VARS-fUSI

Functional ultrasound imaging (fUSI) infers neural activity from cerebral blood volume changes, with higher sensitivity and spatial resolution than fMRI and without the intracortical electrodes that electrophysiology requires. Its bottleneck is data volume: a single power Doppler (DOP) image takes hundreds of ultrasound pulse emissions, beamformed into in-phase and quadrature (IQ) frames, filtered by singular value decomposition to remove tissue motion, then squared and averaged. That pipeline caps the imaging framerate, consumes storage and bandwidth, and can heat the probe past FDA safety guidance during long sessions.

VARS-fUSI (VARiable Sampling fUSI) reconstructs the same DOP images from 10-15% of those frames — 32 instead of 250 or 300 — and is the first fUSI method built on neural operators, models that learn mappings between function spaces rather than between fixed-size arrays. That choice is what makes the "variable sampling" in the name work: because the temporal axis is treated as a continuous function, a single trained model runs at IQ sampling rates and acquisition durations it never saw during training, without retraining. It was developed at Caltech by Bahareh Tolooshams and colleagues with the Andersen and Shapiro labs, NVIDIA, and clinical collaborators at USC and Rancho Los Amigos, and posted as a preprint in April 2025.

#Key Features

• Discretization-agnostic temporal filtering: Tissue-clutter removal is performed globally in the Fourier domain by a Fourier neural operator, so the model accepts a different frame count or sampling rate at inference than it was trained on — something fixed-framerate convolutional predecessors such as Deep-fUS cannot do.
• Two independent acceleration modes: Reduced-time takes the first K frames of each acquisition, raising the effective DOP framerate roughly eightfold in monkey recordings and tenfold in mouse and human recordings; reduced-sampling keeps the full acquisition window but drops the IQ sampling rate by up to 90%.
• One base model, few-shot adaptation across species: A model trained only on mouse data generalised to unseen sessions from held-out subjects without fine-tuning, and adapted to monkeys and humans from 2 to 120 example images in minutes on a single GPU with under 24 GB of memory.
• Complex-valued input with spatial guidance: The network preserves the complex structure of the IQ signal instead of discarding phase, and is steered toward blood-flow regions by an SVD-based spatial guidance module.

#Technical Details

The architecture is a U-shaped encoder-decoder with five levels and channel widths of 16, 32, 64, 128 and 256, ReLU nonlinearities, residual connections and 0.2 dropout. Each layer decomposes processing into spatial filtering by convolution and global temporal filtering by a Fourier neural operator retaining nine spectral modes. Training used the MultiMice dataset — fUSI acquired through intact skull and skin from five mouse brains — for 600 epochs with Adam at a learning rate of 1e-4, weight decay 1e-5 and gradient clipping at 5.0, with SVD-processed full-frame DOP images as ground truth. Fine-tuning warm-starts from those weights: 2 images and 1,600 iterations for a new mouse, 120 images and 3,600 iterations for the macaque sessions, 32 images and 5,120 iterations for the human sessions.

On the MultiMice test set VARS-fUSI beat SVD-fUSI, DnCNN, Deep-fUS-base and Deep-fUS on NMSE, PSNR and SSIM at K = 32 frames. In macaque lateral intraparietal cortex and in a human participant with an acoustic cranial window, saccade-direction and movement-effector decoding from reconstructed images tracked the full-frame ground truth closely and exceeded Deep-fUS. End-to-end, one DOP image takes 0.253 s versus 2.295 s for the standard pipeline: collection wait time falls from 2.2 s to 0.23 s and processing from 0.095 s to 0.023 s on an RTX 4090.

#Applications

The clearest beneficiary is fUSI-based brain-computer interfaces, where a higher DOP framerate directly reduces decoder latency and the sampling reduction eases the memory and emission budgets that currently block large fields of view and volumetric imaging. The same efficiency argument applies to bedside neonatal monitoring, intraoperative functional mapping, and freely-moving animal recordings, all settings where probe heating and data throughput constrain session length.

#Impact

VARS-fUSI is a preprint and has not yet been peer reviewed. The code repository is MIT licensed and provides the model implementation, training, fine-tuning and prediction scripts and the experiment configurations; it does not host checkpoint files, so reproducing the base model currently requires training on the authors' data. The authors are candid about the ceiling: VARS-fUSI does not improve image quality or decodability over the full-frame SVD standard, only the cost of getting there, and they caution against using its reconstructions directly for functional brain mapping because of noise, signal dampening and the possibility of hallucinated activity at the most aggressive sampling reductions. As an application of operator learning to a live neurotechnology bottleneck, though, it is a clean demonstration that discretization-invariance is worth architectural effort in biomedical imaging.