Xu Zhen, Wang Yu-Hong, Wang Ya-Lin, Feng You-Zhen, Ye Jin-Shao, Cheng Zhong-Yuan, Cai Xiang-Ran
Medical Imaging Center, First Affiliated Hospital of Jinan University, Guangzhou, China.
Department of Radiology, Academy of Orthopedics Guangdong Province, Third Affiliated Hospital of Southern Medical University, Guangzhou, China.
Quant Imaging Med Surg. 2024 May 1;14(5):3628-3642. doi: 10.21037/qims-23-1194. Epub 2024 Apr 7.
Due to the variations in surgical approaches and prognosis between intraspinal schwannomas and meningiomas, it is crucial to accurately differentiate between the two prior to surgery. Currently, there is limited research exploring the implementation of machine learning (ML) methods for distinguishing between these two types of tumors. This study aimed to establish a classification and regression tree (CART) model and a random forest (RF) model for distinguishing schwannomas from meningiomas.
We retrospectively collected 88 schwannomas (52 males and 36 females) and 51 meningiomas (10 males and 41 females) who underwent magnetic resonance imaging (MRI) examinations prior to the surgery. Simple clinical data and MRI imaging features, including age, sex, tumor location and size, T1-weighted images (T1WI) and T2-weighted images (T2WI) signal characteristics, degree and pattern of enhancement, dural tail sign, ginkgo leaf sign, and intervertebral foramen widening (IFW), were reviewed. Finally, a CART model and RF model were established based on the aforementioned features to evaluate their effectiveness in differentiating between the two types of tumors. Meanwhile, we also compared the performance of the ML models to the radiologists. The receiver operating characteristic (ROC) curve, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were used to evaluate the models and clinicians' discrimination performance.
Our investigation reveals significant variations in ten out of 11 variables in the training group and five out of 11 variables in the test group when comparing schwannomas and meningiomas (P<0.05). Ultimately, the CART model incorporated five variables: enhancement pattern, the presence of IFW, tumor location, maximum diameter, and T2WI signal intensity (SI). The RF model combined all 11 variables. The CART model, RF model, radiologist 1, and radiologist 2 achieved an area under the curve (AUC) of 0.890, 0.956, 0.681, and 0.723 in the training group, and 0.838, 0.922, 0.580, and 0.659 in the test group, respectively.
The RF prediction model exhibits more exceptional performance than an experienced radiologist in discriminating intraspinal schwannomas from meningiomas. The RF model seems to be better in discriminating the two tumors than the CART model.
由于椎管内神经鞘瘤和脑膜瘤的手术方式及预后存在差异，术前准确区分两者至关重要。目前，探索机器学习（ML）方法用于区分这两种肿瘤的研究有限。本研究旨在建立用于区分神经鞘瘤和脑膜瘤的分类回归树（CART）模型和随机森林（RF）模型。
我们回顾性收集了88例神经鞘瘤（男性52例，女性36例）和51例脑膜瘤（男性10例，女性41例），这些患者在手术前行磁共振成像（MRI）检查。回顾了简单的临床数据和MRI影像特征，包括年龄、性别、肿瘤位置和大小、T1加权像（T1WI）和T2加权像（T2WI）信号特征、强化程度和方式、硬脑膜尾征、银杏叶征以及椎间孔扩大（IFW）。最后，基于上述特征建立CART模型和RF模型，以评估它们在区分这两种肿瘤类型方面的有效性。同时，我们还将ML模型的性能与放射科医生的进行了比较。采用受试者操作特征（ROC）曲线、准确率、敏感性、特异性、阳性预测值（PPV）和阴性预测值（NPV）来评估模型和临床医生的鉴别性能。
我们的研究表明，在比较神经鞘瘤和脑膜瘤时，训练组11个变量中有10个、测试组11个变量中有5个存在显著差异（P<0.05）。最终，CART模型纳入了5个变量：强化方式、IFW的存在、肿瘤位置、最大直径和T2WI信号强度（SI）。RF模型结合了所有11个变量。CART模型、RF模型、放射科医生1和放射科医生2在训练组中的曲线下面积（AUC）分别为0.890、0.956、0.681和0.723，在测试组中分别为0.838、0.922、0.580和0.659。
在区分椎管内神经鞘瘤和脑膜瘤方面，RF预测模型比经验丰富的放射科医生表现更出色。RF模型在区分这两种肿瘤方面似乎比CART模型更好。
