Background: This study analyzed different classifier models for differentiating pancreatic adenocarcinoma from surrounding healthy pancreatic tissue based on radiomic analysis of magnetic resonance (MR) images. Methods: We observed T2W-FS and ADC images obtained by 1.5T-MR of 87 patients with histologically proven pancreatic adenocarcinoma for training and validation purposes and then tested the most accurate predictive models that were obtained on another group of 58 patients. The tumor and surrounding pancreatic tissue were segmented on three consecutive slices, with the largest area of interest (ROI) of tumor marked using MaZda v4.6 software. This resulted in a total of 261 ROIs for each of the observed tissue classes in the training–validation group and 174 ROIs in the testing group. The software extracted a total of 304 radiomic features for each ROI, divided into six categories. The analysis was conducted through six different classifier models with six different feature reduction methods and five-fold subject-wise cross-validation. Results: In-depth analysis shows that the best results were obtained with the Random Forest (RF) classifier with feature reduction based on the Mutual Information score (all nine features are from the co-occurrence matrix): an accuracy of 0.94/0.98, sensitivity of 0.94/0.98, specificity of 0.94/0.98, and F1-score of 0.94/0.98 were achieved for the T2W-FS/ADC images from the validation group, retrospectively. In the testing group, an accuracy of 0.69/0.81, sensitivity of 0.86/0.82, specificity of 0.52/0.70, and F1-score of 0.74/0.83 were achieved for the T2W-FS/ADC images, retrospectively. Conclusions: The machine learning approach using radiomics features extracted from T2W-FS and ADC achieved a relatively high sensitivity in the differentiation of pancreatic adenocarcinoma from healthy pancreatic tissue, which could be especially applicable for screening purposes.
背景：本研究基于磁共振（MR）图像的影像组学分析，探讨不同分类器模型在区分胰腺导管腺癌与周围正常胰腺组织中的效能。方法：我们收集了87例经组织学证实的胰腺导管腺癌患者的1.5T磁共振T2加权脂肪抑制序列（T2W-FS）和表观扩散系数（ADC）图像用于训练与验证，并将所得最优预测模型在另一组58例患者中进行测试。使用MaZda v4.6软件在连续三层图像上勾画肿瘤及周围胰腺组织，并以肿瘤最大感兴趣区域（ROI）为标记，最终训练-验证组中每类组织获得261个ROI，测试组获得174个ROI。软件从每个ROI提取共304个影像组学特征，分为六类。通过六种不同分类器模型结合六种特征降维方法，并采用五折受试者交叉验证进行分析。结果：深入分析表明，基于互信息评分进行特征降维的随机森林分类器（所有九个特征均来自共生矩阵）取得最佳结果：验证组中T2W-FS/ADC图像的准确度为0.94/0.98，敏感度为0.94/0.98，特异度为0.94/0.98，F1分数为0.94/0.98；测试组中T2W-FS/ADC图像的准确度为0.69/0.81，敏感度为0.86/0.82，特异度为0.52/0.70，F1分数为0.74/0.83。结论：基于T2W-FS和ADC图像提取影像组学特征的机器学习方法在区分胰腺导管腺癌与正常胰腺组织时具有较高的敏感度，尤其适用于筛查场景。
肿瘤(癌症)患者之家