PaperlayerView exact revision on arXiv

Source-linked AI summary

UniFLM: United Segmentation and Measurement on Fetal Limb Ultrasonic Image

Zeen Zhou, Qiuhua Chen, Xiaojun Cao, Changmao Chen, Chao Sun, Bo Du

arXiv:2608.27240v1cs.CV

TL;DR

Fetal long-bone assessment lacks high-quality annotated data and a unified framework for multiple bones, despite its importance in prenatal skeletal evaluation. The paper constructs the FLB dataset and proposes UniFLM, combining semantic alignment, noise-robust sampling, and clinician-pattern-based landmark localization. UniFLM achieves state-of-the-art performance across four bone categories, including a 2.35% Dice improvement on forearm structures, processes images at 105 FPS, and keeps measurement errors within human inter-observer variability.

  • Problem

    Fetal limb assessment lacks high-quality annotated data and systematic, unified methods for precise measurement across multiple long bones.

  • Method

    UniFLM combines the FLB multi-category dataset with SASC, PoSamp, and PRM for unified segmentation and clinician-style measurement.

  • Results

    2.35% Dice improvement on challenging forearm structures; UniFLM achieves state-of-the-art performance across four fetal long bone categories and processes images at 105 FPS.

  • Takeaways & Limitations

    UniFLM provides a real-time decision support tool for prenatal skeletal assessment, with measurement errors within human expert inter-observer variability.

  • Takeaways & Limitations

    Further validation on larger, multi-vendor datasets containing pathological cases is required to strengthen generalization claims.

Abstract

from arXiv · show

Prenatal ultrasound examination is crucial for assessing fetal limb development and detecting congenital anomalies. However, existing artificial intelligence models often overlook fetal lethal skeletal dysplasias due to the lack of high-quality annotated data and a unified framework for multiple long bones. Moreover, generic segmentation models struggle with the inherent noise and semantic gaps in ultrasound images. To address these challenges, we construct the Fetal Limb Bones (FLB) dataset, comprising high-quality annotations for the humerus, femur, tibia-fibula, and radius-ulna. Furthermore, we propose UniFLM, a unified framework for automatic cross-plane segmentation and measurement. UniFLM incorporates a Semantic-Aware Skip Connection module to bridge the semantic gap between encoder and decoder features, and a Positive Sampling strategy to adaptively filter noise and extract essential semantic information. Finally, a Point Regression Mapping module is introduced to learn clinician annotation patterns for precise bone length measurement. Extensive experiments conducted on the FLB dataset demonstrate that the proposed UniFLM achieves superior accuracy and enhanced generalization capabilities in fetal long bone assessment compared to current state-of-the-art models.

1. Introduction

The paper identifies gaps in fetal long-bone assessment and introduces the FLB dataset and UniFLM, a unified framework for cross-category segmentation and measurement. UniFLM combines semantic feature alignment, noise suppression, and clinician-pattern-based landmark localization, with experiments showing superior cross-category generalization.

  • Motivation: Fetal limb long-bone morphology and length are clinically significant for diagnosing lethal skeletal dysplasias during prenatal assessment.The relevant bones are the humerus, femur, tibia-fibula, and radius-ulna.
  • Research gaps: Low-resolution, uneven-contrast, noisy ultrasound images and multiple bones in one frame complicate unified long-bone segmentation.These conditions make paired-bone configurations particularly challenging for a common framework.
  • Research gaps: Existing datasets and models provide limited support for precise, multi-bone fetal limb measurement and early skeletal-dysplasia diagnosis.FPUS23 supplies bounding boxes, while DeepGA predicts gestational age from femur length only.
  • Contributions: The FLB dataset provides high-quality, multi-category ultrasound annotations curated by senior experts for clinically relevant model training.It covers the humerus, femur, tibia-fibula, and radius-ulna.
  • Contributions: UniFLM integrates SASC for encoder-decoder semantic alignment, PoSamp for ultrasound-noise suppression, and PRM for clinician-style landmark localization.The framework is designed for end-to-end automated segmentation and biometric measurement.
  • Results: Extensive evaluations report superior generalization in cross-category fetal bone measurement, while the FLB dataset and source code are publicly available.The source code is provided through the project repository.

2. Related Work

Related work spans U-Net variants, Transformer-based segmentation, foundation models, attention mechanisms, and landmark-regression strategies. The paper positions unified multi-bone fetal segmentation and measurement as an unresolved gap under ultrasound-specific noise and resource constraints.

