Abstract
Standard-view acquisition is the most operator-dependent step of transthoracic echocardiography, and automating it is an attractive target for human–AI collaboration. Recent methods pose it as supervised regression: from the current frame a network predicts the probe motion to an expert-annotated target pose and is trained on the error to that pose. Yet the goal of acquisition is not to reproduce a particular hand pose but to obtain a specific standard-view image; pose error is only a proxy for this, and a network trained on recorded trajectories cannot act from probe poses outside its data. We therefore formulate acquisition as reinforcement learning (RL) and reward the agent by the similarity between the current and target views, computed in the patchtoken feature space of a frozen, general-purpose vision foundation model (FM). We build a patient-specific ultrasound simulator from public chest CT, in which a virtual probe on the skin surface renders B-mode frames. On the parasternal long-axis (PLAX) view, the FM (DINOv2) similarity decreases smoothly and monotonically with probe offset, providing a dense, label-free reward. A single soft actor–critic policy trained with it acquires PLAX on five unseen patients with a 12.4 mm mean positional error and 81.2% clinical-tolerance success, and at times reaches valid views from poses that differ from the expert annotation, showing that a foundation-model similarity reward captures an acquisition goal that pose error alone does not.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/HAIC26_018.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=WLsaTlTppV
BibTex
@InProceedings{JeoHyu_Toward_MICCAISAT2026,
author = { Jeong, Hyunseok AND Yoon, Jong-um AND Lee, Seohwa AND Jeon, Byunghwan},
title = { { Toward Agentic Echocardiography: Reinforcement Learning with Foundation-Model Similarity Rewards in a CT-Derived Simulation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17279},
month = {pending},
page = {pending}
}