Lung-DDPM+: Efficient thoracic CT image synthesis using diffusion probabilistic model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Generative AI is crucial for synthetic data generation in medical diagnostics, particularly for lung cancer detection using CT scans.
• Current generative models for lung cancer diagnosis exhibit limitations in efficiency and anatomical accuracy, hindering clinical adoption.

Purpose of the Study:

• To introduce Lung-DDPM+, an optimized generative model for thoracic CT images with lung nodules.
• To address the efficiency and anatomical precision issues of previous generative models.

Main Methods:

• Lung-DDPM+ utilizes a denoising diffusion probabilistic model (DDPM) guided by nodule semantic layouts.
• The model incorporates a pulmonary DPM-solver for accelerated sampling and improved focus on lesion areas.

Main Results:

• Lung-DDPM+ demonstrates significant improvements: 8x fewer FLOPs, 6.8x lower GPU memory usage, and 14x faster sampling than Lung-DDPM.
• The model achieves comparable sample quality to existing SOTA models in segmentation tasks and is validated by radiologist assessment.
• Generated synthetic CT images with lung nodules exhibit high quality and fidelity.

Conclusions:

• Lung-DDPM+ effectively generates high-quality synthetic thoracic CT images with lung nodules.
• The model shows potential for broader applications in medical imaging, including general tumor and lesion synthesis.
• The improved efficiency and quality of Lung-DDPM+ enhance its clinical applicability in lung cancer diagnosis.