ADHD Subtype Diagnosis Through Convolutional Neural Networks Evaluation of the Connectivity Networks in Brain fMRI

Attention-deficit hyperactivity disorder (ADHD) affects at least 5% of the world population. ADHD typically starts in childhood and can affect normal development that can cause serious consequences during adulthood. Moreover, there are three subtypes of ADHD, each with its symptomatic differences that require particular treatment. The development of tools capable of detecting the brain differences between each subtype and a control group can help in directing future neuropsychological studies and guide treatments. Here we investigate the hypothesis of using Convolutional Neural Networks (CNN) as a tool to detect ADHD in connectivity networks of the brain. This research used the Yeo brain parcellation maps from 28 healthy volunteers, 36 combined subtype ADHD, and 26 inattentive type ADHD. The CNN achieved a 64.71% accuracy at detecting the multi-class diagnosis (Control, ADHD combined, or ADHD inattentive) when using the complete connectivity network matrix, surpassing the individual accuracies achieved when using only the within network connectivity measures. However, the separate analysis of each network showed that the Dorsal, Default and Frontoparietal networks have better accuracy of performing the classification compared to using other within brain connectivity measures. The Default Network has shown greater accuracy at identifying ADHD-I than the other within networks or whole brain connectivity measures. These results suggest that the neurological differences of the presence of ADHD and its subtypes may be present and, therefore, affect more than one connectivity network functionality.