Guided synthesis of annotated lung CT images with pathologies using a multi-conditioned denoising diffusion...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Generative Models for Medical Data Synthesis
• Computational Pathology

Background:

• AI models for medical diagnostics require large, annotated datasets, which are scarce due to privacy and annotation costs.
• Existing generative models struggle with the precision and annotation fidelity needed for medical applications.

Purpose of the Study:

• To develop a controlled framework for generating synthetic, annotated medical images for AI training.
• To address the limitations of data scarcity and annotation overhead in medical AI development.

Main Methods:

• Utilized a denoising diffusion probabilistic model trained unconditionally on lung CT scans.
• Extended the model with classifier-free sampling for constrained, multi-conditional generation.
• Validated the framework by generating anatomically consistent lung CT images with specific pathologies.

Main Results:

• Generated diverse, annotated lung CT images with high anatomical fidelity and consistency.
• Successfully produced images with specific pathologies (e.g., lung nodules) at designated locations.
• Demonstrated superior performance compared to state-of-the-art generative models on comparable datasets.

Conclusions:

• The proposed framework effectively generates high-quality synthetic medical images with annotations.
• This method facilitates the creation of robust training datasets for AI diagnostic tools.
• The approach shows generalizability across medical imaging modalities and applications.