Abstract
Echocardiographic segmentation remains challenging due to speckle noise, low contrast boundaries, and substantial domain shift across devices and acquisition views. While foundation models such as SAM2 offer stable generalization, they exhibit systematic failure in zero-shot echocardiographic segmentation, which we term prompt-invariant collapse. We attribute this to the difficulty of finding consistent spatiotemporal representations across diverse data distributions in the pixel domain.
We observe that frequency-domain cardiac signals, driven by periodic systolic-diastolic motion, maintain stable and device-invariant structure across patients and datasets where pixel-domain statistics vary wildly. Motivated by this, we propose EchoFAM, a lightweight framework that adapts a fully frozen SAM2 backbone via frequency-domain physical priors. Two progressive modules are introduced: a Frequency Adapter (FA) for spatially differentiated residual correction of encoder embeddings, and a Memory-Aware Temporal Gate (MATG) for phase-aware temporal regulation via historical context fusion. Experiments demonstrate that EchoFAM achieves state-of-the-art performance on in-distribution benchmarks while exhibiting notably robust generalization on zero-shot segmentation across unseen data distributions, validating frequency-domain cardiac motion as an effective and efficient prior for echocardiographic segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_005.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=eE0FcfLvpk
BibTex
@InProceedings{LiQin_EchoFAM_MICCAISAT2026,
author = { Li, Qingya AND Wu, Miao AND Zhang, Yu AND Zou, Chuanqi AND Li, Lei},
title = { { EchoFAM: Frequency Prior-Driven Adaptation of SAM2 for Echocardiography Video Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}