SegSRGAN: Super-resolution and segmentation using generative adversarial networks - Application to neonatal brain MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Clinical neonatal brain MRI analysis faces challenges due to low anisotropic resolution.
• Current MRI analysis pipelines often separate image reconstruction and segmentation, potentially limiting accuracy.

Purpose of the Study:

• To propose a novel methodology and software for simultaneous high-resolution reconstruction and segmentation of neonatal brain MRI data.
• To address the limitations of separate image processing steps in MRI analysis.

Main Methods:

• Utilizes generative adversarial networks (GANs) for simultaneous super-resolution reconstruction and segmentation.
• Details the GAN architecture, implementation, and crucial technical aspects.
• Demonstrates the method's application in neonatal cortex analysis.

Main Results:

• Quantitative evaluation using Dice, PSNR, SSIM, and connected components.
• Qualitative and quantitative assessments on research (dHCP) and clinical (Epirmex) datasets.
• Highlights the method's ability to leverage data-augmentation strategies.

Conclusions:

• The proposed method and software show significant potential for practical medical applications.
• Provides a freely available tool for super-resolution reconstruction and segmentation of cerebral structures from arbitrary MRI datasets.