BIBSNet: A Deep Learning Baby Image Brain Segmentation Network for MRI Scans.

Area of Science:

• Neuroimaging
• Developmental Neuroscience
• Artificial Intelligence in Medicine

Background:

• Infant brain segmentation using MRI is crucial for understanding typical and atypical development.
• Existing algorithms struggle with the rapid changes in infant brains.
• Developing robust segmentation models is essential for pediatric neuroimaging research.

Purpose of the Study:

• Introduce BIBSNet (Baby and Infant Brain Segmentation Neural Network), an open-source deep learning model.
• Achieve robust and generalizable brain segmentation for infant MR images.
• Provide a faster and more accurate alternative to current segmentation methods.

Main Methods:

• Trained a deep neural network (BIBSNet) using a large dataset of manually annotated infant brain MR images (n=90, age 0-8 months).
• Utilized data augmentation and a combination of real and synthetic images from the BOBs repository and SynthSeg.
• Employed a 10-fold cross-validation procedure for model training and testing.
• Assessed performance using Dice Similarity Coefficient (DSC) and analysis of derived metrics (cortical thickness, connectivity, volumes) via the infant-ABCD-BIDS pipeline.

Main Results:

• BIBSNet segmentations demonstrated superior performance over Joint Label Fusion (JLF) across all brain regions based on DSC.
• BIBSNet outperformed JLF in nearly all derived metrics, including cortical thickness, resting state connectivity, and brain region volumes.
• The BIBSNet model is 600x faster than JLF.

Conclusions:

• BIBSNet offers a significant improvement for infant brain MRI segmentation across all analyzed age groups.
• The model is computationally efficient, produces FreeSurfer-compatible labels, and integrates easily into existing pipelines.
• BIBSNet presents a viable and advanced solution for segmenting the developing infant brain.