3D whole brain segmentation using spatially localized atlas network tiles.

Area of Science:

• Medical Image Analysis
• Neuroimaging
• Deep Learning

Background:

• Detailed whole brain segmentation is crucial for quantitative analysis of structural magnetic resonance imaging (MRI).
• Current deep convolution neural network (CNN) methods face challenges including GPU memory limitations and insufficient training data.
• Existing 3D CNN approaches often underperform compared to multi-atlas segmentation (MAS) methods.

Purpose of the Study:

• To introduce the Spatially Localized Atlas Network Tiles (SLANT) method for high-resolution whole brain segmentation.
• To address limitations in spatial/contextual information learning and limited training data in CNN-based segmentation.
• To develop an efficient and accurate deep learning solution for detailed brain segmentation.

Main Methods:

• The SLANT method utilizes multiple independent 3D fully convolutional networks (FCNs), each learning contextual information for specific spatial locations.
• Auxiliary labels were generated for over 5000 unlabeled scans using MAS to augment the training dataset.
• A containerized pipeline was developed for deploying the integrated deep learning and traditional medical image processing solution.

Main Results:

• The proposed SLANT method demonstrated superior performance compared to state-of-the-art multi-atlas segmentation methods.
• Computational time for segmentation was significantly reduced from over 30 hours to 15 minutes.
• The method achieved high-resolution whole brain segmentation with improved accuracy.

Conclusions:

• The SLANT method offers a significant advancement in automated high-resolution whole brain segmentation.
• Integrating multiple FCNs and leveraging large-scale auxiliary data effectively addresses key challenges in deep learning for medical imaging.
• The open-source availability of SLANT facilitates further research and clinical application in neuroimaging analysis.