Deep Learning via Fused Bidirectional Attention Stacked Long Short-Term Memory for Obsessive-Compulsive Disorder Diagnosis and Risk Screening

摘要

The compulsive urges to perform stereotyped behaviors are typical symptoms of obsessive-compulsive disorder (OCD). OCD has certain hereditary tendencies and the direct OCD relatives (i.e., sibling (Sib)) have 50% of the same genes as patients. Sib has a higher probability of suffering from the same disease. Resting-state functional magnetic resonance imaging (R-fMRI) has made great progress by diagnosing OCD and identifying its high-risk population. Accordingly, we design a new deep learning framework for OCD diagnosis via R-fMRI data. Specifically, the fused bidirectional attention stacking long short-term memory (FBAS-LSTM) is exploited. First, we obtain two independent time series from the original R-fMRI by frame separation, which can reduce the length of R-fMRI sequence and alleviate the training difficulty. Second, we apply two independent BAS-LSTM learning on the hidden spatial information to obtain preliminary classification results. Lastly, the final diagnosis results are obtained by voting from the two diagnostic results. We validate our method on our in-house dataset including 62 OCD, 53 siblings (Sib) and 65 healthy controls (HC). Our method achieves average accuracies of 71.66% for differentiating OCD vs. Sib vs. HC, and outperforms the related algorithms.