Voxelwise Visual Modeling of Natural Scenes BOLD fMRI responses using weighted-neighbor based regressor
Subhrasankar Chatterjee, Shreyansh Verma, Debasis Samanta, Monalisa Sarma · 2024
Deciphering the neural substrates underlying visual processing is fundamental in medical imaging, with computational models playing a pivotal role. Voxelwise encoding models are essential in unraveling the complex relationship between visual stimuli and neural responses across cortical regions. However, while feature extraction methodologies have advanced with deep learning, translating features into neural responses through regression remains challenging. Traditional regression methods struggle with the escalating dimensionality of feature spaces, necessitating novel approaches. This paper introduces a novel weighted neighbor-based regression framework for voxelwise visual encoding models to improve predictive accuracy and interpretability. Through systematic experiments, we evaluate the framework’s performance across varying feature dimensions, support neighbor values, and regression methodologies. Leveraging two standard deep-learning models, ALexNet and ResNet18, we extract features representing early and late visual areas and systematically vary the number of support neighbors (n). Results demonstrate the framework’s superiority over traditional linear regression approaches in accurately predicting voxel-wise neural responses. Furthermore, we identify optimal configurations that maximize predictive accuracy while minimizing overfitting or underfitting. Our findings underscore the framework’s potential as a valuable tool in voxelwise visual encoding, shedding light on the complex neural mechanisms underlying visual processing. The proposed framework offers a nuanced and interpretable representation of feature-to-response mapping, enhancing our understanding of visual perception in the human brain. Future research directions include further refinement of the framework, exploration of alternative regression methodologies, and extension to incorporate additional neuroimaging modalities. Ultimately, advancements in voxelwise visual encoding will deepen our understanding of visual cognition and perception.