Automatic Classification of Signal and Noise in Functional Magnetic Resonance Imaging Scans Using Convolutional Neural Networks

The integration of Artificial Intelligence (AI), particularly deep learning models like VGG16 and ResNet50, in the analysis of functional magnetic resonance imaging (fMRI) data has significantly advanced our understanding of brain functionality and the diagnosis of neurological disorders. This paper explores the application of Convolutional Neural Networks (CNNs) to enhance the accuracy and efficiency of fMRI data analysis, addressing challenges such as high dimensionality, noise, and the need for complex data preprocessing. Our study evaluates the performance of VGG16 and ResNet50 models and one 3D CNN on 2D and 3D fMRI datasets, highlighting the limitations of VGG16 in handling 2D data and demonstrating the superior performance of ResNet50 on balanced and unbalanced datasets. Additionally, we investigate the impact of using 3D data from the Human Connectome Project (HCP), achieving up to 98% accuracy on the validation set. The results indicate that CNNs can effectively replace traditional Independent Component Analysis (ICA) methods by leveraging their capability for automatic feature extraction and end-to-end learning.