Echo-POSED: Geometric Self-Distillation for Echocardiography Guidance

2026-05-30

AI Summary

This study addresses the lack of real-time probe guidance in transthoracic echocardiography (TTE) by proposing a self-supervised framework that operates without expert annotations. Leveraging routinely acquired 3D ultrasound data, the method employs geometric self-distillation to learn probe pose representations from 2D slices. It is the first to jointly model, in a label-free setting, the equivariance of probe motion and the invariance of the cardiac cycle, thereby constructing a pose space on the SO(3)×SO(3) manifold. By integrating 3D-to-2D reconstruction with virtual perturbation consistency constraints, the model enables cross-patient and cross-device probe adjustment recommendations. Evaluated on a held-out test set and an external 3D-TTE dataset, the approach achieves a mean angular error of 8.2 degrees, demonstrating strong geometric consistency and generalization capability.

Abstract

We introduce Echo-POSED, a self-supervised framework for real-time transthoracic echocardiography (TTE) guidance that recommends probe adjustments directly from 2D ultrasound images, without the need for expert-labelled views or tracked probe trajectories. Instead, it trains on 2D views sliced from routinely acquired 3D echocardiography volumes, enforcing equivariance to probe motions while remaining invariant to cardiac phase, yielding a pose representation on $\mathrm{SO}(3)\times\mathrm{SO}(3)$. Across a held-out split and public external 3D--TTE datasets (including vendor shift), Echo-POSED maintains geometric consistency under virtual perturbations and enables intra- and inter-patient guidance simulations, achieving a combined mean angular error of 8.2 degrees between the guided and target views in intra-patient simulations with cardiac motion.

Problem

• echocardiography guidance
• self-supervised learning
• probe pose estimation
• real-time ultrasound
• geometric consistency

Innovation

• self-supervised learning
• geometric equivariance
• echocardiography guidance
• pose estimation
• SO(3) representation