Chen Xiaowen, Tang Mingyue, Liu Aimin, Wei Xiaoqin
School of Medical Imaging, North Sichuan Medical College, Nanchong, China.
School of Basic Medicine and Forensic Medicine, North Sichuan Medical College, Nanchong, China.
Ann Transl Med. 2022 Jul;10(14):765. doi: 10.21037/atm-22-2961.
Alzheimer's disease (AD) is a widespread neurodegenerative disease that mostly affects the elderly population. Given its prevalence, a precise and efficient stratification system based on AD symptomology that uses functional magnetic resonance imaging (MRI) has great potential in the clinical diagnosis and prognosis estimation of AD patients. It was evident that deep learning methods have performed extremely well in the field of automated stratification of AD based on MRI because of their high predicting accuracy and reliability.
We proposed a deep convolutional neural network (CNN) and iterated random forest (RF) architecture for MRI image stratification by both anatomical location and image modality using the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. We employed 3 cross-sectional data sets from the ADNI to conduct our binary-stratification [AD and normal controls (NCs), or AD and mild cognitive impairment (MCI)], and multi-stratification (AD, MCI, and NCs) process using MRI. And the accuracy, recall, specificity, area under the curve of receiver operating characteristic curve (AUC), F1 and Matthew's correlation coefficient (MCC) scores to assess accuracy of auxiliary clinical diagnoses.
Compared to other combinations of algorithms, our model obtained remarkable overall stratification accuracies in all different classification sets. In terms of AD MCI, the mean training AUC of the 3 runs were 85.1% in 95% confidence intervals (CIs). In terms of AD NC, the mean training AUC of the 3 runs was 90.6% in 95% CIs. In terms of the 3 stratifications of AD, MCI, and NC, relative precision, recall, and specificity for each category in the training test (TS) were all near 89%, while the F1 and MCC scores of both sets were 59.9% and 59.5%, respectively.
Using a deep CNN and iterated RF architecture, we showed that brain image stratification is a promising means for evaluating AD, and examining the underlying etiology of the disease, by applying computer and medical images to achieve the early auxiliary diagnosis of AD. However, we still have a long way to go from the discovery of image markers to clinical application.
阿尔茨海默病（AD）是一种广泛存在的神经退行性疾病，主要影响老年人群。鉴于其患病率，基于AD症状学并使用功能磁共振成像（MRI）的精确高效分层系统在AD患者的临床诊断和预后评估中具有巨大潜力。很明显，深度学习方法因其高预测准确性和可靠性，在基于MRI的AD自动分层领域表现极为出色。
我们使用阿尔茨海默病神经影像倡议（ADNI）数据库，提出了一种深度卷积神经网络（CNN）和迭代随机森林（RF）架构，用于按解剖位置和图像模态对MRI图像进行分层。我们采用来自ADNI的3个横断面数据集，使用MRI进行二元分层[AD与正常对照（NC），或AD与轻度认知障碍（MCI）]以及多分层（AD、MCI和NC）过程。并使用准确性、召回率、特异性、受试者操作特征曲线（ROC）下面积（AUC）、F1和马修斯相关系数（MCC）分数来评估辅助临床诊断的准确性。
与其他算法组合相比，我们的模型在所有不同分类集中均获得了显著的总体分层准确性。在AD与MCI方面，3次运行的平均训练AUC在95%置信区间（CI）内为85.1%。在AD与NC方面，3次运行的平均训练AUC在95%CI内为90.6%。在AD、MCI和NC的三重分层方面，训练测试（TS）中每个类别的相对精度、召回率和特异性均接近89%，而两组的F1和MCC分数分别为59.9%和59.5%。
通过使用深度CNN和迭代RF架构，我们表明脑图像分层是评估AD以及通过应用计算机和医学图像来检查疾病潜在病因以实现AD早期辅助诊断的一种有前景的手段。然而，从发现图像标志物到临床应用，我们仍有很长的路要走。
