BDFM: Foundation Model for Segmentation and Classification Tasks of Brain Diseases

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Neurology

Background:

• AI-assisted diagnosis for brain diseases is limited by scarce annotated data and poor transferability of specialized models.
• Foundation models offer a promising avenue to overcome these challenges in medical AI.
• Developing robust models for brain disease analysis requires addressing data limitations and improving feature extraction.

Purpose of the Study:

• To develop a self-supervised foundation model for brain disease analysis.
• To enhance the model's ability to extract detailed lesion features.
• To improve the generalizability of AI models for brain disease diagnosis using limited annotated datasets.

Main Methods:

• A masked image modeling (MIM)-based foundation model for brain disease (BDFM) was established.
• A large-scale dataset (BD-15k) comprising over ten brain diseases was curated for pre-training.
• A spatial-frequency dual-domain decoder and a spatial mean masking strategy were introduced to improve lesion feature extraction.

Main Results:

• The proposed methods enhanced BDFM's performance in reconstructing intricate lesion details compared to baseline approaches.
• BDFM demonstrated excellent generalization capabilities across segmentation and classification tasks on small annotated datasets.
• Extensive experiments confirmed the model's effectiveness in downstream brain disease analysis tasks.

Conclusions:

• BDFM surpasses the performance of task-specific models trained from scratch, without requiring complex, specialized designs.
• This research advances the field of medical foundation models, enabling more effective brain disease analysis.
• The developed BDFM offers a powerful tool for AI-assisted diagnosis and analysis of brain pathologies.