Title: Cohort-Scale Neural Atlases of Ultrasound Video

Abstract

Despite being the most prevalent real-time imaging technique in clinical settings, ultrasound faces a significant hurdle in per-frame video annotation. This bottleneck stems from the high cost and scarcity of expert labels, compounded by image variability caused by operator-dependent probe positioning, attenuation, shadowing, and speckle. These challenges are particularly acute because critical clinical data is often dynamic, such as left-ventricular movement in echocardiography or bone and muscle kinematics in musculoskeletal studies. While population atlases can reduce annotation expenses by mapping observations to a common coordinate system, current neural atlas approaches are largely restricted to individual videos, small test-time datasets, or collections focused on specific objects.

To address this, we present a cohort-scale neural atlas designed for ultrasound video. This approach utilizes a unified canonical chart, incorporating per-video Generative Latent Optimization embeddings, and is trained jointly across thousands of frames within the DINOv3 feature space. Evaluated across five datasets involving cardiac and musculoskeletal imaging, which include point landmarks and segmentation masks, our method successfully learns coherent canonical templates and facilitates precise annotation transfer within the atlas space.

On the EchoNet-Dynamic and MSK-Bone datasets, the system enables single- and few-shot transfer with accuracy levels comparable to robust dense-correspondence baselines, all while requiring only minutes of training time on a single consumer-grade GPU. The resulting embeddings are highly interpretable: linear projections uncover structured variations within the cohort, interpolating the image decoder yields anatomically realistic intermediate frames, and latent inversion at test time allows for the reconstruction of held-out frames via the atlas. These findings indicate that cohort-scale neural atlases provide a practical and interpretable framework for alleviating the expert annotation burden in ultrasound video analysis.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC