MR-contrast-aware image-to-image translations with generative adversarial networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Magnetic resonance imaging (MRI) exams require multiple sequences for different image contrasts.
• Acquisition parameters influence image quality, contrast, and scan time.
• Time-consuming MRI acquisition and motion artifacts necessitate methods for image synthesis.

Purpose of the Study:

• To develop a method for synthesizing MR images with adjustable contrast properties.
• To address the need for efficient contrast generation in MRI.
• To overcome limitations of time-consuming MRI sequences and motion corruption.

Main Methods:

• Trained an image-to-image generative adversarial network (GAN).
• Conditioned the GAN on MR acquisition parameters: repetition time and echo time.
• Utilized a style transfer network approach where acquisition parameters define image 'style'.

Main Results:

• Successfully synthesized MR images with adjustable contrast.
• Outperformed the benchmark pix2pix approach on the fastMRI dataset.
• Achieved superior peak signal-to-noise ratio (24.48) and structural similarity (0.66) compared to the benchmark.

Conclusions:

• Developed the first model for fine-tuned MR contrast synthesis.
• Enables synthesis of missing MR contrasts and data augmentation for AI training.
• Applicable to other medical imaging tasks like intermodality translation and multi-field-strength synthesis.