Source-linked AI summary
Expert-like Bone Ultrasound Segmentation through Expert-in-the-loop Mask-conditioned Progressive Learning
Arash Tavangar, Larissa K. Chiu, Hamidreza Khodashenas, Gregory K. Berry, Amir Hooshiar
TL;DR
Manual ultrasound bone segmentation is labor-intensive, while existing interactive methods do not fully model brush-based refinement or evolving expert intent. ExiL learns structured completion from imperfect brush masks and expert-approved updates, improving segmentation quality and reducing annotation effort.
Problem
Manual ultrasound bone segmentation is labor-intensive, and existing interactive methods do not fully support dense brush-mask completion or learning evolving expert annotation behavior.
Method
ExiL combines synthetic expert-like brush-mask simulation, mask-conditioned refinement, and online adaptation from expert-approved updates.
Results
The expert-in-the-loop framework improves segmentation quality over rough annotations while reducing structured annotation errors and annotation time.
Takeaways & Limitations
Modeling expert correction behavior provides a practical pathway for accelerating medical-image annotation and improving segmentation reliability in clinical ultrasound applications.
Takeaways & Limitations
Learning expert-like policies requires intermediate annotation states and correction trajectories, but such datasets are limited.
Abstract
from arXiv · showhide
Manual annotation remains a major bottleneck in ultrasound (US) bone segmentation, where experts typically iteratively refine rough brush masks rather than delineating precise contours in a single pass. We present ExiL, a mask-conditioned progressive learning framework that models annotation as a structured refinement trajectory. ExiL combines a synthetic expert-like brush simulator based on signed distance fields with a lightweight 7.8M-parameter U-Net that learns to complete and refine imperfect masks from US images. During deployment, an expert mode updates the model directly from accepted refinements, enabling continual adaptation to expert behavior. Evaluated using UltraBones100k cadaver data for quantitative segmentation and a prospective volunteer dataset for annotation-efficiency analysis, ExiL reduced single-expert average annotation time from 60 to 20 seconds per frame (66.7\%) and improved mean Dice by approximately 0.045 over non-progressive training, while achieving 0.87 Dice and 2.7 px boundary error in the best trajectory-aware setting. With 10--50 ms inference, ExiL enables real-time, self-improving annotation for US-guided orthopedic workflows in practical clinical labeling.
1 Introduction
Ultrasound bone segmentation is clinically valuable but difficult to scale because manual annotation is laborious and existing interactive methods do not model evolving expert behavior. ExiL addresses these gaps with progress-aware refinement, brush-behavior modeling, and expert-approved adaptation.
- Clinical Need and Workflow: Ultrasound supports low-cost, real-time, non-ionizing imaging, making bone surface segmentation important for image-guided orthopedic procedures.US-based bone registration has been proposed for preoperative and intraoperative workflows.
- Clinical Need and Workflow: Speckle noise, low contrast, acoustic shadowing, and observer variability make precise manual bone-boundary delineation time-intensive and difficult to scale.Manual expert annotation remains the clinical gold standard despite these constraints.
- Related Works: Most existing methods treat segmentation as static prediction or independent interactions, relying on high-quality labels and failing to learn experts’ evolving annotation behavior.Interactive systems commonly use clicks or scribbles without representing progress-aware correction trajectories.
- Contributions: ExiL integrates progress conditioning, behavioral modeling, iterative expert-approved adaptation, synthetic brush simulation, and mask-conditioned boundary refinement.The framework is designed to address the mismatch between existing interaction models and real expert workflows.
- Limitations: Existing approaches generally do not model dense brush-like annotations, infer missing anatomy from incomplete masks, or personalize behavior from expert-approved updates.Intermediate annotation states and authentic correction trajectories are rarely available in medical imaging.
2 Materials and Methods
The study trains and evaluates a mask-conditioned completion model using synthetic expert-like brush states and expert-in-the-loop refinement. It combines geometric and boundary cues with cadaver segmentation evaluation and prospective volunteer annotation-efficiency testing.
- Materials and Methods: Approximately 2.5k UltraBones100k frames were used to study completion learning under realistic data scarcity despite the dataset containing over 100k images.The ground truth was generated through CT-to-US spatial alignment.
- Synthetic Dataset: Expert-imitating Annotation Synthesis: Synthetic brush masks use signed distance fields, sampled coverage ratios of 10–30%, 40–60%, and 70–90%, and smoothed spatial bias to mimic rough and refined painting.Gaussian smoothing with σ = 5 px creates uneven inward painting patterns resembling manual annotation.
- Model: The completion model is a mask-conditioned U-Net with approximately 7.8M trainable parameters, using image, rough-mask, distance-to-brush, and edge information.Distance-to-brush supplies geometric context, while edge maps emphasize boundary cues.
- Expert-in-the-Loop Framework: ExiL predicts a refined mask from an ultrasound image and expert brush mask, then re-invokes the model after iterative expert edits until acceptance.Expert-approved final masks support deployment-time adaptation, while the objective combines overlap, contour alignment, and false-positive suppression.
- Validation Studies: Quantitative Dice and boundary error were evaluated on UltraBones100k cadaver data, while annotation efficiency was measured on 360 frames from a prospective three-participant volunteer dataset.The volunteer sequences were not used during training or model development.
3 Results and Discussion
ExiL improved annotation efficiency and segmentation refinement through expert-in-the-loop updates, with benefits most evident for low-coverage prompts and trajectory-aware interaction.
- Annotation efficiency: Expert-in-the-loop variants substantially reduced annotation time, while trajectory-aware refinement required fewer interaction cycles to reach acceptable masks.These efficiency gains were consistent with improved completion quality.
- Ablation findings: Expert-mode updates produced the largest ablation gains, identifying real expert interaction patterns as the primary improvement driver.Boundary-aware losses reduced boundary misalignment, and trajectory-aware supervision improved refinement stability in low-coverage cases.
- Input-channel analysis: Distance-to-brush features improved low-coverage completion, edge cues improved boundary snapping, and combining both achieved the best boundary alignment.The channel ablation compared the US image, rough mask, distance-to-brush map, and edge map.
- Coverage sensitivity: Performance differences were more pronounced under low-coverage prompts, where ExiL v2 produced better boundary alignment and more complete masks.At high coverage, all models performed comparably well.
- Coverage sensitivity: Medium-coverage inputs produced results comparable to high-coverage inputs, whereas extremely sparse or anatomically inconsistent masks caused performance declines.Trajectory-aware learning supported spatial inference with partial supervision but could prioritize trajectory advancement over completing sparse regions.
- Online adaptation: Sequential expert-mode updates yielded consistent improvement without degradation, suggesting stable adaptation to expert editing patterns.The evaluation tracked model behavior across annotated frames and measured Dice improvement as annotated frames increased.
4 Conclusion
The paper presents ExiL as an expert-in-the-loop framework for refining brush-based ultrasound bone annotations through progressive interaction, improving segmentation accuracy and reducing annotation effort.
- Conclusion: ExiL combines synthetic annotation simulation with expert-mode online updates to refine brush-based ultrasound bone annotations through progressive interaction.The framework is presented as a practical autocorrection approach for medical image annotation.
- Conclusion: The framework improves overlap and boundary accuracy while reducing structured annotation errors.The authors position expert correction modeling as a scalable pathway for accelerating annotation workflows and enhancing segmentation reliability.