Measurement Guidance in Diffusion Models: Insight from Medical Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Augmentation

Background:

• Acquiring medical samples faces significant challenges like cost, privacy, and radiation.
• Diffusion models show promise for image synthesis, but current methods primarily focus on general metrics (FID, IS).
• Existing guidance methods for diffusion models have limited impact on downstream task performance.

Purpose of the Study:

• To analyze the limitations of current guidance methods in diffusion models for medical image synthesis.
• To develop an uncertainty-guided diffusion model for generating high-quality synthetic medical data.
• To evaluate the practical utility of the proposed method in downstream medical applications.

Main Methods:

• Analysis of data distribution effects from previous guidance techniques.
• Introduction of uncertainty guidance at each sampling step within diffusion models.
• Extensive experimental validation on four medical datasets using ten classic networks.

Main Results:

• The uncertainty-guided diffusion model demonstrates practical contributions to downstream tasks.
• Experiments show improved performance on disease grading and diagnosis when trained on augmented datasets.
• Theoretical guarantees for general gradient guidance in diffusion models are provided.

Conclusions:

• Uncertainty guidance in diffusion models offers a significant improvement over existing methods for medical image synthesis.
• The developed model effectively generates synthetic data that benefits downstream medical applications.
• This work paves the way for more controllable and effective generative tasks in healthcare.