3D-DGGAN: A Data-Guided Generative Adversarial Network for High Fidelity in Medical Image Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Limited availability of 3D medical images hinders research in classification, segmentation, and detection.
• Existing 2D generative methods struggle with 3D anatomical structures, causing slice discontinuities.
• Current 3D generative networks require extensive data, which is often unavailable, leading to inadequate training and low-fidelity outputs.

Purpose of the Study:

• To propose a novel data-guided generative adversarial network (GAN) for high-fidelity 3D medical image generation.
• To address the challenge of generating realistic 3D medical images from limited datasets.
• To improve the accuracy and detail of AI-generated medical imaging data.

Main Methods:

• A data-guided GAN utilizes a generator that creates images from reference codes extracted from real data.
• The generator produces both noisy and noise-free decoded images to assess reference code fidelity against real images.
• A multi-component discriminator (volume, slab, and slice) is employed to enhance fidelity by analyzing images at different granularities.

Main Results:

• The proposed method successfully generates high-fidelity 3D medical images using a small amount of real training data.
• Quantitative comparisons using Fréchet Inception Distance (FID) and Maximum Mean Discrepancy (MMD) show superior performance over existing methods.
• The multi-component discriminator effectively differentiates real from generated images, improving overall realism.

Conclusions:

• The data-guided GAN offers a robust solution for generating high-fidelity 3D medical images, particularly in data-scarce scenarios.
• The proposed approach overcomes limitations of previous 2D and 3D generative methods.
• This advancement has significant implications for medical imaging research, enabling more robust model training and development.