Institute for Nuclear Medicine and Clinical Molecular Imaging, Eberhard Karls University, Tuebingen, Germany.
German Center of Neurodegenerative Diseases, Eberhard Karls University, Tuebingen, Germany.
Eur J Nucl Med Mol Imaging. 2019 Oct;46(11):2370-2379. doi: 10.1007/s00259-019-04400-w. Epub 2019 Jul 24.
The pattern expression score (PES), i.e., the degree to which a pathology-related pattern is present, is frequently used in FDG-brain-PET analysis and has been shown to be a powerful predictor of conversion to Alzheimer's disease (AD) in mild cognitive impairment (MCI). Since, inevitably, the PES is affected by non-pathological variability, our aim was to improve classification with the simple, yet novel approach to identify patterns of non-pathological variance in a separate control sample using principal component analysis and removing them from patient data (controls-based denoising, CODE) before calculating the PES.
Multi-center FDG-PET from 220 MCI patients (64 non-converter, follow-up ≥ 4 years; 156 AD converter, time-to-conversion ≤ 4 years) were obtained from the ADNI database. Patterns of non-pathological variance were determined from 262 healthy controls. An AD pattern was calculated from AD patients and controls. We predicted AD conversion based on PES only and on PES combined with neuropsychological features and ApoE4 genotype. We compared classification performance achieved with and without CODE and with a standard machine learning approach (support vector machine).
Our model predicts that CODE improves the signal-to-noise ratio of AD-PES by a factor of 1.5. PES-based prediction of AD conversion improved from AUC 0.80 to 0.88 (p= 0.001, DeLong's method), sensitivity 69 to 83%, specificity 81% to 88% and Matthews correlation coefficient (MCC) 0.45 to 0.66. Best classification (0.93 AUC) was obtained when combining the denoised PES with clinical features.
CODE, applied in its basic form, significantly improved prediction of conversion based on PES. The achieved classification performance was higher than with a standard machine learning algorithm, which was trained on patients, explainable by the fact that CODE used additional information (large sample of healthy controls). We conclude that the proposed, novel method is a powerful tool for improving medical image analysis that offers a wide spectrum of biomedical applications, even beyond image analysis.
模式表达评分（PES），即病理相关模式的存在程度，常用于 FDG-脑-PET 分析，已被证明是轻度认知障碍（MCI）向阿尔茨海默病（AD）转化的有力预测因子。由于 PES 不可避免地受到非病理变化的影响，我们的目标是通过使用主成分分析在单独的对照样本中识别非病理变异模式的简单而新颖的方法来提高分类效果，并在计算 PES 之前从患者数据中去除这些模式（基于对照的去噪，CODE）。
从 ADNI 数据库中获得了 220 名 MCI 患者（64 名非转化者，随访时间≥4 年；156 名 AD 转化者，转化时间≤4 年）的多中心 FDG-PET。从 262 名健康对照中确定了非病理变异的模式。从 AD 患者和对照中计算出 AD 模式。我们仅基于 PES 以及基于 PES 结合神经心理学特征和 ApoE4 基因型来预测 AD 转化。我们比较了使用和不使用 CODE 以及使用标准机器学习方法（支持向量机）时的分类性能。
我们的模型预测，CODE 将 AD-PES 的信噪比提高了 1.5 倍。基于 PES 的 AD 转化预测从 AUC 0.80 提高到 0.88（p=0.001，DeLong 法），敏感性从 69%提高到 83%，特异性从 81%提高到 88%，马氏相关系数（MCC）从 0.45 提高到 0.66。当将去噪后的 PES 与临床特征相结合时，获得了最佳分类（0.93 AUC）。
在基本形式下，CODE 显著提高了基于 PES 的转化预测。所获得的分类性能高于使用患者训练的标准机器学习算法，这可以解释为 CODE 使用了额外的信息（大量健康对照）。我们得出结论，所提出的新颖方法是一种强大的医学图像分析工具，具有广泛的生物医学应用，甚至超越了图像分析。
