The Developing Human Connectome Project: A fast deep learning-based pipeline for neonatal cortical surface...

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Medical Image Analysis

Background:

• The Developing Human Connectome Project (dHCP) studies perinatal brain development using structural MRI.
• Current automated pipelines for cortical surface reconstruction are time-consuming, hindering large-scale studies.

Purpose of the Study:

• To develop a significantly faster deep learning (DL) based pipeline for neonatal cortical surface reconstruction within the dHCP.
• To improve the efficiency and scalability of processing neonatal brain MRI data.

Main Methods:

• Implemented a DL pipeline incorporating DL-based brain extraction, surface reconstruction, and spherical projection.
• Utilized a multiscale deformation network for end-to-end diffeomorphic cortical surface reconstruction from T2-weighted MRI.
• Integrated a fast, unsupervised spherical mapping approach to minimize distortions.
• Employed GPU acceleration for cortical surface inflation and feature estimation.

Main Results:

• The DL-based pipeline processes a single MRI scan in 24 seconds, a nearly 1000-fold speed increase over the original dHCP pipeline.
• Cortical surface reconstruction quality was superior or equal to the original pipeline in 82.5% of test samples.
• Achieved high-quality cortical surface extraction for the dHCP neonatal dataset.

Conclusions:

• The proposed DL-based pipeline offers a highly efficient and accurate solution for neonatal cortical surface reconstruction.
• This accelerated method facilitates large-scale neuroimaging studies of early brain development.
• The pipeline demonstrates the potential of deep learning to revolutionize medical image processing workflows.