Missing MRI Pulse Sequence Synthesis Using Multi-Modal Generative Adversarial Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic resonance imaging (MRI) is crucial for disease assessment, diagnosis, and treatment planning.
• Acquiring multiple MRI sequences provides valuable, complementary information for physicians and automated systems.
• Challenges like long scan times, image corruption, and contrast agent allergies can limit the acquisition of necessary MRI sequences.

Purpose of the Study:

• To propose a generative adversarial network (GAN) capable of synthesizing missing MRI sequences.
• To leverage redundant information from available sequences to generate missing ones.
• To address the loss of complementary information caused by incomplete MRI scans.

Main Methods:

• Developed a multi-input, multi-output generative adversarial network (GAN).
• The GAN combines information from all available MR pulse sequences.
• Synthesizes one or more missing sequences in a single forward pass.

Main Results:

• Demonstrated and validated the method on two brain MRI datasets (four sequences each).
• Successfully synthesized all missing sequences in scenarios with one, two, or three sequences missing.
• Outperformed competing unimodal and multi-modal methods quantitatively and qualitatively.

Conclusions:

• The proposed GAN effectively generates missing MRI sequences, enhancing diagnostic capabilities.
• This approach mitigates challenges associated with incomplete MRI scans, improving data utility.
• The method offers a robust solution for reconstructing missing MR imaging data, benefiting both clinical practice and automated analysis.