  • Segmentation methods: U-Net and its variants established encoder-decoder segmentation, while attention and dense skip pathways expanded feature fusion.Examples include Attention U-Net, UNet++, and UNet 3+.
  • Segmentation methods: Vision Transformer models capture global context but require substantial computational resources, motivating more efficient segmentation alternatives.TransUNet and Swin-Unet are cited as representative examples.
  • Ultrasound challenges: Fetal ultrasound analysis must address low contrast, acoustic shadows, and speckle noise, prompting increased use of channel and spatial attention.SE-Net is described as recalibrating channel importance, while spatial attention focuses on relevant anatomical regions.
  • Open gap: A unified framework that simultaneously segments and measures multiple fetal limb bones remains a significant research gap.The gap concerns coordinated use of feature alignment and attention strategies across bone categories.
  • Measurement methods: Heatmap-based landmark methods provide spatial uncertainty but require costly coordinate extraction, whereas direct regression is efficient but may lack spatial context and geometric robustness.The related work contrasts these trade-offs as motivation for hybrid approaches.

3. Methodology

UniFLM unifies fetal long-bone segmentation and measurement through semantic feature alignment, noise suppression, and coarse-to-fine landmark refinement. Its encoder–decoder pipeline produces segmentation masks and precise measurement points from ultrasound images.

  • UniFLM uses a 6-stage encoder and 5-stage decoder for unified fetal long-bone segmentation and biometric measurement.
  • Semantic Alignment Skip Connection (SASC): SASC aggregates multi-scale encoder features, aligns them through cross-attention, and redistributes them to the decoder to bridge semantic discrepancies.
  • Positive Sampling (PS): Positive Sampling adaptively filters background noise from the deepest encoder feature and produces a robust representation for subsequent decoding.A residual connection preserves essential spatial information while enhancing salient features.
  • Point Regression Mapping (PRM): PRM refines initial keypoints extracted from coarse segmentation masks into precise landmarks using a patch-based refinement network.The refinement network predicts landmark locations relative to local patch centers, while cross-entropy treats localization as classification over a spatial grid.
  • Segmentation and measurement: The decoder first generates an initial mask, from which rough bone endpoints are extracted and passed to PRM for final measurement-point prediction.

4. Experimental Results

The FLB dataset contains diverse, clinically annotated ultrasound images spanning four fetal long-bone categories and gestational ages. Annotation agreement was high, and patient-wise splitting was used to evaluate generalization to unseen clinical cases.

  • Dataset composition: 1,690 ultrasound images cover the humerus, femur, forearm, and leg categories, with 600, 500, 295, and 295 images respectively.
  • Dataset diversity: Images were acquired across multiple ultrasound systems and gestational ages ranging from 14 to 40 weeks.
  • Annotation protocol: Three experienced sonographers provided pixel-level masks, endpoint coordinates, and image-quality scores for each image.
  • Annotation reliability: Mean inter-annotator agreement reached a Dice coefficient of 0.92 and an endpoint distance of 1.1 mm.Final annotations used majority voting with senior expert adjudication for persistent disagreements.
  • Dataset split: A patient-wise 7:1:2 training/validation/test split prevents images from the same patient appearing in multiple subsets.This setup evaluates generalization to completely new clinical cases.

4.2. Implementation Details

UniFLM is evaluated against diverse segmentation baselines using overlap, boundary, measurement, and statistical metrics. It achieves the best overall performance, with larger gains on challenging paired-bone structures.

  • Evaluation setup: Experiments use Dice, IoU, HD95, MED, MAE, and MSE to evaluate segmentation overlap, boundary accuracy, endpoint localization, and measurement error.The setup also compares classic, Transformer-based, foundation-model, Mamba-based, and U-KAN baselines.
  • Quantitative results: UniFLM achieves the best overall performance across all fetal bone categories, with particularly notable improvements on challenging anatomical structures.Table 1 reports the comprehensive quantitative comparison on the FLB dataset.
  • Quantitative results: UniFLM improves Dice by 1.25% on Femur and 1.26% on Humerus over the best baseline.These gains occur on single-bone structures and indicate consistent improvements on relatively easier tasks.
  • Quantitative results: UniFLM improves Dice by 2.35% on Forearm and 0.70% on Leg over the best baseline.The larger Forearm gain highlights performance on paired-bone anatomical configurations.
  • Statistical significance: UniFLM shows statistically significant improvements with p < 0.05 over all baselines on Forearm and Leg datasets.Improvements on Femur and Humerus are consistent but not always statistically significant because baseline performance is already high.

4.4. Ablation Study

Ablation experiments isolate the contributions of SASC, PoSamp, and PRM. Their combination improves segmentation and measurement, with especially strong effects on challenging bone structures.

  • Module Contribution Analysis: SASC alone improves Femur Dice by 0.86% but decreases Forearm Dice by 1.34%, whereas PoSamp alone improves every bone category.PoSamp’s largest individual gain is +0.14% Dice on Humerus.
  • Module Contribution Analysis: Combining SASC and PoSamp yields synergistic gains, including a 1.19% increase over baseline on Forearm.The combined improvement exceeds the sum of their individual contributions on the challenging Forearm category.
  • Module Contribution Analysis: Adding PRM further elevates all evaluation metrics and produces a 0.70% Dice gain on Leg.The result demonstrates PRM’s role in refining boundary predictions when endpoint localization is difficult.
  • Measurement Module Analysis: PRM reduces MED by 16-23% and MAE by 15-18% compared with geometric post-processing methods.Table 4 evaluates measurement accuracy with and without PRM integration.

