Trevor Chan — Ultrasound Guidance

Overview

Point-of-care transthoracic echocardiography (TTE) is a low-cost, non-invasive tool for diagnosing common cardiac diseases, but acquiring and interpreting TTE images requires specialized training, limiting patient access. While extensive prior work focuses on TTE interpretation, the equally critical problem of image acquisition remains largely unaddressed. Here, we introduce a novel method for flow-based ultrasound guidance that enables even untrained operators to acquire diagnostically valuable TTE imaging studies. Specifically, we formulate the task of ultrasound transducer placement as a distribution mapping problem over the space of $SE(3)$ poses. Using flow matching, we train a network to learn a velocity field transporting from the distribution of arbitrary poses to that of canonical view poses. Sampling from this field constitutes a closed-loop process: conditioned on the current image and desired view class, the network predicts a velocity in $se(3)$, which is conveyed to the operator as real-time probe adjustment instructions via a graphical interface. As in generative image models, we find in simulation that classifier-free guidance yields improvements in convergence and pose alignment. Using this approach, we collected 1 hour of training data, consisting of ultrasound images and tracked transducer poses, on a single subject and evaluated performance on 3 unseen subjects. In qualitative and quantitative comparisons with 5 novice and intermediate operators, we found that the addition of model-based guidance significantly decreased acquisition time for the 6 most common TTE views. These findings indicate that flow-based guidance models may substantially close the performance gap between novice and trained sonographers, expanding access to point-of-care cardiac imaging in resource-limited settings.

TTE image acquisition can be viewed as a distribution transport problem, where samples from the distribution of random probe placements are transformed into optimal probe placements corresponding to specific canonical views.

AP4C (top) and PLAX (bottom) views acquired with model guidance. At the start of sampling, the transducer is placed in a neutral position on the chest wall. The operator then attempts to align the reticle with the target to acquire the selected view.

In prospective <i>in vivo</i> evaluations, novice and intermediate operators required significantly less time to acquire images of comparable diagnostic quality.

In prospective in vivo evaluations, novice and intermediate operators required significantly less time to acquire images of comparable diagnostic quality.

Annealed diffusion autoguidance accelerates convergence to the target view by straightening sampled trajectories. A weighting of 1.0 is identical to no guidance, and higher weightings more heavily emphasize the strong model prediction.