Department of Neurology (H.Z., L.T., B.X.), Xiangya Hospital of Central South University, Changsha, Hunan, China.
From the School of Computer Science and Engineering (R.H., B.Z., C.Z.).
AJNR Am J Neuroradiol. 2020 Jul;41(7):1279-1285. doi: 10.3174/ajnr.A6621.
Differentiating the types of pediatric posterior fossa tumors on routine imaging may help in preoperative evaluation and guide surgical resection planning. However, qualitative radiologic MR imaging review has limited performance. This study aimed to compare different machine learning approaches to classify pediatric posterior fossa tumors on routine MR imaging.
This retrospective study included preoperative MR imaging of 288 patients with pediatric posterior fossa tumors, including medulloblastoma ( = 111), ependymoma ( = 70), and pilocytic astrocytoma ( = 107). Radiomics features were extracted from T2-weighted images, contrast-enhanced T1-weighted images, and ADC maps. Models generated by standard manual optimization by a machine learning expert were compared with automatic machine learning via the Tree-Based Pipeline Optimization Tool for performance evaluation.
For 3-way classification, the radiomics model by automatic machine learning with the Tree-Based Pipeline Optimization Tool achieved a test micro-averaged area under the curve of 0.91 with an accuracy of 0.83, while the most optimized model based on the feature-selection method χ score and the Generalized Linear Model classifier achieved a test micro-averaged area under the curve of 0.92 with an accuracy of 0.74. Tree-Based Pipeline Optimization Tool models achieved significantly higher accuracy than average qualitative expert MR imaging review (0.83 versus 0.54, < .001). For binary classification, Tree-Based Pipeline Optimization Tool models achieved an area under the curve of 0.94 with an accuracy of 0.85 for medulloblastoma versus nonmedulloblastoma, an area under the curve of 0.84 with an accuracy of 0.80 for ependymoma versus nonependymoma, and an area under the curve of 0.94 with an accuracy of 0.88 for pilocytic astrocytoma versus non-pilocytic astrocytoma.
Automatic machine learning based on routine MR imaging classified pediatric posterior fossa tumors with high accuracy compared with manual expert pipeline optimization and qualitative expert MR imaging review.
在常规影像学上区分小儿后颅窝肿瘤的类型有助于术前评估和指导手术切除计划。然而，定性放射影像学 MR 成像检查的性能有限。本研究旨在比较不同的机器学习方法在小儿后颅窝肿瘤的常规 MR 成像上的分类表现。
本回顾性研究纳入了 288 例小儿后颅窝肿瘤患者的术前 MR 成像资料，包括髓母细胞瘤（ = 111 例）、室管膜瘤（ = 70 例）和毛细胞星形细胞瘤（ = 107 例）。从 T2 加权图像、对比增强 T1 加权图像和 ADC 图中提取放射组学特征。通过机器学习专家进行标准手动优化生成的模型与通过 Tree-Based Pipeline Optimization Tool 的自动机器学习进行比较，用于性能评估。
对于 3 分类，基于 Tree-Based Pipeline Optimization Tool 的自动机器学习的放射组学模型在测试中的微平均曲线下面积为 0.91，准确率为 0.83，而基于特征选择方法 χ 评分和广义线性模型分类器的最优化模型在测试中的微平均曲线下面积为 0.92，准确率为 0.74。Tree-Based Pipeline Optimization Tool 模型的准确率显著高于平均定性专家 MR 成像评估（0.83 与 0.54，<.001）。对于 2 分类，Tree-Based Pipeline Optimization Tool 模型在髓母细胞瘤与非髓母细胞瘤的分类中，曲线下面积为 0.94，准确率为 0.85；在室管膜瘤与非室管膜瘤的分类中，曲线下面积为 0.84，准确率为 0.80；在毛细胞星形细胞瘤与非毛细胞星形细胞瘤的分类中，曲线下面积为 0.94，准确率为 0.88。
与手动专家管道优化和定性专家 MR 成像评估相比，基于常规 MR 成像的自动机器学习对小儿后颅窝肿瘤的分类具有较高的准确性。
