Using Functional or Structural Magnetic Resonance Images and Personal Characteristic Data to Diagnose ADHD and Autism

Abstract

A clinical tool that can diagnose psychiatric illness using functional or structural magnetic resonance (MR) brain images would greatly assist physicians. Here, we propose a learning algorithm that uses the histogram of oriented gradients (HOG) features of MR brain images, as well as personal characteristic data, as features. We show that this learner can produce effective classifiers when run on two large public datasets. It is able to diagnose ADHD with hold-out accuracy of 0.696 (over baseline = 0.550) using personal characteristics and structural brain scan features when trained on the ADHD-200 global competition dataset and is also able to diagnose autism with hold-out accuracy of 0.650 (over baseline = 0.516) using functional images with personal characteristic data when trained on the Autism Brain Imaging Data Exchange (ABIDE) dataset. We also show that it is possible to diagnose ADHD and autism by using just structural brain images with accuracies of 0.661 and 0.601 respectively. Our imaging-based accuracy on the ADHD-200 dataset is about 8% higher than the best imaging-based accuracy in the ADHD-200 competition. While these results are not yet at the level of clinical relevance, they outperform all previously presented methods on both datasets. To our knowledge, this is the first demonstration of a single automated process being able to produce an effective diagnostic system for two different psychiatric illnesses (ADHD and autism). These results suggest that the learning approach using HOG features as input may produce diagnostic classifiers (from functional and/or structural brain images) that perform well for other psychiatric disorders.