Prediction of fetal brain gestational age using multihead attention with Xception

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Fetal Medicine

Background:

• Accurate gestational age (GA) prediction is vital for fetal development monitoring and prenatal care.
• Traditional GA prediction methods often lack precision and efficiency.
• Deep learning (DL) offers a promising avenue for enhancing GA prediction accuracy.

Purpose of the Study:

• To develop and evaluate a novel DL approach for GA prediction using fetal brain MRI.
• To combine the Xception pretrained model with a multihead attention (MHA) mechanism for feature extraction and prediction.
• To assess the model's performance across different anatomical views and compare it with existing state-of-the-art (SOTA) methods.

Main Methods:

• Utilized a dataset of 52,900 fetal brain MRI images from 741 patients (GA 19-39 weeks).
• Employed the Xception model for feature extraction, followed by configurable Multihead Attention (MHA) mechanisms.
• Trained the model to predict GA in days, optimizing parameters like attention heads and key/value space dimensionality.

Main Results:

• Achieved high accuracy on the test set: R-squared (R²) of 96.5%, Mean Absolute Error (MAE) of 3.80 days, and Pearson Correlation Coefficient (PCC) of 98.50%.
• 5-fold cross-validation confirmed reliability with average R² of 95.94%, MAE of 3.61 days, and PCC of 98.02%.
• Demonstrated superior performance across axial and sagittal views, outperforming other SOTA models.

Conclusions:

• The proposed DL model integrating Xception and MHA provides a highly accurate and reliable method for GA prediction from fetal brain MRI.
• The model's effectiveness across multiple anatomical views suggests its robustness for clinical application.
• This approach holds significant potential to aid clinicians in precise GA determination, optimizing prenatal care.