基于加速扩散张量成像的低级别胶质瘤分子亚型分类
Molecular Subtype Classification in Lower-Grade Glioma with Accelerated DTI.
机构信息
From the Departments of Radiation Oncology (E.A., H.N.)
From the Departments of Radiation Oncology (E.A., H.N.).
出版信息
AJNR Am J Neuroradiol. 2019 Sep;40(9):1458-1463. doi: 10.3174/ajnr.A6162. Epub 2019 Aug 14.
BACKGROUND AND PURPOSE
Image-based classification of lower-grade glioma molecular subtypes has substantial prognostic value. Diffusion tensor imaging has shown promise in lower-grade glioma subtyping but currently requires lengthy, nonstandard acquisitions. Our goal was to investigate lower-grade glioma classification using a machine learning technique that estimates fractional anisotropy from accelerated diffusion MR imaging scans containing only 3 diffusion-encoding directions.
MATERIALS AND METHODS
Patients with lower-grade gliomas ( = 41) (World Health Organization grades II and III) with known () mutation and 1p/19q codeletion status were imaged preoperatively with DTI. Whole-tumor volumes were autodelineated using conventional anatomic MR imaging sequences. In addition to conventional ADC and fractional anisotropy reconstructions, fractional anisotropy estimates were computed from 3-direction DTI subsets using DiffNet, a neural network that directly computes fractional anisotropy from raw DTI data. Differences in whole-tumor ADC, fractional anisotropy, and estimated fractional anisotropy were assessed between -wild-type and -mutant lower-grade gliomas with and without 1p/19q codeletion. Multivariate classification models were developed using whole-tumor histogram and texture features from ADC, ADC + fractional anisotropy, and ADC + estimated fractional anisotropy to identify the added value provided by fractional anisotropy and estimated fractional anisotropy.
RESULTS
ADC ( = .008), fractional anisotropy ( < .001), and estimated fractional anisotropy ( < .001) significantly differed between -wild-type and -mutant lower-grade gliomas. ADC ( < .001) significantly differed between -mutant gliomas with and without codeletion. ADC-only multivariate classification predicted mutation status with an area under the curve of 0.81 and codeletion status with an area under the curve of 0.83. Performance improved to area under the curve = 0.90/0.94 for the ADC + fractional anisotropy classification and to area under the curve = 0.89/0.89 for the ADC + estimated fractional anisotropy classification.
CONCLUSIONS
Fractional anisotropy estimates made from accelerated 3-direction DTI scans add value in classifying lower-grade glioma molecular status.
背景与目的
基于影像的低级别胶质瘤分子亚型分类具有重要的预后价值。弥散张量成像在低级别胶质瘤亚型分类中有很大的应用前景，但目前需要进行冗长的、非标准的采集。我们的目标是研究一种机器学习技术，该技术可以从仅包含 3 个扩散编码方向的加速扩散 MRI 扫描中估计各向异性分数，从而对低级别胶质瘤进行分类。
材料与方法
对 41 例（WHO 分级 II 级和 III 级）已知存在 IDH1/2 突变和 1p/19q 共缺失状态的低级别胶质瘤患者进行术前弥散张量成像（DTI）检查。使用常规解剖学 MRI 序列自动勾画全肿瘤体积。除了常规的 ADC 和各向异性分数重建外，还使用 DiffNet 从 3 方向 DTI 子集计算各向异性分数，DiffNet 是一种直接从原始 DTI 数据计算各向异性分数的神经网络。在存在和不存在 1p/19q 共缺失的情况下，比较 IDH1/2 野生型和突变型低级别胶质瘤的全肿瘤 ADC、各向异性分数和估计的各向异性分数之间的差异。使用来自 ADC、ADC+各向异性分数和 ADC+估计的各向异性分数的全肿瘤直方图和纹理特征，建立多变量分类模型，以确定各向异性分数和估计的各向异性分数提供的附加值。
结果
在 IDH1/2 野生型和突变型低级别胶质瘤之间，ADC（ =.008）、各向异性分数（ <.001）和估计的各向异性分数（ <.001）差异有统计学意义。在存在和不存在 1p/19q 共缺失的情况下，IDH1/2 突变型低级别胶质瘤之间的 ADC（ <.001）差异有统计学意义。仅 ADC 的多变量分类预测 IDH1/2 突变状态的曲线下面积为 0.81，预测 1p/19q 共缺失状态的曲线下面积为 0.83。当加入各向异性分数后，分类的曲线下面积提高到 0.90/0.94，当加入估计的各向异性分数后，分类的曲线下面积提高到 0.89/0.89。
结论
从加速的 3 方向 DTI 扫描中获得的各向异性分数估计值可提高低级别胶质瘤分子状态的分类价值。
Zotero 插件