MAM-E: Mammographic Synthetic Image Generation with Diffusion Models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Data scarcity is a significant challenge in medical imaging research.
• Generative models, particularly diffusion models, show promise for data augmentation.
• Diffusion models offer high-quality image generation with less complex training than GANs.

Purpose of the Study:

• To explore diffusion models for synthesizing high-quality, full-field digital mammograms.
• To utilize stable diffusion models for inpainting synthetic mass-like lesions.
• To introduce MAM-E, a text-prompt-controlled pipeline for mammography synthesis and lesion generation.

Main Methods:

• Implementation of state-of-the-art conditional diffusion pipelines for mammogram generation.
• Application of stable diffusion models for synthetic lesion inpainting.
• Development of the MAM-E pipeline with text-prompt control and region specification.
• Quantitative and qualitative assessment of generated images.

Main Results:

• Successful generation of high-quality synthetic full-field digital mammograms.
• Effective inpainting of synthetic mass-like lesions on healthy mammograms using stable diffusion.
• Demonstration of MAM-E's capability for controlled mammography synthesis and lesion generation.
• Provision of assessment metrics and user-friendly graphical interfaces.

Conclusions:

• Diffusion models are a viable tool for medical image synthesis and data augmentation.
• The MAM-E pipeline provides a controllable method for generating realistic mammograms with synthetic lesions.
• Further research and implementation of these generative models can advance medical imaging analysis and training.