Abstract:It is important to be able to predict the probability of whether MCI patients would convert to AD. The research goal is to use AI approach to develop automatic, neuroimage-based diagnostic and, especially, prognostic classifiers to identify MCI subjects who will convert to AD within several years. Multilayer perceptron (MLP) is used for developing our classifiers as MLP has been shown to be able to address large nonlinear problems effectively. MRI images of 288 MCI participants were selected according to the first visit diagnosis from the ADNI database and divided into a stable-MCI (n=112) and a convert-MCI (n=176) group based on the third years’ follow-up examination as the ground truth (Table. 1, Fig. 1). To reduce the dimensionality of feature space, all the first-visit images of were segmented into 18 structures (Table.2) according to MNI152 atlas by FSL and their volumes were calculated. A five-level MLP classifier was created, whose nodes of three hidden layers were 1024, 512, and 256 respectively. The volumes of 18 structures and the subject's age were chosen as features. The MLP was trained with backpropagation algorithm combined with Adam optimization algorithm. Clinical characteristic of ANDI participants. Note: BL: baseline, 3yt: 3 years timepoint There were no significant differences in the 18 structure volumes between the two groups except brainstem, right amygdala, WM and GM (Fig.2). Using only the first-visit images, the classify accuracy is 87.89% with MLP classifier. MCI is a heterogeneous cohort. The gray-scaled brain images are high-dimensional (>10 millions) and contain large variations on top of the individual anatomy and mixed pathologies. This would require a greater “big-data” set for AI applications. Using the volumes of major brains structures allowed us to drastically reduce the dimensions. With only 288 samples, we were able to build a MLP classifier with high prediction accuracy. This work presented a novel deep-learning approach where ground-truth is poorly defined. Further validation of classifiers to identify and stage subjects likely to progress to AD is required in future studies. AI with feature space reduction using quantitative MRI is an effective method for development of accurate prognostic classifier to predict MCI subjects convert to AD.

IC‐P‐171: PREDICTING MCI‐TO‐AD CONVERSION USING NEUROIMAGING CLASSIFIER WITH ARTIFICIAL INTELLIGENCE (AI)