Department of Radiology, The People's Hospital of Guangxi Zhuang Autonomous Region & Research Center of Radiology, Guangxi Academy of Medical Sciences, Nanning, 530021, Guangxi, China.
Department of Radiology, First Affiliated Hospital, School of Medicine, Zhejiang University, 79 Qingchun Road, Hangzhou, 310007, Zhejiang, China.
Abdom Radiol (NY). 2022 Nov;47(11):3782-3791. doi: 10.1007/s00261-022-03639-6. Epub 2022 Aug 17.
A log-combined model was developed to predict the invasive behavior of pancreatic solid pseudopapillary neoplasm (pSPN) based on clinical and radiomic features extracted from multiparametric magnetic resonance imaging (MRI).
A total of 111 patients with pathologically confirmed pSPN who underwent preoperative plain and contrast-enhanced MRI were included, and divided into an invasive group (n = 34) and non-invasive group (n = 77). Clinical features and laboratory data related to pSPN invasive behavior were analyzed. Regions of interest were delineated based on T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced T1WI (CE-T1WI) to extract radiomic features. Correlation analysis was performed for these features, followed by L1_based feature selection (C = 0.15). A logistic regression algorithm was used to construct models based on each of the four sequences and a log-combined model was used to integrate the sequences. A receiver operating characteristic (ROC) curve was plotted to evaluate the model performance, and the Brier score was used to assess the overall accuracy of the model predictions.
The area under the ROC curve was 0.68, 0.73, 0.71, and 0.49 for Log-T1WI, Log-T2WI, Log-DWI, and Log-CE models, respectively, and 0.81 for the log-combined model. The accuracy, precision, sensitivity, and specificity of the log-combined model were 0.77, 0.88, 0.75, and 0.78, respectively. The best performance was obtained with the log-combined model with a Brier score of 0.18. Tumor location was identified as a significant clinical feature in comparison between the two groups (p < 0.05), and invasive pSPN was more frequent in the tail of the pancreas.
The log-combined model based on multiparametric MRI and clinical features can be used as a non-invasive diagnostic tool for preoperative prediction of pSPN invasive behavior and to facilitate the development of individualized treatment strategies and monitoring management plans.
基于多参数磁共振成像（MRI）提取的临床和放射组学特征，建立 Log 联合模型以预测胰腺实性假乳头状肿瘤（pSPN）的侵袭行为。
共纳入 111 例经病理证实的 pSPN 患者，术前均行平扫及增强 MRI 检查，根据术后病理分为侵袭组（n=34）和非侵袭组（n=77）。分析与 pSPN 侵袭行为相关的临床特征和实验室数据。基于 T1 加权成像（T1WI）、T2 加权成像（T2WI）、弥散加权成像（DWI）和对比增强 T1WI（CE-T1WI）勾画感兴趣区，提取放射组学特征。对这些特征进行相关性分析，然后基于 L1 正则化进行特征选择（C=0.15）。使用逻辑回归算法分别基于四个序列构建模型，并使用 Log 联合模型整合序列。绘制受试者工作特征（ROC）曲线评估模型性能，采用 Brier 评分评估模型预测的整体准确性。
ROC 曲线下面积分别为 Log-T1WI、Log-T2WI、Log-DWI 和 Log-CE 模型的 0.68、0.73、0.71 和 0.49，Log 联合模型为 0.81。Log 联合模型的准确性、精确性、敏感性和特异性分别为 0.77、0.88、0.75 和 0.78。Brier 评分为 0.18，联合模型的性能最佳。与两组相比，肿瘤位置是有显著差异的临床特征（p<0.05），胰腺尾部侵袭性 pSPN 更为常见。
基于多参数 MRI 和临床特征的 Log 联合模型可作为术前预测 pSPN 侵袭行为的一种非侵入性诊断工具，有助于制定个体化治疗策略和监测管理方案。
