Comparing 3D, 2.5D, and 2D Approaches to Brain Image Auto-Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep learning models for brain image auto-segmentation typically use 2D, 2.5D, or 3D approaches.
• The comparative performance of these dimensionality approaches for brain structure segmentation remains unclear.

Purpose of the Study:

• To compare the efficacy of 3D, 2.5D, and 2D deep learning approaches for brain image auto-segmentation.
• To evaluate segmentation accuracy, performance with limited data, memory requirements, and computational speed across different models.

Main Methods:

• Utilized 3430 multi-institutional brain MRIs for training and testing.
• Compared three auto-segmentation models: capsule networks, UNets, and nnUNets.
• Assessed performance using Dice scores, limited data scenarios, memory usage, and computational speed.

Main Results:

• 3D models consistently achieved higher Dice scores than 2.5D and 2D models across all evaluated architectures.
• 3D models demonstrated superior performance retention with reduced training dataset sizes.
• 3D models exhibited faster training convergence (20-40%) and deployment speed (30-50%) but required significantly more computational memory (20x).

Conclusions:

• 3D deep learning models represent the most accurate and robust approach for brain image auto-segmentation, particularly when training data is limited.
• Despite higher memory demands, the enhanced accuracy, speed, and data efficiency of 3D models justify their use in clinical and research settings.