Comparative evaluation of interpretation methods in surface-based age prediction for neonates.

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Artificial Intelligence in Medicine

Background:

• Deep learning enhances brain age prediction accuracy in the third trimester.
• The "black box" nature of deep learning hinders clinical interpretability in neuroimaging.
• Existing interpretability methods from other fields are not directly applicable to neuroimaging.

Purpose of the Study:

• To evaluate regional age prediction and perturbation-based saliency maps for neonatal brain age prediction.
• To assess the adaptability of these interpretability methods for neuroimaging data.
• To provide clinically relevant insights into brain development using interpretable AI.

Main Methods:

• Analysis of 664 T1 MRI scans using the NEOCIVET pipeline.
• Extraction of brain surface and cortical features.
• Comparative analysis of saliency index (SI) with relative brain age (RBA) and structural covariance networks.

Main Results:

• The saliency index effectively pinpointed brain regions crucial for age prediction.
• Perturbation techniques demonstrated advantages in handling complex medical data.
• Identified key brain regions contributing to accurate clinical factor indication.

Conclusions:

• Perturbation-based saliency maps offer a promising approach for interpretable neonatal brain age prediction.
• These methods can guide personalized clinical interventions for premature neonates.
• The study provides a framework for developing clinically relevant, interpretable deep learning models in healthcare.