Synthesizing 3D Lung CT scans with Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Healthcare datasets are often private and limited in size, hindering the development of robust AI models.
• Patient data heterogeneity poses a significant challenge for machine learning in healthcare.
• Existing generative models often use 2D data, which is insufficient for complex 3D medical scans like CTs.

Purpose of the Study:

• To develop a 3D Progressive Growing Generative Adversarial Network (3D PGGAN) for synthesizing high-resolution (128^3) thoracic CT volumes.
• To address the limitations of data scarcity and 2D representations in medical imaging synthesis.
• To generate synthetic lung CT data that can improve the generalization capabilities of AI models in healthcare.

Main Methods:

• Development of a 3D Progressive Growing Generative Adversarial Network (3D PGGAN).
• Generation of synthetic thoracic CT volumes at a 128^3 resolution.
• Quantitative evaluation using 3D Multi-Scale Structural Similarity (3D MS-SSIM) and qualitative assessment via a Visual Turing Test.

Main Results:

• Successful generation of 3D synthetic lung CT scan volumes.
• Validation of synthesized data quality through quantitative metrics and human evaluation.
• Demonstration of the model's capability to synthesize entire lung volumes, not just specific pathologies.

Conclusions:

• The 3D PGGAN is a viable tool for synthesizing realistic 3D medical imaging data, specifically thoracic CT scans.
• Synthesized data can help mitigate data scarcity issues in healthcare, a major barrier to AI model generalization.
• This approach offers a promising direction for augmenting medical datasets and advancing AI in diagnostic imaging.