Denoising diffusion probabilistic models for 3D medical image generation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Diffusion probabilistic models have shown success in generating realistic images from text.
• Their application in 3D medical imaging (MRI, CT) remains underexplored.
• Synthetic medical data can aid privacy and augment limited datasets.

Purpose of the Study:

• To evaluate the capability of diffusion models in synthesizing high-quality 3D medical imaging data.
• To assess the utility of synthetic medical images for AI model pre-training and performance enhancement.

Main Methods:

• Utilized diffusion probabilistic models to generate synthetic magnetic resonance imaging (MRI) and computed tomography (CT) data.
• Conducted quantitative evaluation by radiologists assessing image realism, anatomical correctness, and inter-slice consistency.
• Employed synthetic images for self-supervised pre-training of breast segmentation models.

Main Results:

• Diffusion models successfully synthesized high-quality medical imaging data for MRI and CT.
• Radiologists confirmed the realistic appearance, anatomical correctness, and slice consistency of synthetic images.
• Synthetic data pre-training improved breast segmentation model performance, increasing Dice scores from 0.91 to 0.95.

Conclusions:

• Diffusion probabilistic models are effective for generating high-quality synthetic medical imaging data.
• Synthetic medical images can be valuable for privacy-preserving AI and augmenting small datasets.
• The use of synthetic data in pre-training demonstrably improves the performance of medical image analysis models, especially in data-scarce scenarios.