Automatic Localization of the Pons and Vermis on Fetal Brain MR Imaging Using a U-Net Deep Learning Model

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Fetal Development

Background:

• Fetal MRI enhances perinatal developmental disorder identification, improving ultrasound accuracy.
• Manual MRI measurements are time-consuming, require expertise, and have inter-observer variability.
• There is a shortage of pediatric neuroradiologists.

Purpose of the Study:

• To develop a deep learning model for automatic identification of anatomic landmarks on the pons and vermis in fetal brain MRI.
• To create a pipeline for suggesting suitable images for pons and vermis measurements.
• To improve the efficiency and accuracy of fetal brain MRI analysis.

Main Methods:

• Retrospective analysis of fetal brain MRI from 55 pregnant patients using a HASTE protocol.
• Development of a U-Net-based deep learning model for landmark identification (pons and vermis).
• Four-fold cross-validation using gestational age-divided datasets to assess model accuracy.

Main Results:

• The deep learning model achieved ≥90% confidence in 85% of cases with a mean error <2.22 mm.
• Anterior and posterior pons, and anterior vermis landmarks showed superior estimation accuracy and confidence.
• A user-friendly graphic interface was developed for clinical application.

Conclusions:

• The deep learning pipeline significantly reduces the time for selecting fetal brain images and performing anatomic measurements.
• This AI-driven approach enhances efficiency and accuracy in fetal neuroimaging analysis.
• The tool assists radiologists in diagnosing perinatal developmental disorders more effectively.