💻 computer science

R2AoP: Reliable and Robust Angle of Progression Estimation from Intrapartum Ultrasound

The paper proposes R2AoP, a reliable and robust framework for estimating the Angle of Progression from intrapartum ultrasound that integrates structurally informed segmentation, confidence-guided geometric modeling, and test-time adaptation to overcome imaging noise and boundary ambiguities, achieving superior accuracy and reproducibility across multi-center benchmarks.

Original authors: Yuanhan Wang, Yifei Chen, Beining Wu, Mingxuan Liu, Xiaotian Hu, Chunbo Jiang, Yijin Li, Changmiao Wang, Feiwei Qin, Qiyuan Tian

Published 2026-05-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuanhan Wang, Yifei Chen, Beining Wu, Mingxuan Liu, Xiaotian Hu, Chunbo Jiang, Yijin Li, Changmiao Wang, Feiwei Qin, Qiyuan Tian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a doctor trying to figure out if a baby is ready to be born. They use a special ultrasound camera placed on the mother's lower belly to take a picture of the baby's head and a specific bone in the mother's pelvis (the pubic symphysis). To know if the baby is moving down correctly, they need to measure a specific angle between these two parts. This is called the Angle of Progression (AoP).

Think of this like trying to measure the angle of a ladder leaning against a wall. If you can't clearly see where the ladder touches the ground or where it hits the wall, your measurement will be wrong. In ultrasound images, things are often blurry, noisy, or shadowed, making it very hard for computers to find these exact lines. If the computer makes a tiny mistake finding the edge of the bone, that small error gets "amplified" when calculating the angle, leading to a big, wrong result.

The paper introduces a new system called R2AoP (Reliable and Robust Angle of Progression) to fix this. Here is how it works, using simple analogies:

1. The "Three-Eyed" Detective (Better Vision)

Standard computer programs often look at an ultrasound image with just one "pair of eyes," trying to see the whole picture at once. Sometimes they miss the tiny, tricky details.

The R2AoP Solution: This system uses a three-branch backbone. Imagine a detective who doesn't just look at the crime scene once, but uses three different strategies simultaneously:
- One "eye" looks at the big picture to understand the general layout.
- Two other "eyes" zoom in specifically on the tricky, blurry edges of the baby's head and the mother's bone.
- These three views work together to create a much sharper, more accurate map of where the boundaries actually are, even when the image is noisy.

2. The "Trust Score" Filter (Ignoring Bad Data)

Even with good vision, some parts of an ultrasound image are just too blurry to trust. If a computer tries to measure an angle using a blurry, uncertain point, the result will be garbage.

The R2AoP Solution: The system assigns a "Trust Score" (confidence) to every single point it finds.
- If a point on the baby's head is clear, it gets a high score.
- If a point is fuzzy or in a shadow, it gets a low score.
- When calculating the angle, the system ignores the low-score points and focuses only on the high-trust points. It's like asking a group of people for directions but only listening to the ones who are 100% sure they know the way, while ignoring the ones who are guessing.

3. The "Self-Correcting" GPS (Adapting to New Roads)

Usually, a computer is trained on one type of ultrasound machine (the "Source") and then asked to work on a different machine or in a different hospital (the "Target"). This is like driving a car trained on sunny California roads suddenly being asked to drive in a snowy Minnesota winter. The car might get confused.

The R2AoP Solution: The system includes a Test-Time Adaptation feature.
- When the system sees a new, unfamiliar image, it doesn't just guess. It takes a tiny "self-check" step.
- It looks at its own prediction and asks: "Does this make geometric sense? Is the angle I'm calculating reliable?"
- If the answer is "no," it makes a tiny, quick adjustment to its internal settings to fix the error before giving the final answer. It's like a GPS that realizes, "Wait, I'm in a new city with different street signs; let me quickly recalibrate my map so I don't give you the wrong turn."

The Result

The authors tested this system on data from different hospitals and different ultrasound machines. They found that R2AoP was much better at finding the correct angle than previous methods. It made fewer mistakes, especially when the images were difficult or the equipment was different.

In short: R2AoP is a smarter way for computers to measure a baby's position during labor. It uses multiple "views" to see better, filters out the blurry parts it can't trust, and quickly adjusts itself when it encounters new types of images, ensuring the doctor gets a reliable measurement every time.

Problem Statement

Accurate estimation of the Angle of Progression (AoP) from intrapartum transperineal ultrasound is essential for objective labor progression assessment and clinical decision-making. However, automated AoP estimation faces significant challenges:

Imaging Degradation: Ultrasound images suffer from speckle noise, low contrast, acoustic shadowing, and blurred boundaries, leading to variability in the visibility of the Pubic Symphysis (PS) and Fetal Head (FH).
Error Amplification: AoP is a cascaded geometric measure derived from the PS axis and the FH tangent. Even minor segmentation errors at the boundary level can be geometrically amplified, resulting in significant angular deviations and reduced measurement stability.
Domain Shifts: Clinical deployment involves variations across different ultrasound devices, acquisition parameters, and operators. Models trained on source data often degrade when applied to target domains with different image intensities and textures.

