探讨基于CT生境影像组学构建的机器学习模型在术前无创预测肺腺癌间变性淋巴瘤激酶(ALK)基因融合表达中的价值。

回顾性纳入2015年3月至2023年11月于张家港市第一人民医院影像科完成ALK基因检测及术前胸部CT检查的130例肺腺癌患者(ALK阳性45例，阴性85例)。按7∶3随机划分为训练集(n＝90)与测试集(n＝40)。采用K-means聚类将病灶划分为两个生境亚区(Habitat 1和Habitat 2)，提取并筛选14个关键生境影像组学特征，随后分别利用自编码器(AE)、遗传规划(GP)、线性判别分析(LDA)、逻辑回归(LR)、Lasso逻辑回归(LRLasso)和支持向量机(SVM)6种算法构建模型；采用受试者操作特征(ROC)曲线评估模型效能，并以德龙检验比较AUC差异。

基于LR和LRLasso算法构建的机器学习模型训练集AUC分别为0.862(0.788~0.935)和0.854(0.778~0.930)，测试集AUC分别为0.830(0.678~0.930)和0.802(0.646~0.911)。LR模型与LRLasso模型和AE模型的AUC差异无统计学意义(P＝0.182，0.104)，与剩余模型AUC差异有统计学差异(均P<0.05)。测试集中，LR模型的敏感度和特异度为71.4%和96.2%，LRLasso模型的敏感度和特异度为64.3%和88.5%。

基于CT影像生境影像组学模型在肺腺癌ALK基因融合具有一定的预测能力并显示潜在应用价值，基于LR的机器学习模型具有较好的泛化能力与临床应用潜力，有望作为无创预测肺腺癌ALK基因融合的新型影像工具。

To explore the value of machine learning models constructed based on CT habitat radiomics in non-invasive preoperative prediction of anaplastic lymphoma kinase (ALK) gene fusion expression in lung adenocarcinoma.

A total of 130 patients with lung adenocarcinoma who completed ALK gene testing and preoperative chest CT examination in the Imaging Department of the First People′s Hospital of Zhangjiagang from March 2015 to November 2023 were retrospectively included (45 cases were ALK positive and 85 cases were ALK negative). They were randomly divided into the training set (n=90) and the test set (n=40) in a 7∶3 ratio. The lesions were divided into two habitat subregions (Habitat 1 and Habitat 2) by K-means clustering, and 14 key habitat radiomics features were extracted and screened. Subsequently, models were constructed respectively using 6 algorithms: autoencoder (AE), genetic programming (GP), linear discriminant analysis (LDA), logistic regression (LR), Lasso logistic regression (LRLasso), and support vector machine (SVM). The receiver operator characteristic (ROC) curve was used to evaluate the model efficacy, and the DeLong test was used to compare the differences in area under the curve (AUC).

The AUCs of the machine learning model training set constructed based on the LR and LRLasso algorithms were 0.862 (0.788-0.935) and 0.854 (0.778-0.930), respectively, and the AUCs of the test set were 0.830 (0.678-0.930) and 0.802 (0.646-0.911), respectively. There were no statistically significant differences in the AUC between the LR model and the LRLasso model or the AE model (P=0.182, 0.104), but there were statistically significant differences in the AUC between the LR model and the remaining models (all P<0.05). In the test set, the sensitivity and specificity of the LR model were 71.4% and 96.2%, respectively, while those of the LRLasso model were 64.3% and 88.5%, respectively.

The CT image-based habitat radiomics shows a certain predictive capability and potential clinical utility for identifying ALK gene fusion in lung adenocarcinoma. The machine learning model based on LR has a good generalization ability and a potential clinical applicability, and may be used as a new non-invasive imaging tool for predicting ALK gene fusion in lung adenocarcinoma.