sTBI-GAN: An adversarial learning approach for data synthesis on traumatic brain segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain segmentation in severe traumatic brain injury (sTBI) is crucial for clinical assessment and network analysis.
• Manual annotation of sTBI MRIs is challenging due to image deformations and lesion complexities, limiting model training.
• Existing data augmentation methods fail to synthesize realistic traumatic brain features, hindering segmentation performance.

Purpose of the Study:

• To develop a novel adversarial inpainting model, sTBI-GAN, for synthesizing realistic, labeled sTBI MR scans.
• To address the limitations of conventional data augmentation in capturing the unique characteristics of traumatic brain injuries.
• To improve the performance of automatic brain segmentation for sTBI patients.

Main Methods:

• Proposed sTBI-GAN model utilizes adversarial inpainting to generate both sTBI images and their corresponding segmentation labels simultaneously.
• Employed a coarse-to-fine inpainting strategy guided by edge information for image synthesis.
• Integrated a registration-based template augmentation pipeline to enhance data diversity and augmentation capacity.

Main Results:

• sTBI-GAN successfully synthesized high-quality labeled sTBI MR images.
• The generated synthetic data significantly improved the performance of both 2D and 3D traumatic brain segmentation models.
• Experimental results demonstrated superior segmentation accuracy compared to alternative methods.

Conclusions:

• sTBI-GAN offers a powerful solution for generating synthetic labeled sTBI MR data, overcoming annotation scarcity.
• The simultaneous generation of images and labels represents a significant advancement in medical image inpainting.
• This method holds promise for advancing automated analysis and understanding of severe traumatic brain injuries.