Department of Neurosurgery, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, 430022 Wuhan, Hubei, China.
J Integr Neurosci. 2021 Sep 30;20(3):623-634. doi: 10.31083/j.jin2003066.
A correct preoperative diagnosis is essential for the treatment and prognosis of necrotic glioblastoma and brain abscess, but the differentiation between them remains challenging. We constructed a diagnostic prediction model with good performance and enhanced clinical applicability based on data from 86 patients with necrotic glioblastoma and 32 patients with brain abscess that were diagnosed between January 2012 and January 2020. The diagnostic values of three regions of interest based on contrast-enhanced T1 weighted images (including whole tumor, brain-tumor interface, and an amalgamation of both regions) were compared using Logistics Regression and Random Forest. Feature reduction based on the optimal regions of interest was performed using principal component analysis with varimax rotation. The performance of the classifiers was assessed by receiver operator curves. Finally, clinical predictors were utilized to detect the diagnostic power. The mean area under curve (AUC) values of the whole tumor model was significantly higher than other two models obtained from Brain-Tumor Interface (BTI) and combine regions both in training (AUC mean = 0.850) and test/validation set (AUC mean = 0.896) calculated by Logistics Regression and in the testing set (AUC mean = 0.876) calculated by Random Forest. Among these three diagnostic prediction models, the combined model provided superior discrimination performance and yielded an AUC of 0.993, 0.907, and 0.974 in training, testing, and combined datasets, respectively. Compared with the brain-tumor interface and the combined regions, features obtained from the whole tumor showed the best differential value. The radiomic features combined with the peritumoral edema/tumor volume ratio provided the prediction model with the greatest diagnostic performance.
正确的术前诊断对于治疗和预测坏死性脑胶质瘤和脑脓肿至关重要，但两者的鉴别仍然具有挑战性。我们基于 2012 年 1 月至 2020 年 1 月间诊断的 86 例坏死性脑胶质瘤和 32 例脑脓肿患者的数据，构建了一个具有良好性能和增强临床适用性的诊断预测模型。利用逻辑回归和随机森林比较了基于增强 T1 加权图像的三个感兴趣区（包括全肿瘤、肿瘤-脑界面和两者的融合区）的诊断价值。利用主成分分析和方差极大旋转对基于最优感兴趣区的特征进行降维。通过接受者操作特征曲线评估分类器的性能。最后，利用临床预测因子检测诊断能力。在训练和测试/验证集中，全肿瘤模型的平均曲线下面积（AUC）值均显著高于基于肿瘤-脑界面（BTI）和融合区的模型（AUC 均值分别为 0.850 和 0.896），在随机森林计算的测试集中 AUC 均值为 0.876。在这三个诊断预测模型中，联合模型提供了卓越的鉴别性能，在训练、测试和联合数据集的 AUC 分别为 0.993、0.907 和 0.974。与肿瘤-脑界面和融合区相比，全肿瘤区的特征显示出最佳的区分价值。肿瘤周围水肿/肿瘤体积比与放射组学特征相结合，为预测模型提供了最佳的诊断性能。
