Title: Echo-POSED: Leveraging Geometric Self-Distillation for Echocardiography Guidance

Abstract:

We present Echo-POSED, a novel self-supervised system designed to provide real-time guidance for transthoracic echocardiography (TTE). This framework suggests necessary probe adjustments directly from 2D ultrasound imagery, eliminating the reliance on expert-labeled views or recorded probe trajectories. The model is trained using 2D slices extracted from standard 3D echocardiography volumes. It enforces equivariance regarding probe movements while maintaining invariance to the cardiac cycle, thereby generating a pose representation within the $\mathrm{SO}(3)\times\mathrm{SO}(3)$ space.

Evaluated on a held-out dataset and public external 3D--TTE collections—which include data with vendor shifts—Echo-POSED demonstrates robust geometric consistency under virtual perturbations. Furthermore, it supports both intra- and inter-patient guidance simulations. In simulations involving cardiac motion, the system achieved a combined mean angular error of 8.2 degrees between the guided and target views.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC