High-Fidelity Functional Ultrasound Reconstruction via a Visual Auto-Regressive Framework.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Engineering

Background:

• Functional ultrasound (fUS) imaging offers high spatiotemporal resolution for neurovascular mapping.
• Data scarcity and limited diversity in fUS datasets hinder machine learning model fairness and performance.
• Ethical constraints and signal attenuation restrict the creation of comprehensive fUS datasets.

Purpose of the Study:

• Introduce UltraVAR, a novel data augmentation framework for fUS imaging.
• Address data scarcity and enhance fairness in machine learning models for fUS data.
• Generate diverse, physiologically plausible fUS samples that preserve neurovascular coupling features.

Main Methods:

• Developed UltraVAR, a data augmentation framework utilizing a pre-trained visual auto-regressive generative model.
• Implemented a scale-by-scale reconstruction mechanism to preserve vascular network topology.
• Integrated Smooth Scaling Layer and Perception Enhancement Module to maintain image fidelity and reduce artifacts.

Main Results:

• UltraVAR successfully generates diverse and physiologically plausible fUS samples.
• Augmented datasets using UltraVAR demonstrated statistically significant improvements in downstream classification accuracy.
• The framework preserves crucial neurovascular coupling features, unlike conventional augmentation methods.

Conclusions:

• UltraVAR effectively mitigates data scarcity and enhances fairness in fUS-based machine learning.
• The framework reconstructs high-fidelity, diverse fUS data, preserving essential physiological correlations.
• This work supports advancements in ultrasound-based neuromodulation and brain-computer interfaces.