High-Fidelity Functional Ultrasound Reconstruction via A Visual Auto-Regressive Framework

TL;DR

This work tackles data scarcity and fairness in functional ultrasound imaging by introducing UltraVAR, a visual auto-regressive augmentation framework built on a VQVAE backbone. It adds two novel components, the Smooth Scaling Layer and Perception Enhancement Module, to enable high-fidelity, multi-scale reconstruction that preserves neurovascular coupling. Empirical results show UltraVAR outperforms competing generative methods on image fidelity and downstream classification tasks, while also improving fairness on imbalanced data. The approach has potential to accelerate clinical translation of fUS, including fetal ultrasound and neonatal brain monitoring, by providing physiologically plausible synthetic data for robust model training and analysis.

Abstract

Functional ultrasound (fUS) imaging provides exceptional spatiotemporal resolution for neurovascular mapping, yet its practical application is significantly hampered by critical challenges. Foremost among these are data scarcity, arising from ethical considerations and signal degradation through the cranium, which collectively limit dataset diversity and compromise the fairness of downstream machine learning models.

Figures (7)

• Figure 1: Overview of the proposed UltraVAR framework. The architecture consists of a VQVAE backbone with encoder and decoder components, integrated with two modules: the Smooth Scaling Layer (SCL) and Perception Enhancement Module (PEM). The hierarchical reconstruction process refines image patches at progressive resolutions.
• Figure 2: Visual comparison of fUS images. This figure displays original fUS images alongside synthetic samples generated by UltraVAR and various baseline methods. Comparisons are shown for different brain activity classes (Class0 and Class1) across two experimental sessions (S1 stands for Session1 and S2 stands for Session2).
• Figure 3: Quantitative comparison of image quality metrics across different generative models. The charts display SSIM, MS-SSIM, and FID values for Session1 (top row) and Session2 (bottom row). UltraVAR consistently outperforms other methods with higher SSIM and MS-SSIM values and lower FID scores.
• Figure 4: PCA visualization comparing feature distributions of original fUS data, DiT-generated data, and UltraVAR-generated data. The closer clustering of UltraVAR samples to the original data, relative to DiT samples, indicates that UltraVAR generates data with representations more faithful to the original dataset.
• Figure 5: Performance comparison on the downstream classification task. The radar charts illustrate classification metrics for different generative models on Session1 (left) and Session2 (right), with the proposed UltraVAR model showing superior performance.