Abstract：
Extracellular volume fraction (ECV) by cardiac magnetic resonance (CMR) allows for the non-invasive quantification of diffuse myocardial fibrosis. Texture analysis and machine learning are now gathering attention in the medical field to exploit the ability of diagnostic imaging for various diseases. This study aimed to investigate the predictive value of texture analysis of ECV and machine learning for predicting response to guideline-directed medical therapy (GDMT) for patients with non-ischemic dilated cardiomyopathy (NIDCM). A total of one-hundred and fourteen NIDCM patients [age: 63 ± 12 years, 91 (81%) males] were retrospectively analyzed. We performed texture analysis of ECV mapping of LV myocardium using dedicated software. We calculated nine histogram-based features (mean, standard deviation, maximum, minimum, etc.) and five gray-level co-occurrence matrices. Five machine learning techniques and the fivefold cross-validation method were used to develop prediction models for LVRR by GDMT based on 14 texture parameters on ECV mapping. We defined the LVRR as follows: LVEF increased ≥ 10% points and decreased LVEDV ≥ 10% on echocardiography after GDMT > 12 months. Fifty (44%) patients were classified as non-responders. The area under the receiver operating characteristics curve for predicting non-responder was 0.82 for eXtreme Gradient Boosting, 0.85 for support vector machine, 0.76 for multi-layer perception, 0.81 for Naïve Bayes, 0.77 for logistic regression, respectively. Mean ECV value was the most critical factor among texture features for differentiating NIDCM patients with LVRR and those without (0.28 ± 0.03 vs. 0.36 ± 0.06, p < 0.001). Machine learning analysis using the support vector machine may be helpful in detecting high-risk NIDCM patients resistant to GDMT. Mean ECV is the most crucial feature among texture features.
摘要：
通过心脏磁共振（CMR）的细胞外体积分数（ECV）允许对弥漫性心肌纤维化进行非侵入性量化。纹理分析和机器学习现在正在医学领域引起人们的关注，以利用各种疾病的诊断成像能力。本研究旨在探讨ECV纹理分析和机器学习对非缺血性扩张型心肌病（NIDCM）患者指南指导药物治疗（GDMT）反应的预测价值。共有114名NIDCM患者[年龄：63±12岁，91（81%）男性]进行回顾性分析。我们使用专用软件对LV心肌的ECV作图进行了纹理分析。我们计算了九个基于直方图的特征(平均值，标准偏差，最大值,minimum,等。)和五个灰度共生矩阵。使用五种机器学习技术和五次交叉验证方法，基于ECV映射上的14个纹理参数，通过GDMT开发了LVRR的预测模型。我们将LVRR定义为：GDMT>12个月后，超声心动图显示LVEF增加≥10%，LVEDV降低≥10%。50名（44％）患者被归类为无应答者。用于预测无反应者的接收器操作特征曲线下的面积对于极限梯度增强为0.82，支持向量机的0.85，0.76用于多层感知，对于朴素贝叶斯，为0.81，逻辑回归为0.77，分别。平均ECV值是区分具有LVRR和无LVRR的NIDCM患者的纹理特征中最关键的因素（0.28±0.03vs.0.36±0.06，p<0.001)。使用支持向量机的机器学习分析可能有助于检测对GDMT耐药的高风险NIDCM患者。平均ECV是纹理特征中最关键的特征。
