A Memory-efficient Deep Framework for Multi-Modal MRI-based Brain Tumor Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor segmentation from MRI is crucial for diagnosis and treatment.
• Current 3D U-Net models offer high accuracy but demand substantial GPU resources, limiting clinical adoption.

Purpose of the Study:

• To develop a memory-efficient deep learning framework for accurate brain tumor segmentation.
• To overcome the hardware limitations of existing 3D U-Net models for wider clinical accessibility.

Main Methods:

• Customized 2D U-Net architecture decomposing multi-label segmentation into sequential binary tasks.
• Extensive data augmentation and batch dice-loss function for enhanced generalization.
• Framework designed for low-memory GPUs (minimum 2GB RAM).

Main Results:

• Achieved near state-of-the-art performance on the BraTS 2020 dataset.
• Dice scores of 0.905 (whole tumor), 0.903 (tumor core), and 0.822 (enhancing tumor) on the testing set.
• Framework is executable on budget GPUs, demonstrating significant memory reduction.

Conclusions:

• The developed framework offers a viable, memory-efficient alternative for brain tumor segmentation.
• Reduced hardware requirements enhance clinical applicability, especially in resource-limited settings.
• The tool is planned for release as part of a free clinical analysis package with publicly available code.