Adaptive 3D Augmentation in StyleGAN2-ADA for High-Fidelity Lung Nodule Synthesis from Limited CT Volumes

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) require large datasets, posing challenges for volumetric medical imaging due to data scarcity.
• Existing methods struggle with limited volumetric data, impacting the quality of synthetic medical image generation.

Purpose of the Study:

• To extend Adaptive Discriminator Augmentation (ADA) for 3D StyleGAN2 to improve generative performance on limited volumetric medical data.
• To enable high-quality synthetic medical image generation from extremely limited datasets.

Main Methods:

• Adapted 2D operations for volumetric processing in 3D StyleGAN2, incorporating a full set of augmentation techniques.
• Evaluated two augmentation pipelines (color-based and comprehensive 3D augmentations) on the NoduleMNIST3D dataset (lung CT scans).
• Compared performance using Kernel Inception Distance (KID) and 3D Structural Similarity Index Measure (SSIM) against a baseline without augmentation.

Main Results:

• Volumetric ADA significantly improved training stability and reduced mode collapse risk, even with severe data constraints.
• A strong augmentation strategy enhanced the realism of generated 3D samples and better preserved anatomical structures.
• Adaptive 3D augmentations enabled high-quality synthetic image generation from limited volumetric datasets.

Conclusions:

• Adaptive 3D augmentations effectively address data scarcity in volumetric medical imaging.
• The proposed 3D StyleGAN2-ADA approach facilitates the generation of realistic and anatomically accurate synthetic medical images.
• This method opens possibilities for improved training of AI models in medical imaging with limited data.