Identifying who has reading disability from brain structure images has been challenging because reading disability is a complex disorder. Using a deep learning approach, we were able to predict reading disability cases from controls with a ~77% accuracy using brain morphology data. An autoencoder was used to first reduce the dimensionality of whole brain image data and then a convolutional neural network was trained to identify reading disability cases form controls using the latent representations of the images from the autoencoder. Perhaps most exciting, an image perturbation approach was used to identify the brain regions that contributed to accurate classification and has the potential to identify the specific regions in each participants image that contributed to classification. That is, the approach can identify the brain regions in individuals that may be atypical and contribute to reading disability or that are protective from risk for reading disability.