OSA: Echocardiography Video Segmentation via
Orthogonalized State Update and Anatomical Prior-aware Feature Enhancement

Rui Wang¹,  Huisi Wu^1*,  Jing Qin²

¹College of Computer Science and Software Engineering, Shenzhen University
²Centre for Smart Health, School of Nursing, The Hong Kong Polytechnic University

Contents

01

Problem Setting

02

Method

03

Experiments

04

Discussion

2

Problem Setting

Medical

(a)Speckle Noise, (b)Blurred Contours, and (c-f)Pronounced Variations in the Target's Morphology Across the Cardiac Cycle

Video

(a)Extended Temporal Contexts, (b)Efficiency–Accuracy Trade-off in Recall, (c)Computational Burden

Task

Clinical Simulation Setting — Absence of Ground Truth at Inference; Training via Boundary-Frame Prediction and Loss Computation

3

Method

— Overview

Fig.1: Overall pipeline of OSA — Orthogonalized State Update + Anatomical Prior-aware Feature Enhancement

4

Method

— Orthogonalized State Update

$$\mathbf{S}_t^{\text{Euc}} = \mathbf{S}_{t-1}\left(\alpha_t (\mathbf{I}_{C_k} - \beta_t \mathbf{k}_t \mathbf{k}_t^\top)\right) + \beta_t \mathbf{v}_t \mathbf{k}_t^\top.$$

$$\mathcal{V}_{C_v, C_k} = \{\mathbf{S} \in \mathbb{R}^{C_v \times C_k} : \mathbf{S}^\top \mathbf{S} = \mathbf{I}_{C_k}\}.$$

$$\mathbf{S}_t = \operatorname{Proj}_{\mathcal{V}}(\mathbf{S}_t^{\text{Euc}}) = \underset{\mathbf{S} \in \mathcal{V}_{C_v, C_k}}{\arg \min} \frac{1}{2} \|\mathbf{S} - \mathbf{S}_t^{\text{Euc}}\|_F^2.$$

$$\mathbf{X}^{(0)} = \frac{\mathbf{S}_t^{\text{Euc}}}{\|\mathbf{S}_t^{\text{Euc}}\|_F + \epsilon}.$$

5

Method

— Anatomical Prior-aware Feature Enhancement

$$\mathbf{M}_t = \mathrm{AvgPool}_{K \times K}(\mathbf{X}_t)$$

$$\mathbf{X}_t^{+} = \mathrm{ReLU}(\mathbf{X}_t - \mathbf{M}_t), \quad \mathbf{X}_t^{-} = \mathrm{ReLU}(\mathbf{M}_t - \mathbf{X}_t)$$

$$\mathbf{H}_t^{+} = \phi^{+}(\mathbf{X}_t^{+}), \quad \mathbf{H}_t^{-} = \phi^{-}(\mathbf{X}_t^{-})$$

$$\begin{aligned}\lambda_t &= \sigma\!\left( \mathbf{W}_g \left[\mathbf{H}_t^{+};\, \mathbf{H}_t^{-}\right] \right), \\\mathbf{Z}_t &= \lambda_t \odot \mathbf{H}_t^{+} + (1 - \lambda_t) \odot \mathbf{H}_t^{-}\end{aligned}$$

6

Experiments

01

OSA achieves state-of-the-art Dice on CAMUS & EchoNet-Dynamic.

02

OSU stabilizes hidden-state trajectories on Stiefel manifold.

03

APFE enhances feature discrimination via anatomical priors.

04

Lightweight design: only +3.2% params vs. baseline.

7

Experiments

8

Discussion

•

Generality: Extend to broader ultrasound datasets.

•

Harder video tasks: Tackle longer sequences, complex rhythms, and difficult cases.

•

Hardware-aware: Optimize the matrix state for parallel acceleration.

9

Thanks!

Paper: arXiv:2501.0xxxx

Code: github.com/wangrui2025/osa

10