Volumetric Imitation Generative Adversarial Networks for Anatomical Human Body Modeling.

Area of Science:

• Medical imaging
• Computer vision
• 3D modeling

Background:

• Volumetric representation is crucial for 3D object visualization in medicine.
• Tomography images containing personal information limit data utilization.
• Existing GANs generate privacy-preserving medical images but lack inter-slice correlation and detailed internal structure representation.

Purpose of the Study:

• To propose Volumetric Imitation GAN (VI-GAN) for generating realistic 3D human anatomical models.
• To address limitations in existing methods regarding vertical slice correlations and internal feature accuracy.
• To create a model that captures external shape, internal slices, and their relationships.

Main Methods:

• Developed VI-GAN, a generative adversarial network incorporating a 3D U-Net and ResNet for feature extraction and up-sampling.
• Implemented a 3D-convolution-based local feature fusion block (LFFB) for enhanced feature integration.
• Utilized a 3D-convolution discriminator and devised reconstruction loss (feature and similarity) for model convergence.

Main Results:

• VI-GAN successfully generated volumetric data that realistically imitates human anatomical models.
• The model effectively captured external shape, internal slices, and crucial vertical slice relationships.
• Evaluated using CT data from 234 individuals, VI-GAN outperformed existing methods in generating accurate volumetric representations.

Conclusions:

• VI-GAN offers a novel approach to generating high-fidelity 3D anatomical models from medical imaging data.
• The method enhances 3D reconstruction accuracy by considering inter-slice correlations, crucial for medical applications.
• VI-GAN provides a valuable tool for research and development in medical imaging and 3D modeling while ensuring patient privacy.