A regression framework to head-circumference delineation from US fetal images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Fetal Development Assessment

Background:

• Accurate measurement of fetal head circumference (HC) from ultrasound (US) images is vital for assessing fetal growth.
• Existing computer-assisted methods aim to reduce variability in HC length measurements.
• Deep learning, particularly convolutional neural networks (CNNs), is increasingly used for fetal head segmentation.

Purpose of the Study:

• To propose a novel deep learning framework for accurate fetal head circumference (HC) delineation.
• To address HC measurement as an edge-delineation problem using regression CNNs, differing from segmentation approaches.
• To reduce intra- and inter-operator variability in HC length measurements.

Main Methods:

• A two-stage deep learning framework combining a region-proposal CNN for head localization and a regression CNN for HC delineation.
• The region-proposal CNN utilizes transfer learning for initial training.
• A novel training strategy based on distance fields is proposed for the regression CNN.

Main Results:

• The framework achieved a mean absolute difference of 1.90 mm and a Dice similarity coefficient of 97.75% on the HC18 Challenge dataset.
• Performance surpassed existing approaches in the literature for HC delineation.
• The system was validated on a dataset comprising 999 training and 335 testing images.

Conclusions:

• The proposed regression CNN framework demonstrates high effectiveness in accurately delineating fetal head circumference.
• The method shows significant potential to assist clinicians in routine clinical practice for fetal growth assessment.
• The approach offers a promising alternative to segmentation-based deep learning methods for HC measurement.