[Study of attention deficit/hyperactivity disorder classification based on convolutional neural networks]

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Developmental Neuroscience

Background:

• Attention deficit/hyperactivity disorder (ADHD) diagnosis relies on subjective methods, leading to misdiagnosis.
• Objective diagnostic tools for ADHD are critically needed, especially in school-aged populations.

Purpose of the Study:

• To develop and validate an objective classification algorithm for ADHD using convolutional neural networks (CNNs).
• To improve diagnostic accuracy and overcome limitations of traditional ADHD assessment methods.

Main Methods:

• Preprocessing of brain MRI scans including skull stripping and smoothing.
• Coarse segmentation to identify key brain regions: right caudate nucleus, left precuneus, and left superior frontal gyrus.
• A three-level CNN model for classifying ADHD and normal groups.

Main Results:

• The CNN algorithm effectively classified ADHD and normal groups.
• Classification accuracy for the right caudate nucleus and left precuneus exceeded the top accuracy from the ADHD-200 competition (62.52%).
• The right caudate nucleus yielded the highest classification accuracy among the tested regions.

Conclusions:

• The proposed coarse segmentation and deep learning method offers a useful, objective approach for ADHD classification.
• This method achieves high accuracy, is computationally simple, and extracts subtle image features effectively.
• It overcomes the time-consuming and complex nature of traditional MRI segmentation techniques, providing a valuable diagnostic tool.