Are Diffusion Models Effective Good Feature Extractors for MRI Discriminative Tasks?

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Diffusion models (DMs) are effective for 2D image tasks but their potential in 3D MRI is underexplored.
• Extracting meaningful features from 3D MRI data is crucial for discriminative tasks like disease diagnosis.

Purpose of the Study:

• To evaluate the effectiveness of diffusion models as feature extractors for 3D MRI discriminative tasks.
• To assess the performance of pretrained diffusion models on brain tumor classification and Alzheimer's disease diagnosis.

Main Methods:

• A denoising diffusion probabilistic model (DDPM) was pretrained on UK Biobank T1-weighted MR images (T1WIs).
• The DDPM was fine-tuned on BraTS2020 (brain tumor) and ADNI1 (Alzheimer's) datasets.
• A novel fusion module, CATS, was introduced to enhance U-Net representations from the DDPM.

Main Results:

• The DDPM generated high-quality synthetic images consistent with the raw data distribution.
• Features from middle blocks and smaller timesteps of the DDPM demonstrated high quality.
• The CATS module, with minimal additional parameters, achieved competitive classification scores on BraTS2020 (0.7704) and ADNI1 (0.9217).

Conclusions:

• Pretraining diffusion models on large-scale datasets and fine-tuning on smaller, task-specific datasets is a viable strategy for MRI data.
• Diffusion models, particularly when enhanced with the CATS module, excel as feature extractors for 3D MRI discriminative tasks beyond generation.