一、成果内容

1、Alzheimer's disease (AD) is a chronic, progressive, and irreversible neurodegenerative disease. According to the World Alzheimer Report 2018 published by Alzheimer’s Disease International (ADI) , the number of patients with AD worldwide is expected to increase from the current 47million to 152million by 2050. Moreover, the pathogenesis of AD is still not fully elucidated and no currently available therapy can cure AD or completely prevent disease progression. Structural MRI (sMRI) is primarily applied to non-invasively capture regional brain atrophy and contribute to understanding the anatomical changes of brain. In particular, the degree of atrophy of hippocampus and entorhinal cortex captured by sMRI can reflect the severity of disease to certain extent and predict the progression from Mild cognitive impairment (MCI) into AD. Hence, sMRI has been widely applied in the researches related to AD diagnosis. The effect of traditional feature extraction methods for sMRI structural changes largely depends on the preprocessing of imaging data, and the preprocessing largely relies upon the empirical knowledge. These subjective factors will exert significant effect on the results of AD diagnosis.
2、Relatively, as a data-driven machine learning method, deep learning does not yield these problems since it less depends on empirical experience. We propose a novel framework for the early diagnosis of AD using multi-slice sMRI learning features(Fig. 2), which utilizes CNN combined with GA and ensemble classification strategy to improve the performance of AD/NC, MCIc/NC, and MCIc/MCInc classification. Through the GA algorithm, the key base classifiers were selected, which can effectively learn the significant classification features contained in the corresponding slices. According to the slice coordinates on the three anatomical planes, the intersection positions can be determined, and then the brain regions where these intersection points are located can be determined by using Brainnetome Atlas, which can be regarded as ROI on MRI image. We carry out these three classification experiments on the sMR images of 787 subjects, compared the experimental results with the data used by other methods(Table 1). The experimental results demonstrate that the performance of our method has been significantly improved, and the obtained brain ROI include hippocampus and amygdala(Fig. 3), which are consistent with the existing AD biomarkers widely recognized by the medical community. At the same time, in this study, Brainnetome Atlas was utilized obtain the behavior domains corresponding to the ROI, and subsequently compared with the symptoms of AD to verify the potential correlation between the ROI and the incidence of AD. According to the statistical results of behavior domains (Fig. 4), the behavior domains corresponding to the ROI mainly included emotion, memory and language, which are basically consistent with the clinical symptoms of AD, such as apathy, amnesia, loss of mobility and language ability, etc., indicating that the ROI obtained in this study are significantly correlated with AD. In addition, we extracted the known ROI(such as hippocampus, amygdala, etc.) from 3D MRI and trained 3D-CNN with SVM (Random Forest, RF) in 3D data format(Fig. 6,7). Compared with the experimental results of other methods(Table 2), the classification effect of our model has also been greatly improved, which proves the feasibility and robustness of the method.