MRI影像组学结合临床特征的机器学习模型对结直肠癌肝转移的预测价值

The predictive value of MRI imaging omics combined with clinical features in machine learning models for colorectal cancer liver metastasis

摘要:

背景与目的 结直肠癌肝转移（CRCLM）是影响患者预后的主要原因，术前无创、精准诊断对制定治疗方案至关重要。传统临床标志物特异性有限，本研究旨在基于多模态MRI影像组学特征，结合机器学习算法，构建预测CRCLM的高效模型，并评价其临床价值。方法 收集2022年5月—2024年5月于河南省南阳市第一人民医院行术前MRI检查并经病理证实的150例结直肠癌患者，随机分为训练集（n=120）和验证集（n=30）。其中CRCLM 57例，无CRCLM 93例。采用单因素与多因素分析筛选CRCLM独立危险因素，建立临床诊断模型。提取多模态MRI影像组学特征，经LASSO筛选后分别构建Logistic回归（LR）、支持向量机（SVM）、随机森林（RF）模型，并比较其诊断效能。建立临床及影像组学联合诊断模型，并通过受试者操作特征和决策曲线（DCA）评估效能与临床获益。结果 癌胚抗原（OR=1.323，95% CI=1.079~1.567）、糖类抗原19-9（OR=2.512，95% CI=1.225~3.799）及中性粒细胞/淋巴细胞比值（OR=1.881，95% CI=1.354~2.409）是CRCLM独立危险因素（均P<0.05），以上3个因素构建的临床诊断模型曲线下面积（AUC）为0.793。RF模型在训练集与验证集AUC最高（0.770、0.763），优于LR和SVM。基于RF的联合诊断模型在训练集与验证集AUC分别为0.913和0.947，明显优于单独临床或影像组学诊断模型，DCA显示联合诊断模型具有最高临床净获益。结论 RF模型在影像组学预测中表现最佳，其与临床特征结合的联合模型能显著提高CRCLM的无创诊断效能，具备较高的临床应用价值。

Abstract:

Background and Aims Colorectal cancer liver metastasis (CRCLM) is a major cause of poor prognosis in patients with colorectal cancer. Accurate and noninvasive preoperative diagnosis is essential for treatment planning. Conventional clinical biomarkers have limited specificity. This study aimed to develop an efficient predictive model for CRCLM by integrating multimodal MRI imaging omics features with machine learning algorithms, and to evaluate its clinical value.Methods A total of 150 patients with colorectal cancer who underwent preoperative MRI and were pathologically confirmed at Nanyang First People's Hospital between May 2022 and May 2024 were retrospectively analyzed. Patients were randomly divided into a training set (n=120) and a validation set (n=30), including 57 cases with CRCLM and 93 cases without. Univariate and multivariate analyses were performed to identify independent risk factors for CRCLM and to construct a clinical diagnostic model. Radiomics features were extracted from multimodal MRI, and the least absolute shrinkage and selection operator (LASSO) method was used for feature selection. Logistic regression (LR), support vector machine (SVM), and random forest (RF) models were built and compared for diagnostic performance. A combined clinical-imaging omics model was further established, and its performance and clinical utility were assessed using receiver operating characteristic curves and decision curve analysis (DCA).Results Carcinoembryonic antigen (OR=1.323, 95% CI=1.079-1.567), carbohydrate antigen 19-9 (OR=2.512, 95% CI=1.225-3.799), and neutrophil-to-lymphocyte ratio (OR=1.881, 95% CI=1.354-2.409) were identified as independent risk factors for CRCLM (all P<0.05). The clinical model constructed with these three factors achieved an AUC of 0.793. Among radiomics models, the RF model demonstrated the highest AUC in both training and validation sets (0.770 and 0.763), outperforming LR and SVM. The combined RF-based model yielded AUC of 0.913 and 0.947 in the training and validation sets, respectively, significantly exceeding the performance of the clinical or imaging omics models alone. DCA confirmed the superior net clinical benefit of the combined model.Conclusion The RF model showed the best diagnostic performance among imaging omics models. When integrated with clinical features, the combined RF model significantly improved the noninvasive diagnostic efficacy of CRCLM and demonstrated high potential for clinical application.