Enhanced Magnetic Resonance Image Synthesis with Contrast-Aware Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic resonance imaging (MRI) exams require multiple pulse sequences for diagnosis.
• Generative deep learning models can synthesize MR images for various applications, including AI training.
• Current methods synthesize specific MR contrasts, limiting flexibility.

Purpose of the Study:

• To develop a method for generating synthetic MR images with adjustable contrast.
• To condition synthetic MR image generation on specific acquisition parameters.
• To evaluate the quality and utility of the generated synthetic MR images.

Main Methods:

• Trained a generative adversarial network (GAN) with an auxiliary classifier (AC) network.
• Conditioned image generation on repetition time, echo time, and image orientation.
• Evaluated AC performance in predicting acquisition parameters and assessed image quality via a visual Turing test.

Main Results:

• The AC accurately predicted repetition time (MAE 239.6 ms), echo time (MAE 1.6 ms), and image orientation (100% accuracy).
• Synthetic MR images demonstrated comparable quality to real images, with experts mislabeling 40.5% in a visual Turing test.
• The method allows for adjustable MR image contrast generation.

Conclusions:

• The developed GAN-AC model successfully generates high-quality synthetic MR images with controllable contrast.
• This technology can assist in MR sequence parameterization, radiology training, and customized AI dataset generation.
• The approach offers a flexible alternative to current generative methods for MR image synthesis.