Subspace-Guided Semantic and Topological Invariant Registration for Annotation-Free Ultrasound Plane Quality Control

Chunzheng Zhu, Jianxin Lin, Feng Wang, Cheng Jiang, Guanghua Tan, Zhenyu Zhou, Shengli Li, Kenli LiarXiv 2026 · arXiv

TL;DR

STRIQ is a new framework that automates ultrasound quality control without requiring expensive manual annotations. By recasting quality assessment as a registration problem—measuring how well a query image aligns with automatically selected high-quality "anchors" in orthogonal feature subspaces—it achieves state-of-the-art clinical correlation in real-time (~180 FPS).

What Problem Does This Paper Address?

Ultrasound (US) imaging is the backbone of prenatal screening, but its effectiveness is notoriously dependent on the operator's skill and the patient's anatomy. Suboptimal image quality can lead to misdiagnosis rates as high as 30%.

While automated Quality Control (QC) could help, existing methods face three major hurdles:

  1. Annotation Bottleneck: Most models require thousands of per-plane quality labels from expert sonographers, which is costly and slow.
  2. Geometric Sensitivity: Ultrasound images are prone to complex spatial deformations. Current "pseudo-labeling" methods often fail to distinguish between a poor-quality image and a good-quality image that is simply rotated or slightly deformed.
  3. Negative Transfer: In multi-plane models (e.g., one model for heart, abdomen, and face), features from one anatomical plane often interfere with another, degrading overall performance.

Methods

STRIQ (Subspace-Guided Registration for Ultrasound Quality Control) introduces a registration-driven paradigm to solve these issues. Instead of predicting a quality score directly, the model asks: "How well can I align this query image to a known high-quality reference?"

Comparison of ultrasound QC paradigms. STRIQ (3) avoids the instability of pseudo-labeling and the errors of structure-weighted schemes.

The framework consists of three novel components:

  • Variance-Spectrum Anchor Selection: The system autonomously distills "gold standard" reference images from unlabeled data. It uses a variance criterion to find the most morphologically stable and high-confidence prototypes for each anatomical plane.
  • Latent Registration Aligner (LRA): This module performs cascaded affine transformations in a hierarchical feature space. It captures correspondences ranging from fine tissue boundaries to global organ positioning, ensuring the quality metric is grounded in topological consistency.
  • Orthogonal Knowledge Subspace (OKS): To solve the "negative transfer" problem, OKS decomposes the model's internal representations into mutually orthogonal low-rank subspaces. Each anatomical plane (e.g., fetal heart vs. kidneys) gets its own dedicated "knowledge partition," preventing inter-plane interference while allowing for controlled shared learning.
Overview of the STRIQ architecture, highlighting the LRA for alignment and the OKS for plane-discriminative subspaces.

Datasets

The authors evaluated STRIQ on two diverse datasets:

  • US4QA (In-house): A large fetal ultrasound dataset containing 30,757 images across four planes (4-chamber heart, abdomen, kidneys, and face) from 324 examinations.
  • CAMUS (Public): A cardiac dataset featuring Apical 2-Chamber (A2C) and 4-Chamber (A4C) views, used to test the model's cross-anatomy robustness.

Key Insights and Ideas

  • Registrability as a Proxy for Quality: The core "aha moment" is that an image's ability to be registered to a high-quality anchor is a robust, annotation-free indicator of its own quality. If an image is too blurry or missing key structures, the registration will fail, resulting in a low quality score.
  • Orthogonality Prevents Confusion: By forcing different anatomical planes into orthogonal subspaces, the model avoids the "jack of all trades, master of none" trap. The t-SNE visualizations in the paper show that as the subspace rank is optimized, different organ planes form distinct, compact clusters.
t-SNE visualization showing how OKS (at rank r=16) creates distinct, optimal clusters for different anatomical planes.
  • Topological Invariance: Because the LRA uses affine transformations and smoothness regularizers, the model learns to be invariant to standard clinical deformations (like probe tilting) while remaining highly sensitive to actual quality degradation (like signal loss or anatomical absence).

Key Results

STRIQ sets a new benchmark for annotation-free ultrasound QC:

  • Superior Correlation: On the US4QA dataset, STRIQ achieved an average Spearman Rank Correlation Coefficient (SRCC) of 0.845, outperforming the previous state-of-the-art (CRL-UIQA) by 6.6%.
  • Cross-Anatomy Robustness: On the cardiac CAMUS dataset, it outperformed the baseline by 7.8%, proving it isn't just limited to fetal imaging.
  • Deformation Resilience: Qualitative results show the model correctly assigns high scores to images with rigid/non-rigid deformations that preserve clinical structures, but sharply penalizes images where structures are missing or severely distorted.
  • Real-Time Performance: The system runs at approximately 5.6 ms per frame on an NVIDIA RTX 4090, making it suitable for real-time guidance during live scans.
STRIQ's quality score response under various rigid and non-rigid deformations. The model remains robust to minor perturbations but penalizes severe structural loss.

Limitations and Future Work

While STRIQ is highly effective for 2D plane quality control, the paper focuses on standard diagnostic planes. Extending this framework to 3D ultrasound volumes or video-based temporal consistency could be a promising next step. Additionally, while the anchor selection is automated, the system still relies on having a pool of unlabeled data that contains at least some high-quality examples to serve as anchors. Exploring how the system behaves in extremely low-resource settings where high-quality images are rare remains an open question.

Summary updated

Abstract

Reliable quality control (QC) of ultrasound images is essential for both real-time acquisition guidance and retrospective clinical audit, yet existing approaches rely heavily on per-plane annotations, or employ pseudo-labeling prone to systematic bias under spatial deformations inherent in clinical acquisition. We present STRIQ, a registration-driven framework that recasts annotation-free US plane quality control as a subspace-guided consistency measurement problem. Specifically, STRIQ introduces a Latent Registration Aligner (LRA) to establish hierarchical feature space correspondences between query images and variance-driven anchors, which are autonomously distilled from unlabeled data via a variance spectrum criterion to serve as structurally stable prototypes. To further disambiguate anatomical planes and mitigate negative knowledge transfer, we propose an Orthogonal Knowledge Subspace (OKS) module. The OKS decomposes plane-specific representations into mutually orthogonal subspaces, enabling fine-grained expert collaboration while preventing inter-plane interference, ensuring that the quality metric is grounded in principled subspace proximity. Extensive experiments on the in-house US4QA and public CAMUS datasets demonstrate that STRIQ achieves state-of-the-art correlation with clinical quality scores, establishing a new paradigm for annotation-free, real-time reliable ultrasound quality control. Our code is available at https://github.com/zhcz328/STRIQ.
BibTeX
@misc{zhu2026subspace,
  author = {Chunzheng Zhu and Jianxin Lin and Feng Wang and Cheng Jiang and Guanghua Tan and Zhenyu Zhou and Shengli Li and Kenli Li},
  title = {{Subspace-Guided Semantic and Topological Invariant Registration for Annotation-Free Ultrasound Plane Quality Control}},
  year = {2026},
  eprint = {2605.25396},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.25396},
}
Figure