fMRI volume classification using a 3D convolutional neural network robust to shifted and scaled neuronal activations

Area of Science:

• Neuroimaging and Machine Learning
• Cognitive Neuroscience
• Artificial Intelligence in Healthcare

Background:

• Deep learning, particularly deep neural networks (DNNs), shows promise in analyzing neuroimaging data.
• Convolutional neural networks (CNNs) are effective DNNs for feature extraction from spatial data.
• Functional MRI (fMRI) data analysis faces challenges due to spatial misalignment and BOLD response variability.

Purpose of the Study:

• To evaluate a 3D CNN model for classifying task information from single functional MRI (fMRI) volumes.
• To assess the 3D CNN's robustness against variations in spatial alignment and preprocessing.
• To compare the 3D CNN's performance against 1D fully connected DNN (1D-fcDNN) and support vector machine (SVM) classifiers.

Main Methods:

• A 3D CNN model was developed to extract features from fMRI volumes for task classification.
• fMRI data from four sensorimotor tasks were used, with evaluations on minimally and fully preprocessed data.
• Classification performance was compared using various cross-validation schemes, including leave-one-subject-out (LOOCV).

Main Results:

• The 3D CNN model significantly outperformed 1D-fcDNN and SVM models in classifying task information.
• With LOOCV and full preprocessing, 3D CNN achieved a mean error rate of 2.1%, compared to 3.1% for 1D-fcDNN and 4.1% for SVM (p=0.041).
• Performance degraded substantially with raw, un-preprocessed fMRI data, highlighting the importance of preprocessing.

Conclusions:

• 3D CNN models are highly effective for decoding task-related information from single fMRI volumes.
• The 3D CNN's architecture inherently handles spatial variations, offering robustness in neuroimaging analysis.
• This approach holds potential for real-time fMRI analysis and advancing brain-computer interfaces.