Figure 1. Overview of the proposed OSA framework. Top: unconstrained linear recurrent models suffer from rank collapse under severe speckle noise. Middle: our OSU mechanism constrains state evolution on the Stiefel manifold, preventing singular value decay. Bottom: the APFE module decouples anatomical structures from speckle noise via physics-driven decomposition, providing noise-resilient cues for temporal tracking.
Abstract
Accurate and temporally consistent segmentation of the left ventricle from echocardiography videos is essential for estimating the ejection fraction and assessing cardiac function. However, modeling spatiotemporal dynamics remains difficult due to severe speckle noise and rapid non-rigid deformations.
Existing linear recurrent models offer efficient in-context associative recall for temporal tracking, but rely on unconstrained state updates, which cause progressive singular value decay in the state matrix—a phenomenon known as rank collapse—resulting in anatomical details being overwhelmed by noise.
To address this, we propose OSA, a framework that constrains the state evolution on the Stiefel manifold. We introduce the Orthogonalized State Update (OSU) mechanism, which formulates memory evolution as Euclidean projected gradient descent on the Stiefel manifold to prevent rank collapse and maintain stable temporal transitions.
Furthermore, an Anatomical Prior-aware Feature Enhancement module explicitly separates anatomical structures from speckle noise through a physics-driven process, providing the temporal tracker with noise-resilient structural cues. Comprehensive experiments on the CAMUS and EchoNet-Dynamic datasets show that OSA achieves state-of-the-art segmentation accuracy and temporal stability, while maintaining real-time inference efficiency for clinical deployment.
Motivation
Echocardiography video segmentation faces severe challenges from speckle noise, low contrast, rapid non-rigid cardiac deformation, and sparse clinical annotations. Existing methods struggle to jointly satisfy accuracy, temporal consistency, and real-time efficiency.
Figure 2. Challenges in echocardiography video segmentation. (a) Severe speckle noise obscures anatomical boundaries. (b) Low contrast between myocardium and blood pool. (c) Rapid non-rigid deformation between systole and diastole. (d) Depth-dependent acoustic attenuation. (e) Sparse annotations (ED/ES only). (f) Real-time clinical requirements.
Method
Orthogonalized State Update (OSU)
OSU constrains the memory evolution to the Stiefel manifold via an orthogonalized update. By enforcing strict orthogonality, the Frobenius norm of the state remains constant, preventing singular value decay and rank collapse. A parameterized high-order Newton-Schulz iteration achieves efficient projection without explicit SVD, reducing computational cost from O(Cv Ck^2) to a fixed number of matrix multiplications.
Figure 3. Architecture of the Orthogonalized State Update (OSU) mechanism. Left: standard Euclidean recurrence allows the state to deviate from the manifold, causing rank collapse. Right: OSU explicitly projects the unconstrained state back onto the Stiefel manifold, preserving orthogonality and preventing singular value decay.
Anatomical Prior-aware Feature Enhancement (APFE)
APFE reformulates spatial feature extraction as a physics-driven decoupling process. It decomposes the feature map into a low-frequency acoustic bias field and high-frequency structural residuals via large-kernel average pooling. Dual Conv-BN-ReLU pathways separately handle structural geometry and region-level semantics, fused via an adaptive gating mechanism. This provides noise-resilient structural anchors for the temporal tracker.
Figure 4. Architecture of the Anatomical Prior-aware Feature Enhancement (APFE) module. A large-kernel average pooling decomposes the feature map into a low-frequency acoustic bias field and high-frequency structural residuals. Dual Conv-BN-ReLU pathways separately process structural geometry and region-level semantics, fused via an adaptive gating mechanism to provide noise-resilient structural anchors for temporal tracking.
Results
OSA achieves state-of-the-art performance on CAMUS and EchoNet-Dynamic datasets, with 94.82 mDice and 35 fps real-time inference.
Table 1. Quantitative comparison with state-of-the-art methods.
Table 2. Ablation study results.
Figure 5. Qualitative segmentation results on CAMUS and EchoNet-Dynamic. OSA consistently yields more accurate contours across varying imaging qualities, acoustic shadowing, and ambiguous boundaries.
BibTeX
@InProceedings{Wang_2026_CVPR,
author = {Wang, Rui and Wu, Huisi and Qin, Jing},
title = {{OSA}: Echocardiography Video Segmentation via Orthogonalized State Update and Anatomical Prior-aware Feature Enhancement},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2026},
pages = {1428-1438}
}