Classification of schizophrenia and normal controls using 3D convolutional neural network and outcome visualization.

Area of Science:

• Neuroimaging
• Machine Learning
• Psychiatry

Background:

• Deep learning for schizophrenia (SCZ) classification often uses manual feature extraction, losing crucial 3D MRI data structure.
• Preserving 3D spatial information is vital for accurate SCZ identification and biomarker discovery.

Purpose of the Study:

• To evaluate a novel 3D convolutional neural network (CNN) model for classifying schizophrenia spectrum disorders (SSDs).
• To compare the proposed 3D CNN model's accuracy against Support Vector Machine (SVM) and other 3D CNNs.
• To identify key brain regions for classification using a class saliency visualization (CSV) method.

Main Methods:

• Task-based functional MRI (fMRI) data from 103 SSD patients and 41 healthy controls were utilized.
• A 3D convolutional autoencoder (3D-CAE) was used for unsupervised pretraining of the CNN model.
• The 3D-CAE-based CNN model was trained and tested, with performance compared against SVM and other 3D CNNs.

Main Results:

• The proposed 3D-CAE-based CNN achieved classification accuracies of 84.15%–84.43% using fMRI data.
• The model demonstrated superior performance compared to SVM and other 3D CNN models.
• The inferior and middle temporal lobes were identified as critical regions for distinguishing SSDs from controls.

Conclusions:

• The 3D-CAE-based CNN model effectively classifies SSD patients with high accuracy, surpassing existing methods.
• Saliency map visualization offers valuable clinical insights by highlighting discriminative brain regions.