Three dimensional convolutional neural network-based classification of conduct disorder with structural MRI.

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Conduct disorder (CD) is a prevalent child and adolescent psychiatric condition with significant long-term societal and patient burdens.
• Neuroimaging combined with deep learning offers potential for identifying biomarkers in psychiatric disorders.

Purpose of the Study:

• To apply an optimized 3D AlexNet convolutional neural network (CNN) model for the classification of CD from healthy controls (HCs) using structural magnetic resonance imaging (sMRI).
• To automatically extract multi-layer, high-dimensional features from sMRI data for biomarker discovery in CD.

Main Methods:

• Acquired high-resolution sMRI data from 60 male adolescents with CD and 60 age- and gender-matched HCs.
• Utilized a 5-fold cross-validation strategy to train and test an optimized 3D AlexNet CNN model.
• Compared the performance of the AlexNet model against a support vector machine (SVM) using receiver operating characteristic (ROC) curves and feature visualization (saliency maps).

Main Results:

• The AlexNet model achieved high classification performance with an accuracy of 0.85, specificity of 0.82, and sensitivity of 0.87.
• The area under the ROC curve (AUC) for AlexNet (0.86) was significantly higher than that for SVM (0.78, p=0.046).
• Saliency maps highlighted key brain regions, including the frontal lobe, superior temporal gyrus, parietal lobe, and occipital lobe, as discriminative features for CD.

Conclusions:

• The deep learning-based 3D AlexNet CNN model effectively extracts hidden features from sMRI, demonstrating its capability in classifying CD.
• This approach shows potential as an assistive tool for clinicians in the diagnosis of conduct disorder.