EN
Shenzhen University logo Hong Kong Polytechnic University logo

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

Rui Wang1    Huisi Wu1*    Jing Qin2
1Shenzhen University
2The Hong Kong Polytechnic University
CVPR 2026 conference logo

Problem Setting

LV segmentation from echocardiography is essential for cardiac function assessment, yet challenging due to:
  • Severe speckle noise & low contrast
  • Rapid non-rigid deformation (systole ↔ diastole)
  • Sparse annotations (ED/ES frames only)
Challenges
Figure 1. Challenges: speckle noise, blurred contours, and shape variations.

Motivation

Existing video segmentation models ignore ultrasound-specific noise statistics, while temporal models suffer from state degradation. OSA addresses both jointly:
  • Speckle resilience via physics-driven anatomical decoupling
  • Temporal stability via orthogonalized state updates
  • Sparse annotation handled by ED/ES frame supervision

Contributions

  • A novel unified temporal + ultrasound-aware architecture
  • Stiefel constraint prevents rank collapse
  • Physics-driven decoupling for noise resilience
  • Stiefel manifold: St ∈ VCv, Ck with S⊤S = I
  • Newton-Schulz projection: avoids exact SVD, O(1) fixed steps

Our Approach

We propose OSA — joint temporal stabilization + anatomy-aware enhancement:
  • OSU: Stiefel manifold constraint prevents rank collapse
  • APFE: Physics-driven decoupling separates anatomy from speckle
OSA Teaser
Figure 2. Paradigm comparison: Memory Bank, LRM, and OSA.

Method Overview

OSA segments LV using a ResNet-50 backbone.
OSA Architecture
Figure 3. OSA architecture: APFE → OSU → decoder.
Pipeline:
  1. ResNet-50 extracts features Ft
  2. APFE decomposes into keys/values
  3. OSU maintains state St on Stiefel manifold
  4. Query Qt ⊗ St
  5. Decoder outputs mask Mt

Orthogonalized State Update

Linear recurrent models suffer rank collapse: unconstrained updates decay singular values in St.
𝒱Cv, Ck = { S ∈ ℜCv×Ck : STS = ICk }
StEuc = St−1(αt(ICk − βtktktT)) + βtvtktT
X(0) = StEuc / (∥StEuc∥F + ε)
X(j+1) = aX(j) + bX(j)(X(j))TX(j) + cX(j)((X(j))TX(j))2
OSU projects state updates onto the Stiefel manifold, ensuring stable transitions across the cardiac cycle.
OSU Detail
Figure 4. OSU: orthogonalized update preserves manifold structure.

Anatomical Prior-aware Feature Enhancement

APFE decouples anatomy from speckle noise via physics-driven decomposition, building stable structural anchors.
APFE Detail
Figure 5. APFE: contrast decomposition provides noise-resilient cues.

Quantitative Results

SOTA on CAMUS & EchoNet-Dynamic in terms of mDice and mHD95. OSA achieves 94.82% mDice and competitive mHD95 performance, reflecting superior boundary precision crucial for clinical measurement. Inference runs at 35 fps, enabling real-time clinical deployment.
Main Results
Table 1. Segmentation performance on CAMUS and EchoNet-Dynamic.

Qualitative Comparison

OSA shows superior boundary adherence and fewer false positives in speckle-dense regions. Compared to state-of-the-art methods such as SAMed-2 and EchoVim, OSA maintains consistent contour tracking across the full cardiac cycle even under severe acoustic shadowing and low-contrast apical views.
Qualitative Results
Figure 6. Visual comparison. Green = GT, Red = pred, Yellow = overlap.
Preview: 100% Use browser's "Save as PDF" in the print dialog