Generating 3D TOF-MRA volumes and segmentation labels using generative adversarial networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning for medical imaging requires large labeled datasets, which are challenging to obtain due to privacy and manual labeling efforts.
• 2D Generative Adversarial Networks (GANs) can generate shareable 2D data but miss volumetric information crucial for 3D medical imaging.
• Existing 3D GANs can generate 3D volumes but struggle with corresponding label synthesis due to computational constraints.

Purpose of the Study:

• To develop and evaluate 3D GANs capable of generating 3D medical image volumes with corresponding labels.
• To address computational limitations in 3D GANs by applying mixed precision techniques.
• To synthesize labeled 3D Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) patches for brain blood vessel segmentation.

Main Methods:

• Four variants of 3D Wasserstein GAN (WGAN) were implemented, including those with gradient penalty (GP), spectral normalization (SN), and mixed precision (SN-MP, c-SN-MP).
• Generated 3D TOF-MRA patches and their segmentation labels were quantitatively assessed using Fréchet Inception Distance (FID) and Precision and Recall of Distributions (PRD).
• 3D U-Nets were trained on synthetic data from WGAN models and compared against a benchmark U-Net trained on real data using Dice Similarity Coefficient (DSC) and balanced Average Hausdorff Distance (bAVD).

Main Results:

• 3D WGAN models utilizing mixed precision (SN-MP and c-SN-MP) achieved superior performance with lower FID scores and better PRD curves.
• A 3D U-Net trained on synthetic data generated by the c-SN-MP WGAN model demonstrated competitive segmentation performance for intracranial vessels (DSC: 0.841, bAVD: 0.508).
• The developed method effectively generates realistic 3D TOF-MRA patches and labels, highlighting the benefits of mixed precision for computational efficiency.

Conclusions:

• The proposed 3D GAN approach with mixed precision successfully generates realistic 3D medical image patches and corresponding labels for brain vessel segmentation.
• Mixed precision significantly enhances computational efficiency, leading to improved GAN performance and enabling the generation of high-quality synthetic data.
• This work facilitates the sharing of labeled 3D medical data, potentially improving the generalizability and clinical applicability of deep learning models.