Lung-Area Restricted CT Generation via Conditional Score-Based Diffusion Models

Chest CT scans are vital for diagnosing lung abnormalities, yet their use in Deep Learning is limited by data scarcity, labeling costs, and privacy concerns. This work explores Score-based Diffusion Models for conditional CT slice generation restricted to the lung area. Two cases are considered: conditioning on lung masks, and on both lung and nodule masks. Custom U-Net architectures are trained under Variance Preserving (VP) and Variance Exploding (VE) stochastic differential equations to compare diffusion trajectories. VP SDEs achieve higher fidelity, with better FID, MMD, and SSIM than VE. Generated images follow conditioning closely, producing realistic lung and nodule structures for data augmentation. While current success is limited to 2D slices, future work targets 3D generation and richer anatomical conditioning. In essence, these results highlight the promise of conditional diffusion models for computer-aided diagnosis and Deep Learning in lung disease analysis.