GANs for Medical Image Synthesis: An Empirical Study.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Generative adversarial networks (GANs) excel at producing photorealistic images.
• A key question is GANs' efficacy in generating usable medical imaging data.
• This study evaluates GANs across multiple medical imaging applications.

Purpose of the Study:

• To assess the benefits and limitations of various GAN architectures in medical imaging.
• To compare GAN performance across different modalities (cardiac cine-MRI, liver CT, retina RGB).
• To evaluate the utility of GAN-generated data for downstream tasks like image segmentation.

Main Methods:

• Tested diverse GAN models (DCGAN to style-based GANs) on cardiac cine-MRI, liver CT, and retina RGB datasets.
• Computed Fréchet Inception Distance (FID) scores to quantify visual realism of generated images.
• Assessed segmentation accuracy using a U-Net trained on GAN-generated versus original data.

Main Results:

• GAN performance varied significantly, with some architectures proving ill-suited for medical imaging.
• Top GANs generated visually realistic images (by FID) and could deceive experts in a visual Turing test.
• Despite visual realism, GAN-generated data did not fully capture the complexity of original medical datasets for segmentation tasks.

Conclusions:

• While GANs can produce high-fidelity medical images, their ability to replicate the full data richness for critical applications like segmentation remains limited.
• The choice of GAN architecture is crucial for successful medical imaging applications.
• Further research is needed to enhance GANs' capability in capturing complex medical data characteristics.