Deep learning based gestational age estimation from multi-view fetal brain magnetic resonance imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Fetal Development

Background:

• Accurate gestational age (GA) estimation is crucial for fetal development assessment.
• Conventional methods (LMP, ultrasound) often lack accuracy, especially in late pregnancy.
• Deep learning (DL) and MRI offer improved GA prediction by analyzing fetal brain development.

Purpose of the Study:

• Develop DL models for GA prediction using multi-view fetal brain MRI.
• Compare DL model performance against conventional biometric regression techniques.

Main Methods:

• Trained and evaluated DL models on 1,321 fetal MRI scans.
• Utilized an external test set of 80 public MRI scans.
• Explored transfer learning vs. training from scratch and single-view vs. multi-modality inputs.

Main Results:

• Pre-trained ResNet-101 model achieved MAE of 4.47 days (internal) and 6.57 days (external).
• DL model outperformed biometric regression (MAE 9.42 days) on the external test set.
• Explainability analysis showed focus on lateral ventricles, cerebellum, and surrounding brain regions.

Conclusions:

• Multi-view MRI and transfer learning significantly improved DL model accuracy for GA estimation.
• The proposed DL approach surpasses conventional biometric regression.
• The study highlights potential for DL-based GA estimation in prenatal diagnostics.