Deep Learning for Mental Illness Detection Using Brain SPECT Imaging

Abstract

We apply deep learning to the detection of mental illness with meaningful results, using SPECT (Single Photon Emission Computed Tomography) images of the brain. The data consists in scans from patients with attention deficit hyperactivity disorder (ADHD), major depressive disorder (MDD) and obsessive compulsive disorder (OCD), plus scans of healthy brains. We focus here on the application of a deep convolutional neural network (CNN). The main challenge in using CNN models for medical diagnosis is often the number of samples not being sufficiently large to ensure high accuracy. We propose a soft classifier for using the machine. Instead of a binary output “yes/no” for each condition, we add an intermediate outcome, which says that the machine yields a weak result. The “Red Zone” corresponds to a positive result (condition is present) and the “Green Zone” corresponds to a negative, each with a preassigned statistical confidence level; the “Amber Zone” is an ambiguous outcome, where the scan is assigned a likelihood of having the condition. This information is then passed to the doctors for further analysis of patients.

Daniel Amen and Thomas Ward are co-last authors.