3D-StyleGAN2-ADA: Volumetric Synthesis of Realistic Prostate T2W MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Generating high-resolution 3D prostate T2-weighted (T2W) MRI is crucial for diagnosis and treatment planning.
• Existing generative models like 3D-StyleGAN struggle with stable training at clinically relevant resolutions for anisotropic volumes.

Purpose of the Study:

• To extend the StyleGAN2-ADA architecture for generating high-fidelity 3D prostate T2W MRI volumes.
• To evaluate the performance of the adapted model in terms of image quality, anatomical coherence, and radiomic feature preservation.

Main Methods:

• Adapted StyleGAN2-ADA to handle 3D anisotropic MRI data, enabling stable training at 256×256×24 resolution.
• Quantitatively assessed generative performance using Fréchet Inception Distance (FID), Kernel Inception Distance (KID), and generative Precision-Recall metrics.
• Evaluated synthetic data utility through radiomic feature analysis and downstream prostate segmentation tasks.

Main Results:

• Achieved stable training and high-resolution (256×256×24) 3D prostate MRI generation, overcoming limitations of baseline 3D-StyleGAN.
• Demonstrated significant improvements in quantitative metrics: FID decreased from 114.2 to 27.3, and generative Precision increased from 0.22 to 0.82.
• Synthetic data augmentation yielded prostate segmentation performance comparable to real-data training, with strong global radiomic feature alignment.

Conclusions:

• The 3D extension of StyleGAN2-ADA successfully enables stable, high-resolution volumetric prostate MRI synthesis.
• The generated synthetic MRI volumes preserve anatomically coherent structures and global radiomic characteristics.
• This approach offers a promising method for augmenting medical imaging datasets, potentially improving downstream clinical applications like segmentation.