Automatic Ventriculomegaly Detection in Fetal Brain MRI: A Step-by-Step Deep Learning Model for Novel 2D-3D Linear...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Fetal Medicine

Background:

• Accurate measurement of fetal lateral ventricles is crucial for diagnosing ventriculomegaly.
• Manual measurements can be time-consuming and subject to inter-observer variability.
• Deep learning offers potential for automating and standardizing these measurements.

Purpose of the Study:

• To develop and validate an automated deep learning workflow for measuring fetal lateral ventricles in MRI.
• To classify fetal brain MRIs as normal or indicative of ventriculomegaly.
• To compare the AI model's performance against expert radiologists.

Main Methods:

• A UNet-based deep learning model was trained on fetal T2-weighted MRI data (FeTA 2022 dataset) for brain segmentation.
• An automated workflow was created to measure lateral ventricle diameter at the thalamus and choroid plexus levels.
• AI measurements were compared with manual measurements from a general radiologist and a neuroradiologist on a test dataset of 22 cases.

Main Results:

• The AI model achieved 95% accuracy in classifying fetal brain MRIs as normal or ventriculomegaly.
• Lateral ventricle diameter measurements had an error of less than 1.7 mm in 95% of cases.
• AI measurements showed non-significant differences compared to both general radiologists (p=0.9827) and neuroradiologists (p=0.2378), with comparable inter-observer variability.

Conclusions:

• The developed AI workflow accurately measures lateral ventricles in fetal brain MRI and classifies them.
• This AI approach provides measurements comparable to expert radiologists, with potential to improve diagnostic efficiency.
• This study represents a novel application of AI for 2D linear measurement of ventriculomegaly using a 3D model framework.