Task-Based Adaptive Transmit Beamforming for Efficient Ultrasound Quantification

TL;DR

The paper tackles power and data bottlenecks in wireless and wearable ultrasound by introducing Task-Based Information Gain (TBIG), a Bayesian active perception framework that adaptively steers transmit beamforming to minimize uncertainty in a differentiable downstream task $D_t$. It formulates a perception–action loop where perception uses diffusion-based posterior sampling to generate image realizations, and action minimizes the expected downstream uncertainty via a Jacobian-based saliency map and $K$-Greedy Minimization, enabling task-focused subsampling. The method generalizes to any differentiable downstream task and demonstrates superior performance to the GIG baseline on EchoNet-LVH–derived left-ventricular measurements, achieving accurate results with roughly 2% of the typical scan lines. This approach promises substantial power and data-rate reductions for continuous monitoring with wearable ultrasound, with potential impact on battery life and throughput in practical devices.

Abstract

Wireless and wearable ultrasound devices promise to enable continuous ultrasound monitoring, but power consumption and data throughput remain critical challenges. Reducing the number of transmit events per second directly impacts both. We propose a task-based adaptive transmit beamforming method, formulated as a Bayesian active perception problem, that adaptively chooses where to scan in order to gain information about downstream quantitative measurements, avoiding redundant transmit events. Our proposed Task-Based Information Gain (TBIG) strategy applies to any differentiable downstream task function. When applied to recovering ventricular dimensions from echocardiograms, TBIG recovers accurate results using fewer than 2% of scan lines typically used, showing potential for large reductions in the power usage and data rates necessary for monitoring. Code is available at https://github.com/tue-bmd/task-based-ulsa.

Figures (3)

• Figure 1: Diagram illustrating single iteration of the task-based perception-action loop using EchoNetLVH segmentation for the downstream task. 1 Generate a set of posterior samples from the sparse acquisition using DPS. 2 Pass each posterior sample $\mathbf{x}_t^{(i)}$ through the downstream task model $f$ to produce samples from the downstream task distribution. 3 Compute the Jacobian matrix using each of the posterior samples as inputs. 4 Average those Jacobian matrices and multiply them with the pixel-wise variance of the input images to produce the downstream task saliency map. 5 Apply $K$-Greedy Minimization to select $K$ scan line locations for the next acquisition.
• Figure 2: (a) shows a qualitative comparison of measurement signal recovery for three patients using (i) TBIG and (ii) GIG sampling strategies with 5/256 scan lines. The MAE between the target and reconstruction is provided at the top left. The uncertainty for each reconstruction is quantified as the standard deviation of the measurement values estimated from the samples in the belief set computed by applying $f$ to $\{\mathbf{x}_t^{(i)}\}_{i=0}^{N_p}$ for both methods. (b) shows a sample-wise error comparison for each patient in the evaluation set, where it is clear that TBIG almost always outperforms GIG.
• Figure 3: The distribution of MAE scores between target and reconstructed LVID time series for the first 100 frames from 50 patients from the EchoNetLVH validation set for both TBIG and GIG strategies.