A Multichannel 2D Convolutional Neural Network Model for Task-Evoked fMRI Data Classification.

Area of Science:

• Neuroimaging
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models, particularly convolutional neural networks (CNNs), are effective for analyzing functional MRI (fMRI) data.
• CNNs excel at feature extraction due to shared-weights and spatial invariance.
• Classifying complex 3D fMRI data remains a challenge, requiring efficient and accurate models.

Purpose of the Study:

• To propose and evaluate M2D CNN, a novel multichannel 2D CNN model for 3D fMRI data classification.
• To compare the performance of M2D CNN against established machine learning and deep learning models.
• To assess the model's efficacy in classifying task-based fMRI data from the Human Connectome Project.

Main Methods:

• Developed a multichannel 2D CNN (M2D CNN) model that processes sliced 2D fMRI data.
• Integrated multichannel information within the 2D CNN framework.
• Benchmarked M2D CNN against Support Vector Machine (SVM), 1D CNN, 2D CNN, 3D CNN, and 3D separable CNN using HCP motor task data.

Main Results:

• All CNN models outperformed SVM in classifying high-dimensional whole-brain imaging data.
• 3D CNN models showed higher accuracy than 1D and 2D CNNs but incurred higher computational costs.
• M2D CNN achieved the highest classification accuracy and demonstrated reduced data overfitting due to fewer parameters than 3D CNN.

Conclusions:

• Convolutional operations are beneficial for fMRI data classification.
• M2D CNN offers a computationally efficient and highly accurate approach for 3D fMRI data analysis.
• The proposed M2D CNN model presents a promising advancement in neuroimaging analysis and classification.