Abstract
BACKGROUND: Deep neural network (DNN) has facilitated the record-breaking of classification accuracy in fields such as speech or visual object recognition. However, limited studies have investigated the applicability of DNN to three-dimensional neuroimage data, and the interpretation of deep learning model remains like a black box. Here, we present an explainable DNN framework to identify key structural deficits in schizophrenia. METHODS: Structural brain magnetic resonance images (MRI) were obtained from 200 schizophrenic patients and 200 age- and sex-matched healthy control subjects. The brain MRI images were normalized and segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) space. We introduced an original KL-L1 regularization method as a feature selection layer in the DNN to reduce dimensionality of neuroimage data and automatically identify key brain voxels without prior knowledge of brain pathology. RESULTS: The DNN classifier with KL-L1 regularization achieved an average test accuracy of 91.7% in WM, an average of 87.5% in GM, and 75.5% in CSF. The key GM voxels identified by the DNN were within brain regions including insula, precuneus, and superior temporal pole; WM voxels were associated with neural tracts, such as cingulum/hippocampus, splenium of corpus callosum, and posterior corona radiata. DISCUSSION: The present study shows that the DNN with KL-L1 regularization can identify key structural deficits that are effectively related to the known structural pathology of schizophrenia. We anticipate that this explainable deep learning approach may provide a useful framework for the search of objective biomarkers of mental illness in future studies.