Shen Qian, Xiang Cong, Han Yongliang, Li Yongmei, Huang Kui
Department of Radiology, The Affiliated Stomatology Hospital of Southwest Medical University, Luzhou, China.
School of Artificial Intelligence, Chongqing University of Technology, Chongqing, China.
Front Med (Lausanne). 2025 Apr 22;12:1566555. doi: 10.3389/fmed.2025.1566555. eCollection 2025.
Computed tomography (CT) imaging of parotid pleomorphic adenoma (PA) has been widely reported, nonetheless few reports have estimated the capsule characteristics of PA at length. This study aimed to establish and validate CT-based intratumoral and peritumoral radiomics models to clarify the characteristics between parotid PA with and without complete capsule.
In total, data of 129 patients with PA were randomly assigned to a training and test set at a ratio of 7:3. Quantitative radiomics features of the intratumoral and peritumoral regions of 2 mm and 5 mm on CT images were extracted, and radiomics models of Tumor, External2, External5, Tumor+ External2, and Tumor+External5 were constructed and used to train six different machine learning algorithms. Meanwhile, the prediction performances of different radiomics models (Tumor, External2, External5, Tumor+External2, Tumor+External5) based on single phase (plain, arterial, and venous phase) and multiphase (three-phase combination) were compared. The receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC) were used to evaluate the prediction performance of each model.
Among all the established machine learning prediction radiomics models, the model based on a three-phase combination had better prediction performance, and the model using a combination of intratumoral and peritumoral radiomics features achieved a higher AUC than the model with only intratumoral or peritumoral radiomics features, and the Tumor+External2 model based on LR was the optimal model, the AUC of the test set was 0.817 (95% CI = 0.712, 0.847), and its prediction performance was significantly higher ( < 0.05, DeLong's test) than that with the Tumor model based on LDA (AUC of 0.772), the External2 model based on LR (AUC of 0.751), and the External5 model based on SVM (AUC of 0.667). And the Tumor+External2 model based on LR had a higher AUC than the Tumor+External5 model based on LDA (AUC = 0.817 vs. 0.796), but no statistically significant difference ( = 0.667).
The intratumoral and peritumoral radiomics model based on multiphasic CT images could accurately predict capsular characteristics of parotid of PA preoperatively, which may help in making treatment strategies before surgery, as well as avoid intraoperative tumor spillage and residuals.
腮腺多形性腺瘤（PA）的计算机断层扫描（CT）成像已有广泛报道，但很少有报告详细评估PA的包膜特征。本研究旨在建立并验证基于CT的瘤内和瘤周放射组学模型，以阐明有完整包膜和无完整包膜的腮腺PA之间的特征。
总共129例PA患者的数据以7:3的比例随机分配到训练集和测试集。提取CT图像上瘤内及瘤周2mm和5mm区域的定量放射组学特征，构建肿瘤、External2、External5、肿瘤+External2和肿瘤+External5的放射组学模型，并用于训练六种不同的机器学习算法。同时，比较基于单相（平扫、动脉期和静脉期）和多相（三相联合）的不同放射组学模型（肿瘤、External2、External5、肿瘤+External2、肿瘤+External5）的预测性能。采用受试者操作特征（ROC）曲线分析和曲线下面积（AUC）评估各模型的预测性能。
在所有建立的机器学习预测放射组学模型中，基于三相联合的模型具有更好的预测性能，使用瘤内和瘤周放射组学特征组合的模型比仅使用瘤内或瘤周放射组学特征的模型获得更高的AUC，基于逻辑回归（LR）的肿瘤+External2模型是最优模型，测试集的AUC为0.817（95%可信区间=0.712，0.847），其预测性能显著高于基于线性判别分析（LDA）的肿瘤模型（AUC为0.772）、基于LR的External2模型（AUC为0.751）和基于支持向量机（SVM）的External5模型（AUC为0.667）（P<0.05，DeLong检验）。并且基于LR的肿瘤+External2模型比基于LDA的肿瘤+External5模型具有更高的AUC（AUC = 0.817 vs. 0.796），但差异无统计学意义（P = 0.667）。
基于多期CT图像的瘤内和瘤周放射组学模型可术前准确预测腮腺PA的包膜特征，有助于制定术前治疗策略，避免术中肿瘤溢出和残留。