Existing methods often prioritize segmentation accuracy or assume consistent imaging conditions, failing to address the specific need for cross-condition stability and the suppression of error amplification in geometric calculations.

Methodology: R2AoP Framework

The authors propose R2AoP, a framework integrating structurally informed segmentation, confidence-guided geometric modeling, and a lightweight test-time adaptation strategy.

1. Three-Branch Local-Structure-Enhanced Backbone

To improve the delineation of critical boundaries (PS and FH), the authors replace standard backbones with a three-branch shared encoding architecture:

Global Context: One branch preserves global context.
Local Attention: Two additional branches emphasize complementary local regions using a gating mechanism. Spatial gating maps are predicted on shallow, high-resolution features and multiplied pixel-wise with the input to form complementary views.
Fusion: The three streams share an encoder to extract multi-scale features, which are fused at each scale and passed to the decoder. This architecture is trained end-to-end to enhance boundary prediction accuracy and stability.

2. Confidence-Guided Geometric Modeling

Instead of relying solely on binary segmentation masks, the method utilizes pixel-level confidence maps ( $S$ ) predicted by the network:

Weighted Contour Fitting: Boundary points are weighted by their spatial confidence. A weighted ellipse fitting is performed on the FH contour to derive geometric parameters (center, axes, rotation).
AoP Calculation: The AoP is computed based on the triangular relationship between the PS axis and the tangent of the fitted FH ellipse.
Reliability Metric ( $C_{AoP}$ ): An AoP confidence score is defined as the average spatial confidence of the key boundary points used in the geometric calculation. This metric quantifies the reliability of the specific measurement.

3. Geometry-Reliable Test-Time Adaptation (TTA)

To handle domain shifts without target annotations, the authors introduce a lightweight TTA strategy that updates only a small subset of parameters (affine parameters of normalization layers and linear layers) during inference. The optimization objective ( $L_{tta}$ ) combines three unsupervised constraints:

Entropy Minimization ( $L_{ent}$ ): Encourages confident segmentation predictions by minimizing pixel-level entropy, driving probability distributions from flat to sharp.
Spatial Smoothness Regularization ( $L_{tv}$ ): Uses Total Variation (TV) to penalize probability discrepancies between adjacent pixels, preventing the amplification of local noise and ensuring spatial coherence.
Geometry Reliable Constraint ( $L_{aop}$ ): A novel term that maximizes the AoP confidence ( $C_{AoP}$ ) by minimizing $-\log(C_{AoP} + \epsilon)$ . This explicitly constrains the segmentation updates to ensure the resulting geometric measurement remains reliable and stable under domain shifts.

Key Contributions

Integrated Framework: The first framework to explicitly couple boundary reliability modeling with geometric AoP computation to mitigate error amplification in cascaded pipelines.
Structural Enhancement: A three-branch backbone with gating mechanisms specifically designed to enhance the representation of slender (PS) and discontinuous (FH) structures in noisy ultrasound.
Geometry-Aware Adaptation: A novel TTA strategy that utilizes the AoP confidence as a geometric reliability constraint, enabling stable inference in target domains without annotations.
Lightweight Efficiency: The adaptation process is computationally efficient, updating only normalization and linear layer parameters, making it suitable for online deployment.

Experimental Results

The method was evaluated on two target domain benchmarks: JNU-IFM 23 and IUGC 24, using the PSFHS dataset as the source domain.

Performance on JNU-IFM 23: R2AoP achieved an AoP error of 6.21° (reducing the error by 0.07° compared to the second-best method, AoP-SAM). It also showed significant improvements in boundary metrics, with an Average Symmetric Surface Distance (ASD) of 3.12 mm for PS and 2.00 mm for FH.
Performance on IUGC 24: The method demonstrated strong generalization, achieving the best results across all metrics. The AoP error was reduced to 4.22°, with an overall Dice score of 0.95 and an ASD of 2.32 mm.
Ablation Studies: Removing any component (three-branch backbone, geometry constraint, entropy minimization, or TV smoothing) resulted in performance degradation. Notably, the geometry reliability constraint had the most significant impact on the IUGC 24 dataset, confirming its role in stabilizing measurements under severe domain shifts.

Significance and Claims

The paper claims that R2AoP establishes a geometry-aware paradigm for automated obstetric measurements. Its primary significance lies in addressing the "long-standing instability" caused by error amplification in segmentation-to-measurement pipelines.

The authors assert that:

Enforcing geometric reliability is critical for making automated measurements clinically usable, not just for improving segmentation metrics.
The proposed formulation is generalizable and can be extended to other geometry-dependent measurements in ultrasound and beyond.
The framework provides a principled direction for the robust clinical deployment of learning-based measurement systems, particularly in scenarios involving heterogeneous acquisition conditions and domain shifts.

The source code is publicly available, and the work was supported by the National Natural Science Foundation of China and Tsinghua University.