A Fast and Memory-Efficient Brain MRI Segmentation Framework for Clinical Applications.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Deep Learning

Background:

• Current brain MRI segmentation tools offer quantitative insights for neurological disorder diagnosis but are limited by high memory and time demands.
• Advanced 3D Convolutional Neural Network (CNN) methods, while achieving state-of-the-art results, necessitate high-memory Graphics Processing Units (GPUs).

Purpose of the Study:

• To develop a memory-efficient brain structure segmentation framework (FLBS) adaptable to dynamic memory constraints.
• To reduce the computational burden of multi-label brain segmentation by employing sequential single-label segmentations.
• To enable customizable brain structure segmentation for specific clinical applications.

Main Methods:

• Customized a memory-efficient framework (FLBS) based on nnU-Net architecture, dynamically adapting to memory limitations.
• Reduced multi-label segmentation to sequential single-label segmentations by extracting patches from T1w and segmentation maps.
• Utilized the MNI305 template for structure localization with a safety margin to decrease hardware usage.

Main Results:

• Achieved comparable accuracy to state-of-the-art methods on OASIS-3 and CoRR-BNU1 datasets.
• Demonstrated significant reduction in GPU requirements while maintaining comparable computational time.
• Validated generalizability on unseen datasets and confirmed feasibility on budget GPUs (minimum 4GB RAM).

Conclusions:

• FLBS offers a memory-efficient deep learning solution for brain MRI segmentation, significantly lowering hardware requirements.
• The framework maintains high accuracy and efficiency, making it suitable for widespread clinical adoption, especially in resource-limited settings.
• The publicly available code aims to facilitate broader access to advanced clinical brain imaging analysis tools.