Joint Neuroimage Synthesis and Representation Learning for Conversion Prediction of Subjective Cognitive Decline

摘要

Predicting the progression of preclinical Alzheimer's disease (AD) such as subjective cognitive decline (SCD) is fundamental for the effective intervention of pathological cognitive decline. Even though mul-timodal neuroimaging has been widely used in automated AD diagnosis, there are few studies dedicated to SCD progression prediction, due to challenges of incomplete and limited data. To this end, we propose a Joint neuroimage Synthesis and Representation Learning (JSRL) framework with transfer learning for SCD conversion prediction using incomplete multimodal neuroimaging data. Specifically, JSRL consists of two major components: 1) a generative adversarial network for synthesizing missing neuroimaging data, and 2) a classification network for learning neuroimage representations and predicting the progression of SCD. These two subnetworks share the same feature encoding module, encouraging that the to-be-generated representations are prediction-oriented and also the underlying association among multimodal images can be effectively modeled for accurate prediction. To handle the limited data problem, we further leverage both image synthesis and prediction models learned from a large-scale ADNI database (with MRI and PET acquired from 863 subjects) to a small-scale SCD database (with only MRI acquired from 113 subjects) in a transfer learning manner. Experimental results show that the proposed JSRL can synthesize reasonable PET scans and is superior to several state-of-the-art methods in SCD conversion prediction.