PTNet3D: A 3D High-Resolution Longitudinal Infant Brain MRI Synthesizer Based on Transformers.

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Artificial Intelligence

Background:

• Longitudinal neurodevelopmental studies require high-quality infant brain MRI data.
• Infant MRI data acquisition is challenging due to motion artifacts and limited attention spans.
• Existing analytical approaches and data augmentation methods are insufficient for modeling infant neurodevelopmental trajectories.

Purpose of the Study:

• To develop a novel 3D MRI synthesis framework for generating realistic infant brain images.
• To address the challenges of corrupted or missing MRI scans in longitudinal infant studies.
• To improve the accuracy and generalizability of infant brain MRI analysis.

Main Methods:

• Introduced PTNet3D, a 3D MRI synthesis framework utilizing transformer and performer layers with attention mechanisms.
• Conducted experiments on the Developing Human Connectome Project (dHCP) and Baby Connectome Project (BCP) datasets.
• Compared PTNet3D with existing Convolutional Neural Network-based Generative Adversarial Networks (CNN-based GANs).

Main Results:

• PTNet3D demonstrated superior synthesis accuracy and generalization compared to CNN-based GANs on infant brain MRI datasets.
• PTNet3D generated more realistic scans than CNN-based models when using multi-age input data.
• Synthesizing corrupted scans with PTNet3D significantly improved infant whole brain segmentation.

Conclusions:

• PTNet3D offers a promising solution for enhancing infant neuroimaging data quality.
• The framework has potential applications in reconstructing corrupted or missing MRI scans.
• Improved MRI data quality through synthesis can lead to more accurate neurodevelopmental modeling and analysis.