Conditional Score-based Diffusion Models for Lung CT Scans Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Chest CT scans are vital for diagnosing lung abnormalities like cancer.
• Limited data, high labeling costs, and privacy concerns hinder deep learning model training.

Purpose of the Study:

• To explore score-based diffusion models for conditional generation of lung CT scan slices.
• To address data scarcity challenges in medical imaging for AI development.

Main Methods:

• Utilized score-based diffusion models with custom U-Net architectures.
• Trained models to predict scores in Variance Preserving (VP) and Variance Exploding (VE) Stochastic Differential Equations (SDEs).
• Explored conditional generation using lung segmentation masks and lung/nodule segmentation mappings.

Main Results:

• The VP SDEs model demonstrated superior image generation quality (SSIM: 0.894, PSNR: 28.6).
• Achieved low scores in domain-specific metrics (FID: 173.4, MMD: 0.0133, ECS: 0.78).
• Generated images accurately reflected conditional guidance, producing realistic lung and nodule structures.

Conclusions:

• Score-based diffusion models show promise for data augmentation in medical imaging.
• The VP SDEs approach is effective for generating high-fidelity 2D lung CT slices.
• Future work includes 3D generation and richer conditional mappings for broader applications.