Lung-DDPM: Semantic Layout-Guided Diffusion Models for Thoracic CT Image Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• AI shows promise in lung cancer screening via medical imaging analysis.
• Data scarcity due to annotation costs and privacy concerns hinders AI development in healthcare.
• Existing generative models struggle with high-fidelity medical image synthesis.

Purpose of the Study:

• To develop an AI-driven method for generating high-fidelity 3D synthetic thoracic CT images for lung cancer screening.
• To address data scarcity challenges in medical imaging datasets.
• To improve downstream lung nodule segmentation tasks using synthetic data.

Main Methods:

• Proposed Lung-DDPM, a semantic layout-guided denoising diffusion probabilistic model (DDPM).
• Enabled generation of anatomically reasonable, seamless, and consistent synthetic CT images, even from incomplete layouts.
• Evaluated image quality and downstream performance on lung nodule segmentation tasks.

Main Results:

• Lung-DDPM outperformed state-of-the-art generative models in image quality metrics (FID, MMD, MSE).
• Achieved significant improvements in lung nodule segmentation (8.8% Dice, 18.6% sensitivity) when trained on combined real and synthetic data.
• Demonstrated superior performance with FID of 0.0047, MMD of 0.0070, and MSE of 0.0024.

Conclusions:

• Lung-DDPM effectively generates high-fidelity synthetic thoracic CT images, addressing data scarcity in lung cancer screening.
• The synthetic data significantly enhances the performance of lung nodule segmentation models.
• Lung-DDPM shows potential for broader applications in medical imaging, including tumor segmentation and risk prediction.