A 5.0 T Ultra-High-Field fMRI Dataset for Naturalistic Visual Scene Processing

Area of Science:

• Computational Neuroscience
• Neuroimaging
• Human Visual System

Background:

• Modeling neural responses to naturalistic visual stimuli is crucial for advancing computational neuroscience and brain-computer interfaces.
• Existing neuroimaging datasets often lack the necessary combination of repeated measurements, shared stimuli, and diverse content for robust model evaluation.

Purpose of the Study:

• To introduce the Natural Vision Dataset (NVD), a novel 5.0 Tesla functional Magnetic Resonance Imaging (fMRI) dataset.
• To provide a standardized resource for investigating human visual representations and evaluating computational models of natural image processing.

Main Methods:

• Acquired 5.0T fMRI data from 20 participants viewing 1,268 shared and 10,000 participant-specific natural images, each presented three times.
• Utilized a hybrid shared-unique stimulus design for assessing response reliability, cross-participant alignment, and model generalization.
• Organized data using Brain Imaging Data Structure (BIDS) with volumetric and surface-based derivatives, and derived image-level responses using GLMsingle.

Main Results:

• Characterized dataset utility through analyses of motion, vigilance, temporal signal-to-noise ratio (tSNR), noise-ceiling, and brain-to-CLIP decoding.
• Demonstrated the dataset's suitability for evaluating encoding and decoding models of visual processing.

Conclusions:

• The NVD is a valuable, publicly available resource for the neuroscience community.
• Facilitates research into human visual representations and the development of advanced computational models for natural image understanding.