Fat-saturated image generation from multi-contrast MRIs using generative adversarial networks with Bloch...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Physics

Background:

• Acquiring multiple magnetic resonance (MR) image contrasts for spinal diagnosis is time-consuming and burdensome.
• Standard MR imaging protocols often require multiple sequences, increasing scan times and patient discomfort.

Purpose of the Study:

• To develop a novel deep learning model, BlochGAN, for synthesizing T2 fat saturation (T2 FS) spine images.
• To generate T2 FS images from readily available T1-weighted (T1-w) and T2-weighted (T2-w) images, leveraging fundamental MR physics.

Main Methods:

• Proposed Bloch equation-based autoencoder regularization generative adversarial network (BlochGAN).
• Utilized the Bloch equation to model the physical relationship between MR image contrasts.
• Employed an architecture with encoder, generator, discriminator, and decoder sub-networks for image synthesis and regularization.

Main Results:

• BlochGAN successfully generated high-quality T2 FS spine images from T1-w and T2-w inputs.
• The method demonstrated quantitatively and qualitatively superior performance compared to existing medical image synthesis techniques.
• The Bloch equation provided a physical basis for accurate contrast generation.

Conclusions:

• BlochGAN offers an efficient and effective approach to synthesize specific MR contrasts for spinal imaging.
• This method can potentially reduce scan times and improve diagnostic workflows in clinical settings.
• The integration of physics-based principles enhances the reliability of AI-driven medical image synthesis.