Anatomically guided latent diffusion for high-resolution 3D chest CT synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate lung cancer analysis via chest CT is vital but limited by scarce, diverse 3D CT datasets for deep learning.
• Developing robust deep learning models for lung nodule analysis requires high-quality, well-annotated 3D CT data.

Purpose of the Study:

• To present LAND, an anatomically guided latent diffusion framework for synthesizing high-quality 3D chest CT volumes with lung nodules.
• To address the data scarcity issue in deep learning for lung cancer detection through efficient 3D medical image synthesis.

Main Methods:

• Developed LAND, a diffusion framework conditioning generation on 3D anatomical masks of lungs and nodules for spatial accuracy.
• Utilized a variational autoencoder (VAE) to encode masks into a latent space preserving nodule morphology.
• Incorporated conditional texture modeling for controlled lesion appearance variation.

Main Results:

• LAND generates 256x256x256 volumes at 1 mm isotropic resolution with reduced computational needs (10-16 GB GPU memory during training).
• Synthesized CT scans exhibit high visual fidelity and anatomical realism.
• Downstream lung nodule segmentation and classification tasks showed improved performance using LAND-generated data.

Conclusions:

• LAND offers a practical and efficient framework for anatomically guided 3D medical image synthesis.
• The method facilitates effective data augmentation for deep learning models in lung cancer analysis.
• LAND demonstrates potential for improving early lung cancer detection and risk stratification.