BrainSegFounder: Towards 3D foundation models for neuroimage segmentation.

Area of Science:

• Artificial intelligence in neuroscience
• Medical imaging analysis
• Foundation models for healthcare

Background:

• Brain health research increasingly uses AI for neuroimaging data analysis.
• Medical foundation models offer improved performance and sample efficiency.
• Neuroimage segmentation requires extensive labeled datasets for AI training.

Purpose of the Study:

• To develop a novel 3D medical foundation model for multimodal neuroimage segmentation.
• To reduce data requirements for AI model training in neuroimage segmentation.
• To enhance the accuracy and predictive capabilities of AI models in brain imaging.

Main Methods:

• A two-stage self-supervised pretraining approach using vision transformers.
• Stage 1: Encoding anatomical structures from large-scale unlabeled multimodal brain MRI data (41,400 participants).
• Stage 2: Identifying disease-specific attributes like tumor/lesion shapes and spatial locations.

Main Results:

• BrainSegFounder demonstrated significant performance gains on the BraTS and ATLAS v2.0 datasets.
• The model surpassed previous winning solutions that used fully supervised learning.
• Scaling model complexity and unlabeled healthy brain data volume enhanced accuracy.

Conclusions:

• The proposed dual-phase pretraining methodology significantly reduces data requirements for neuroimage segmentation.
• The findings highlight the impact of large-scale unlabeled data and model complexity on AI performance in brain imaging.
• BrainSegFounder offers a flexible and accurate solution for multimodal neuroimage segmentation adaptable to various imaging modalities.