Shi Yan-Jie, Zhu Hai-Tao, Li Xiao-Ting, Zhang Xiao-Yan, Liu Yu-Liang, Wei Yi-Yuan, Sun Ying-Shi
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China.
Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Radiology, Peking University Cancer Hospital & Institute, No.52 Fu Cheng Road, Hai Dian District, Beijing 100142, China.
Clin Imaging. 2023 Apr;96:15-22. doi: 10.1016/j.clinimag.2023.01.008. Epub 2023 Jan 26.
This study aimed to investigate the diagnostic performance of the histogram array and convolutional neural network (CNN) based on diffusion-weighted imaging (DWI) with multiple b-values under magnetic resonance imaging (MRI) to distinguish pancreatic ductal adenocarcinomas (PDACs) from solid pseudopapillary neoplasms (SPNs) and pancreatic neuroendocrine neoplasms (PNENs).
This retrospective study consisted of patients diagnosed with PDACs (n = 132), PNENs (n = 45) and SPNs (n = 54). All patients underwent 3.0-T MRI including DWI with 10 b values. The regions of interest (ROIs) of pancreatic tumor were manually drawn using ITK-SNAP software, which included entire tumor at DWI (b = 1500 s/m). The histogram array was obtained through the ROIs from multiple b-value data. PyTorch (version 1.11) was used to construct a CNN classifier to categorize the histogram array into PDACs, PNENs or SPNs.
The area under the curves (AUCs) of the histogram array and the CNN model for differentiating PDACs from PNENs and SPNs were 0.896, 0.846, and 0.839 in the training, validation and testing cohorts, respectively. The accuracy, sensitivity and specificity were 90.22%, 96.23%, and 82.05% in the training cohort, 84.78%, 96.15%, and 70.0% in the validation cohort, and 81.72%, 90.57%, and 70.0% in the testing cohort. The performance of CNN with AUC of 0.865 for this differentiation was significantly higher than that of f with AUC = 0.755 (P = 0.0057) and α with AUC = 0.776 (P = 0.0278) in all patients.
The histogram array and CNN based on DWI data with multiple b-values using MRI provided an accurate diagnostic performance to differentiate PDACs from PNENs and SPNs.
本研究旨在探讨基于磁共振成像（MRI）下多b值扩散加权成像（DWI）的直方图阵列和卷积神经网络（CNN）区分胰腺导管腺癌（PDAC）与实性假乳头状瘤（SPN）及胰腺神经内分泌肿瘤（PNEN）的诊断性能。
本回顾性研究纳入了诊断为PDAC（n = 132）、PNEN（n = 45）和SPN（n = 54）的患者。所有患者均接受3.0-T MRI检查，包括具有10个b值的DWI检查。使用ITK-SNAP软件手动绘制胰腺肿瘤的感兴趣区域（ROI），其中包括DWI（b = 1500 s/m）时的整个肿瘤。通过多个b值数据的ROI获得直方图阵列。使用PyTorch（版本1.11）构建CNN分类器，将直方图阵列分类为PDAC、PNEN或SPN。
在训练、验证和测试队列中，直方图阵列和CNN模型区分PDAC与PNEN及SPN的曲线下面积（AUC）分别为0.896、0.846和0.839。训练队列中的准确率、敏感性和特异性分别为90.22%、96.23%和82.05%，验证队列中分别为84.78%、96.15%和70.0%，测试队列中分别为81.72%、90.57%和70.0%。在所有患者中，用于这种区分的AUC为0.865的CNN性能显著高于AUC = 0.755的f（P = 0.0057）和AUC = 0.776的α（P = 0.0278）。
基于MRI多b值DWI数据的直方图阵列和CNN在区分PDAC与PNEN及SPN方面具有准确的诊断性能。
