FUGC: Benchmarking Semi-Supervised Learning Methods for Cervical Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Obstetrics & Gynecology

Background:

• Accurate cervical segmentation in transvaginal ultrasound (TVS) is crucial for spontaneous preterm birth (PTB) risk assessment.
• Supervised learning methods are hindered by the limited availability of labeled cervical ultrasound data.
• Developing robust AI models requires standardized benchmarks for evaluating cervical segmentation techniques.

Purpose of the Study:

• Introduce the Fetal Ultrasound Grand Challenge (FUGC), the first benchmark for semi-supervised learning in cervical segmentation.
• Provide a comprehensive dataset and evaluation framework for AI-based cervical analysis.
• Facilitate advancements in AI-assisted PTB risk assessment.

Main Methods:

• Establishment of the FUGC benchmark with 890 TVS images (500 training, 90 validation, 300 testing).
• Evaluation of submitted methods using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and runtime (RT) with a weighted score (0.4/0.4/0.2).
• Ten teams with 82 participants developed and submitted semi-supervised learning solutions.

Main Results:

• The best-performing methods achieved 90.26% mean DSC, 38.88 mean HD, and 32.85 ms mean RT.
• Demonstrated the effectiveness of semi-supervised approaches in scenarios with limited labeled data.
• Highlighted innovative solutions developed by participating teams.

Conclusions:

• FUGC provides a standardized benchmark for cervical segmentation research.
• Semi-supervised learning shows significant promise for cervical analysis in low-data regimes.
• The challenge lays the groundwork for AI-driven clinical tools for PTB risk assessment.