Classification of Subcortical Vascular Cognitive Impairment Using Single MRI Sequence and Deep Learning Convolutional Neural Networks

Deep learning has great potential for imaging classification by extracting low to high-level features. Our aim was to train a convolutional neural network (CNN) with single T2-weighted FLAIR sequence to classify different cognitive performances in patients with subcortical ischemic vascular disease (SIVD). In total, 217 patients with SIVD [including 52 with vascular dementia (VaD), 82 with vascular mild cognitive impairment (VaMCI), and 83 with non-cognitive impairment (NCI)] and 46 matched healthy controls (HCs) underwent MRI scans and neuropsychological assessments. 2D and 3D CNNs were trained to classify VaD, VaMCI, NCI, and HCs based on FLAIR data. For 3D-based model, the loss curves of the training set approached 0.017 after about 20 epochs, while the curves of the testing set maintained at about 0.114. The accuracy of training set and testing set reached 99.7 and 96.9% after about 30 and 35 epochs, respectively. However, the accuracy of the 2D-based model was only around 70%, which performed significantly worse than 3D-based model. This experiment suggests that deep learning is a powerful and convenient method to classify different cognitive performances in SIVD by extracting the shift and scale invariant features of neuroimaging data with single FLAIR sequence. 3D-CNN is superior to 2D-CNN which involves clinical evaluation with MRI multiplanar reformation or volume scanning.