来自整个肿瘤区域和肿瘤-脑界面区域的纹理特征是区分肿瘤类型及其恶性程度的关键指标。然而，这两个区域的纹理特征对于识别胶质母细胞瘤和转移性肿瘤的判别价值尚未得到彻底探索。本研究的目的是开发和验证一种诊断模型，该模型结合了整个肿瘤区域和 10 mm 肿瘤与大脑界面区域的纹理特征，试图识别更稳定和有效的纹理特征。我们回顾性收集了增强的纹理特征2010 年至 2024 年间 97 例胶质母细胞瘤 (GBM) 和单脑转移瘤 (SBM) 患者的 T1 加权成像数据。利用机器学习建立多种诊断模型，根据整个肿瘤和 10 mm 的纹理特征区分 GBM 和 SBM肿瘤与大脑的界面区域。通过 5 倍交叉验证分析对结果进行评估，计算每个模型的受试者工作特征曲线下面积 (AUC)。使用 Delong 测试比较每个模型的性能，并通过采用 Shapley 附加解释 (SHAP) 进一步增强优化模型的可解释性。使用 FeAture Explorer (FAE) 软件比较验证数据集中所有管道的 AUC。在Relief和自动编码器（AE）建立的模型中，使用“一标准误差”规则的AUC最高。 “10mm_glrlm_GrayLevelNonUniformity”被认为是最稳定和最具预测性的特征。训练集、测试集和验证集的最佳模型并不相同。测试集中，Relief19AE模型的AUC最高为0.869，准确率为0.857。结合肿瘤整体和肿瘤-脑界面的纹理特征模型有利于区分胶质母细胞瘤和孤立性转移瘤，肿瘤界面表现出更高的异质性。版权所有 © 2024 大学放射科医生协会。由爱思唯尔公司出版。保留所有权利。

Texture features, derived from both the entire tumor area and the region of the tumor-to-brain interface, are crucial indicators for distinguishing tumor types and their degrees of malignancy. However, the discriminative value of texture features from both regions for identifying glioblastomas and metastatic tumors has not been thoroughly explored. The aim of this study is to develop and validate a diagnostic model that combines texture features from the entire tumor area and a 10 mm tumor-to-brain interface region, in an attempt to identify more stable and effective texture features.We retrospectively collected enhanced T1-weighted imaging data from 97 patients with glioblastoma(GBM) and single brain metastasis(SBM) between 2010 and 2024. Machine learning is used to establish multiple diagnostic models for discriminating GBM and SBM based on texture features of the entire tumor and 10 mm tumor-to-brain interface regions. Results underwent evaluation through 5-fold cross-validation analysis, calculating the area under the receiver operating characteristic curve (AUC) for each model. The performance of each model was compared using the Delong test, and the interpretability of the optimized model was further augmented by employing Shapley additive explanations (SHAP).The AUCs for all pipelines in the validation dataset were compared using FeAture Explorer (FAE) software. Among the models established by Relief and autoencoder (AE), the AUC was highest using the "one-standard error" rule. '10mm_glrlm_GrayLevelNonUniformity' was considered the most stable and predictive feature. The best models in the training set, test set, and validation set were not the same. In the test set, the Relief19AE model had the highest AUC of 0.869 and an accuracy of 0.857.The texture feature model that combines the overall tumor and the tumor-brain interface is beneficial for distinguishing glioblastoma from solitary metastatic tumors, and the texture features of the tumor interface exhibit higher heterogeneity.Copyright © 2024 The Association of University Radiologists. Published by Elsevier Inc. All rights reserved.