC,and EP-70 keV CT values.Based on these variables,six models were constructed using BN and SVM,including clinical parameter models,spectral CT parameter models,and combined models.All the six models demonstrated high predictive performance without overfitting.The combined model using BN showed the best predictive performance with an AUC range of 0.890 to 0.933,and Delong test showed that it was statistically significant(P<0.05).In contrast,the SVM combined model did not show a statistical difference between the model and the other two models(P>0.05).Conclusion Machine learning models combining clinical and spectral CT parameters can efficiently evaluate the LVI and PNI status of GAC patients,with the BN combined model achieving the highest accuracy.