Development of a Deep Learning Algorithm for Posterior Fossa Abnormality Recognition on First-Trimester US Screening...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Fetal Medicine

Background:

• Early detection of fetal anomalies is crucial for timely intervention.
• Assessing the posterior fossa in the first trimester can be challenging.
• Deep learning offers potential for automated analysis of medical images.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for automated posterior fossa assessment.
• To identify open spina bifida (OSB) and cystic posterior fossa (CPF) anomalies on first-trimester ultrasound scans.
• To improve the accuracy and efficiency of first-trimester fetal anomaly screening.

Main Methods:

• Retrospective analysis of midsagittal fetal brain ultrasound images (11-14 weeks gestation).
• Manual annotation of posterior fossa regions.
• Training and validation of three convolutional neural networks using a 70/30 split.
• Ensemble averaging of predictions from cross-validated models.

Main Results:

• The MobileNetV3 Large Weights model achieved an AUC of 0.94.
• The algorithm demonstrated high accuracy (88%), recall (81%), and specificity (93%) in detecting OSB and CPF anomalies.
• OSB was classified with 93% accuracy and higher recall compared to CPF anomalies.

Conclusions:

• Deep learning, specifically MobileNetV3 Large Weights, can accurately assess the fetal posterior fossa.
• The algorithm effectively distinguishes normal scans from those with OSB or CPF anomalies.
• This AI tool shows promise for enhancing first-trimester screening for major fetal abnormalities.