4.5. Clinical Reliability Analysis

Clinical reliability analysis examines measurement errors across bone categories, gestational ages, and noise levels. UniFLM remains accurate across these conditions, including clinically challenging cases.

  • Measurement errors: Over 85% of Femur and Humerus measurements fall within the clinically acceptable 2.0 mm error threshold.The CDF analysis reports the proportion of measurements meeting the threshold across bone categories.
  • Measurement errors: Approximately 75% of Forearm and Leg measurements fall within 2.0 mm, while 90% fall within 3.0 mm.These paired-bone structures show lower compliance at 2.0 mm but broader compliance at 3.0 mm.
  • Gestational age robustness: Performance is relatively stable across Early, Middle, and Late gestational-age groups, with a slight decrease in Late GA.The decrease is associated with increased acoustic shadowing from bone calcification, while PRM helps maintain measurement accuracy.
  • Noise robustness: UniFLM maintains Dice scores above 80% under moderate noise with σ ≤0.1, outperforming standard U-Net as noise increases.The reported robustness is attributed to PoSamp preventing overfitting to noise patterns during training.

4.6. Computational Efficiency

The study evaluates computational efficiency and clinical measurement agreement alongside qualitative segmentation quality. UniFLM combines real-time throughput with measurement errors comparable to expert variability.

  • Computational Efficiency: UniFLM uses 35.8M parameters and achieves 105 FPS, providing the best reported accuracy-efficiency trade-off.The computational comparison uses 256 × 256 inputs and identifies suitability for real-time deployment on standard GPU hardware.
  • Clinical Reliability: UniFLM’s measurement error is comparable to inter-expert variability for Femur and Humerus and remains within 0.5 mm for Forearm and Leg.These results come from comparison against individual expert measurements.
  • Qualitative analysis: Qualitative results show more complete and anatomically accurate UniFLM segmentations, especially in acoustic-shadow regions where other methods produce fragmented masks.Grad-CAM indicates deeper layers capture global bone structure while shallower layers progressively refine representations.

4.9. Analysis of Acoustic Shadows and Bone Calcification

Acoustic shadowing from fetal bone calcification obscures distal bone boundaries and complicates segmentation and measurement. UniFLM addresses these challenges while maintaining stable training across key loss-weight settings.

  • Bone calcification blocks ultrasound waves, creating shadows that obscure distal boundaries and complicate accurate biometric measurements.
  • UniFLM's qualitative results show more complete, anatomically accurate segmentations in shadowed regions, with decoder attention focused on bone boundaries and background noise suppressed.
  • Severe shadowing can make posterior boundaries invisible, causing standard edge-based models to produce incomplete C-shaped masks and measurement errors.
  • PRM regresses anatomical endpoints from global semantic features, using learned bone-shaft shape priors when local visual evidence is missing.
  • Training remains stable for λ1 ∈[0.7, 1.3], with optimal performance at λ1 = 1.0 and smooth convergence without oscillation.

5. Discussion

UniFLM combines complementary modules to address acoustic-shadow ambiguity and noisy supervision, while its broader applicability remains bounded by difficult paired-bone and imaging scenarios.

  • SASC, PoSamp, and PRM jointly improve robustness by filtering encoder noise, focusing supervision on high-confidence regions, and decoupling measurement from pixel-level segmentation.
  • SASC propagates semantic information into shadowed areas, PoSamp reduces sensitivity to abrupt shadow-boundary intensity changes, and PRM infers endpoints from learned shape priors.
  • Severe radius-ulna or tibia-fibula overlap produces merged segmentations in approximately 5% of paired-bone images.
  • The modules are presented as potentially applicable to other noise-sensitive modalities, including low-dose CT and speckle-affected OCT, while PRM targets measurement from imperfect masks.

6. Conclusion

The paper introduces the 1,690-image FLB dataset and UniFLM for unified fetal long-bone segmentation and measurement. Experiments report state-of-the-art performance, real-time processing, and human-level measurement-error range, while broader validation remains necessary.

  • The FLB dataset contains 1,690 images, and UniFLM integrates SASC, PoSamp, and PRM for fetal limb segmentation and measurement.
  • UniFLM achieves state-of-the-art performance across all four fetal long-bone categories.
  • 2.35% Dice improvement on challenging forearm structures accompanies processing at 105 FPS and measurement errors within human expert inter-observer variability.
  • Further validation on larger, multi-vendor datasets containing pathological cases is required to strengthen generalization claims.
Loading 2608.27240v1…