Xie Yuping, Li Xuanzi, Yang Shuai, Jia Fujie, Han Yuanyuan, Huang Mingsheng, Chen Lei, Zou Wei, Deng Chuntao, Liang Zibin
The Cancer Center of The Fifth Affiliated Hospital of Sun Yat-sen University, Zhuhai, China.
Department of Radiotherapy, The Cancer Center of the Fifth Affiliated Hospital of Sun Yat-sen University, Zhuhai, China.
Transl Cancer Res. 2025 Feb 28;14(2):731-742. doi: 10.21037/tcr-24-1355. Epub 2025 Feb 24.
Brain metastases are common brain tumors in adults. Brain metastases from different primary tumors have special magnetic resonance imaging (MRI) features. As a new technology that can extract and quantify medical image data, and with the rapid development of artificial intelligence, the machine learning model based on radiology has been successfully applied to the diagnosis and differentiation of tumors. This study aimed to develop radiomics models from post-contrast T1-weighted images using machine learning algorithms to differentiate lung cancer from breast cancer brain metastases.
A retrospective analysis was conducted on 118 lung cancer brain metastases patients and 62 breast cancer brain metastases patients confirmed by surgery pathology or combined clinical and imaging diagnosis at The Fifth Affiliated Hospital of Sun Yat-sen University from August 2015 to September 2023. Patients were randomly divided into a training set (126 cases) and a validation set (54 cases) at a 7:3 ratio. Enhanced T1-weighted images of all patients were imported into ITK-SNAP software to manually delineate the region of interest (ROI). Radiomic features were extracted based on the ROI and feature selection was performed using the least absolute shrinkage and selection operator. Significant features were used to develop models using logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), multilayer perceptron (MLP), and light gradient boosting machine (LightGBM). The diagnostic performance of the models was assessed using the receiver operating characteristic (ROC) curve.
The LightGBM radiomics model exhibited the best diagnostic performance, with an area under the curve (AUC) of 0.875 [95% confidence interval (CI): 0.819-0.931] in the training set and 0.866 (95% CI: 0.740-0.993) in the validation set.
The enhanced MRI radiomics model, especially the LightGBM model, can accurately predict the primary lesion types of brain metastases from lung cancer and breast cancer origins.
脑转移瘤是成人常见的脑部肿瘤。不同原发肿瘤的脑转移瘤具有特殊的磁共振成像（MRI）特征。作为一种能够提取和量化医学图像数据的新技术，随着人工智能的快速发展，基于放射学的机器学习模型已成功应用于肿瘤的诊断与鉴别诊断。本研究旨在利用机器学习算法从增强T1加权图像中开发放射组学模型，以鉴别肺癌脑转移瘤和乳腺癌脑转移瘤。
对2015年8月至2023年9月在中山大学附属第五医院经手术病理或临床与影像联合诊断确诊的118例肺癌脑转移瘤患者和62例乳腺癌脑转移瘤患者进行回顾性分析。患者按7:3的比例随机分为训练集（126例）和验证集（54例）。将所有患者的增强T1加权图像导入ITK-SNAP软件，手动勾画感兴趣区域（ROI）。基于ROI提取放射组学特征，并使用最小绝对收缩和选择算子进行特征选择。利用逻辑回归（LR）、支持向量机（SVM）、K近邻（KNN）、多层感知器（MLP）和轻梯度提升机（LightGBM），使用显著特征建立模型。采用受试者工作特征（ROC）曲线评估模型的诊断性能。
LightGBM放射组学模型表现出最佳的诊断性能，训练集中曲线下面积（AUC）为0.875 [95%置信区间（CI）：0.819 - 0.931]，验证集中为0.866（95% CI：0.740 - 0.993）。
增强MRI放射组学模型，尤其是LightGBM模型，能够准确预测肺癌和乳腺癌来源的脑转移瘤的原发灶类型。
