DenseFormer-MoE: A Dense Transformer Foundation Model With Mixture of Experts for Multi-Task Brain Image Analysis
IEEE Transactions on Medical Imaging
|March 14, 2025
Summary
This study introduces DenseFormer-MoE, a novel deep learning foundation model for brain image analysis. It effectively diagnoses diseases and predicts brain age using T1-weighted MRI scans.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Diagnostics
Background:
- Deep learning models are crucial for brain image analysis but often task-specific.
- Developing a versatile foundation model for diverse neuroimaging tasks remains a challenge.
Purpose of the Study:
- To propose DenseFormer-MoE, a foundation model integrating dense convolutional networks, Vision Transformers, and Mixture of Experts for multi-task brain imaging analysis.
- To enhance feature representation generalization through Masked Autoencoder and self-supervised pre-training.
- To address multi-task learning optimization conflicts using a Mixture of Experts approach.
Main Methods:
- Integration of DenseNet and Vision Transformer for progressive local and global feature learning from sMRI.
- Masked Autoencoder and self-supervised learning for pre-training the foundation model.
- Mixture of Experts (MoE) for dynamic expert selection in multi-task learning.
Main Results:
- DenseFormer-MoE demonstrates promising performance in predicting brain age.
- The model achieves effective diagnosis of multiple brain diseases.
- Evaluated on UK Biobank, ADNI, and PPMI datasets, showing competitive results.
Conclusions:
- DenseFormer-MoE offers a robust foundation model for diverse brain imaging tasks.
- The proposed architecture effectively handles multi-task learning for neuroimaging.
- This approach advances automated analysis of structural MRI for clinical applications.
