Medical Imaging
A Multicentre Benchmark Dataset for Comprehensive Landmark-Based Fetal Ultrasound Biometry
This paper introduces an open-access, multicentre benchmark dataset for fetal ultrasound biometry, featuring 4,513 de-identified US images from 1,904 subjects across four clinical sites and seven US devices. The dataset includes expert anatomical landmark annotations for head (BPD, OFD), abdomen (TAD, APAD), and femur (FL) measurements. The study demonstrates that models trained on multicentre data generalize more robustly across clinical sites compared to single-centre training, highlighting the critical problem of domain shift in AI-assisted fetal biometry.
Executive Impact
This research addresses a critical need in AI-assisted fetal biometry by providing a robust, diverse dataset and highlighting the importance of multicentre training for generalizable models in clinical practice.
0
Images
0
Subjects
0
Clinical Sites
0
US Devices
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Dataset Overview
The paper introduces the first publicly available multicentre, multi-device, landmark-annotated dataset covering primary fetal biometry measures (BPD, OFD, APAD, TAD, FL). It contains 4,513 de-identified US images from 1,904 subjects across 4 clinical sites using 7 different US devices. This rich diversity helps address domain shift, a major challenge in AI-assisted fetal growth assessment.
Domain Shift & Generalization
The study quantifies domain shift, showing that models trained and evaluated on single-centre data significantly overestimate performance. Multicentre training substantially improves generalization to unseen acquisition settings, demonstrating NME values below 0.1 for head and abdomen within-domain, but increasing for cross-domain. This highlights the need for diverse training data for robust clinical deployment.
Annotation & Preprocessing Pipeline
The dataset standardizes landmark annotations to a consistent format with subject-disjoint train/test splits. Preprocessing ensures consistency across different US devices and protocols. A Dynamic Orientation Determination (DOD) method is used to enforce measurement-specific orientation consistency during training, avoiding manual point-swapping and mitigating annotation protocol differences that cause cross-domain errors.
Clinical Workflow Integration
Direct landmark detection, as opposed to segmentation-based methods, better aligns with clinical workflow, requiring sonographers to mark two anatomical points per measurement. This is a faster and more intuitive task, making AI-assisted biometry systems more practical for real-world application.
4,513
Images for Multicentre Biometry
| Training Scope | Generalization (NME) | Implication |
|---|---|---|
| Single-Centre |
|
|
| Multicentre (M-C) |
|
|
Enterprise Process Flow
Acquire Raw US Images
→
Standardize Annotations (Landmark-based)
→
Image-Specific Preprocessing
→
Dynamic Orientation Determination (DOD)
→
Model Training & Evaluation
Streamlining Fetal Biometry
Traditional fetal biometry is time-consuming and operator-dependent, leading to high inter-operator variability. The proposed landmark-based approach, combined with AI, reduces this variability and improves workflow efficiency. By focusing on direct landmark detection, the system integrates seamlessly into existing clinical practices, requiring simpler input than complex segmentation. This makes AI-assisted fetal biometry more intuitive and efficient for clinicians, ultimately reducing diagnostic uncertainty.
Calculate Your AI Biometry ROI
Estimate the potential time and cost savings for your enterprise by automating fetal ultrasound biometry with our AI solution. Input your operational metrics to see the impact.
Annual Cost Savings
$0
Annual Hours Reclaimed
0 hours
AI Biometry Implementation Roadmap
Our structured approach ensures a smooth integration of the AI biometry system into your clinical workflow.
Phase 1: Data Preparation & Model Training
Gathering and anonymizing existing ultrasound datasets, followed by training the AI model on your specific acquisition protocols to optimize performance.
Phase 2: Integration & Pilot Testing
Seamlessly integrating the AI solution with your existing PACS and EHR systems. Conducting pilot tests with a small group of sonographers to gather feedback and refine the system.
Phase 3: Full Deployment & Continuous Optimization
Rolling out the AI biometry system across all relevant clinical sites. Ongoing monitoring, performance tuning, and updates to ensure sustained accuracy and efficiency.
Ready to Transform Your Workflow?
Discover how our multicentre AI biometry solution can transform your clinical practice. Book a free consultation with our experts today.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat