Texture analysis and machine learning of non-contrast T1-weighted MR images in patients with hypertrophic cardiomyopathy-Preliminary results
Texture analysis and machine learning of non-contrast T1-weighted MR images in patients with hypertrophic cardiomyopathy-Preliminary results
复制标题
DOI：
10.1016/j.ejrad.2018.03.013
复制
发表时间：
2018-05-01
影响因子：
3.3
通讯作者：
Manka, Robert
中科院分区：
文献类型：
--
作者：
Baessler, Bettina;Mannil, Manoj;Manka, Robert
目的：在第一个概念验证研究中测试纹理分析（TA）是否允许使用基于机器学习的方法在非对比T1加权心脏磁共振（CMR）图像上检测肥厚型心肌病（HCM）的心肌组织改变：方法：这项回顾性研究，IRB批准的研究包括32例已知HCM患者。30例CMR正常的患者作为对照。使用免费提供的软件包在短轴非造影T1加权图像上绘制包括左心室的TA感兴趣区域。逐步降维和纹理特征选择进行选择的功能，使检测心肌组织的变化，在HCM患者的非对比T1加权CMR images.Results：比较HCM患者和对照组，四个纹理特征被确定显示组间差异显着（灰度级非均匀性[GLevNonU]：74 +/-17对比38 +/-9，p < .001;低频子带中小波系数的能量[WavEnLL]：58 +/-5对比48 +/-10，p < .001;分数：0.70 +/- 0.07对比0.78 +/- 0.05，p < .001;总和平均值：16.6 +/- 0.4对比17.0 +/- 0.5，p = .007）。包含单参数GLevNonU的模型被证明是区分HCM患者和对照的最佳模型，灵敏度/特异性为91%/93%。GLevNonU >= 46的截止值允许以94%/90%的灵敏度/特异性区分HCM患者和对照。即使在没有晚期钆增强（LGE）的患者中，定义的截止导致LGE患者与健康对照组的分化，具有100%的灵敏度和90% specificity.Conclusions：TA的非对比T1加权图像允许检测心肌组织的变化，在设置HCM具有良好的准确性，提供潜在的新参数，心肌纹理改变的非对比评估。
Purpose: To test in a first proof-of-concept study whether texture analysis (TA) allows for the detection of myocardial tissue alterations in hypertrophic cardiomyopathy (HCM) on non-contrast T1-weighted cardiac magnetic resonance (CMR) images using machine learning based approaches.Methods: This retrospective, IRB-approved study included 32 patients with known HCM. Thirty patients with normal CMR served as controls. Regions-of-interest for TA encompassing the left ventricle were drawn on short-axis non-contrast T1-weighted images using a freely available software package. Step-wise dimension reduction and texture feature selection was performed for selecting features enabling the detection of myocardial tissue alterations in HCM patients on non-contrast T1-weighted CMR images.Results: Comparing HCM patients and controls, four texture features were identified showing significant differences between groups (Grey-level Non-uniformity [GLevNonU]: 74 +/- 17 vs. 38 +/- 9, p < .001; Energy of wavelet coefficients in low-frequency sub-bands [WavEnLL]: 58 +/- 5 vs. 48 +/- 10, p < .001; Fraction: 0.70 +/- 0.07 vs. 0.78 +/- 0.05, p < .001; Sum Average: 16.6 +/- 0.4 vs. 17.0 +/- 0.5, p = .007). A model containing the single parameter GLevNonU proved to be the best for differentiating between HCM patients and controls with a sensitivity/specificity of 91%/93%. A cut-off of GLevNonU >= 46 allowed for distinguishing HCM patients from controls with a sensitivity/specificity of 94%/90%. Even in patients without late gadolinium enhancement (LGE), the defined cut-off led to a differentiation of LGE-patients from healthy controls with 100% sensitivity and 90% specificity.Conclusions: TA on non-contrast T1-weighted images allows for the detection of myocardial tissue alterations in the setting of HCM with excellent accuracy, delivering potential novel parameters for a non-contrast assessment of myocardial texture alterations.