Robust deep MRI contrast synthesis using a prior-based and task-oriented 3D network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Magnetic Resonance Imaging (MRI) provides crucial diagnostic information through various contrasts.
• Acquiring multiple MRI contrasts increases scan time, cost, and patient discomfort.
• Current 2D synthesis methods for missing MRI contrasts suffer from 3D reconstruction artifacts.

Purpose of the Study:

• To develop a 3D deep learning model for synthesizing T2-weighted MRI volumes from T1-weighted images.
• To improve image quality and anatomical detail preservation in synthesized MRI contrasts.
• To enhance the robustness and generalizability of MRI contrast synthesis for clinical applications.

Main Methods:

• A 3D deep learning architecture was employed for T1-to-T2 weighted MRI volume synthesis.
• A novel loss function combining segmentation-oriented and frequency space information was utilized.
• Multi-atlas prior information and a semi-supervised learning framework were integrated for improved performance.
• The method was validated against state-of-the-art approaches, focusing on segmentation tasks.

Main Results:

• The proposed 3D synthesis method significantly improved image quality and anatomical detail.
• The segmentation-oriented and frequency space loss functions enhanced preservation of fine details.
• The integration of multi-atlas and semi-supervised learning improved model generalizability.
• The approach demonstrated superior performance compared to existing methods, especially in challenging segmentation scenarios.

Conclusions:

• The 3D deep learning approach offers an effective solution for synthesizing missing MRI contrasts.
• The novel loss functions and integration of prior knowledge enhance the accuracy and robustness of MRI synthesis.
• This method has the potential to improve clinical efficiency and diagnostic capabilities by reducing the need for multiple acquisitions.