2D medical image synthesis using transformer-based denoising diffusion probabilistic model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Artificial intelligence (AI) methods are increasingly used in medical imaging research.
• A significant challenge is the limited size and scope of available training image datasets for AI model development.
• This limitation hinders the successful deployment of AI models in clinical settings.

Purpose of the Study:

• To introduce a novel medical image synthesis framework to address the challenge of limited training datasets for AI models.
• To evaluate the quality, authenticity, and diversity of synthetically generated medical images.
• To assess the utility of synthetic images in improving AI model performance for medical image classification tasks.

Main Methods:

• A 2D image synthesis framework based on a diffusion model utilizing a Swin-transformer-based network was developed.
• The model incorporates a forward Gaussian noise process and a reverse denoising process.
• Training data comprised four diverse medical image datasets: chest X-rays, heart MRI, pelvic CT, and abdomen CT. Evaluation involved visual Turing assessments by medical physicists and quantitative metrics (IS, FID, FDS, DS).

Main Results:

• Visual Turing assessments indicated a realistic appearance of synthetic images, with an average accuracy of 0.64.
• Quantitative evaluations yielded an Inception Score (IS) of 2.28, Fréchet Inception Distance (FID) of 37.27, Feature Diversity Score (FDS) of 0.20, and Diversity Score (DS) of 0.86.
• For COVID-19 classification, AI models trained on synthetic data achieved comparable (0.89) and mixed data (0.93) performance to models trained on real data (0.88).

Conclusions:

• The developed image synthesis framework successfully generates high-quality medical images across various modalities.
• This framework can effectively supplement existing training datasets, enhancing AI model development and deployment.
• The approach holds significant potential for advancing data-driven medical imaging research.