Fast-DDPM: Fast Denoising Diffusion Probabilistic Models for Medical Image-to-Image Generation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Denoising diffusion probabilistic models (DDPMs) show great promise in computer vision but are underused in medical imaging due to high computational costs.
• Extensive time steps (e.g., 1,000) in DDPMs lead to lengthy training (days/weeks) and sampling (minutes/hours) for medical images.

Purpose of the Study:

• To introduce Fast-DDPM, an efficient method to improve training speed, sampling speed, and generation quality in medical imaging.
• To address the computational challenges hindering the adoption of DDPMs in clinical applications.

Main Methods:

• Developed Fast-DDPM, a novel approach training and sampling with only 10 time steps.
• Introduced two efficient 10-step noise schedulers: uniform and non-uniform time step sampling.
• Optimized time-step utilization by aligning training and sampling procedures.

Main Results:

• Fast-DDPM achieved superior performance over standard DDPM and other state-of-the-art methods in medical image-to-image tasks.
• Demonstrated significant reductions in training time (0.2×) and sampling time (0.01×) compared to DDPM.
• Successfully applied Fast-DDPM to multi-image super-resolution, denoising, and image-to-image translation.

Conclusions:

• Fast-DDPM offers a computationally efficient and effective solution for medical image generation using diffusion models.
• The method holds potential for advancing disease diagnosis and treatment planning through faster, higher-quality medical imaging.