Lung nodule synthesis guided by customized multi-confidence masks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate lung nodule generation is crucial for developing AI-assisted lung cancer diagnosis systems.
• Existing Generative Adversarial Networks (GANs) and Denoising Diffusion Probabilistic Models (DDPMs) have limitations in generating realistic and controllable lung nodules, particularly irregular shapes.
• GANs can produce artifacts and unnatural boundaries, while DDPMs lack precise control for segmentation tasks.

Purpose of the Study:

• To develop a unified generative model that overcomes the limitations of existing GANs and DDPMs for synthesizing controllable lung nodule images.
• To enable the generation of customized lung nodule images with specific shapes, sizes, and details guided by multi-confidence masks.
• To improve the quality and controllability of synthetic lung nodule data for medical imaging applications.

Main Methods:

• A unified model combining GAN and DDPM is proposed, comprising a Rough Lung Nodule Generator (GAN-based) and a Lung Nodule Optimizer (DDPM-based).
• The GAN component synthesizes rough lung nodules based on multi-confidence masks, allowing control over initial shape and size.
• The DDPM component refines these rough nodules to achieve more authentic boundaries and realistic details.

Main Results:

• The unified model achieved the best Fréchet Inception Distance (FID) score compared to existing methods.
• Generated synthetic lung nodules exhibit high quality and controllable features, addressing limitations of prior models.
• The synthetic data proved effective as a valuable supplement for training segmentation tasks.

Conclusions:

• The proposed unified GAN-DDPM model successfully generates high-quality, customized lung nodule images.
• This approach enhances the controllability and realism of synthetic medical data, addressing key limitations in current generative models.
• The model offers a promising solution for augmenting datasets in AI-driven lung cancer diagnosis and segmentation.