Deep-learned short tau inversion recovery imaging using multi-contrast MR images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic Resonance Imaging (MRI) is crucial for diagnosing various conditions.
• Short Inversion Recovery (STIR) sequences are valuable for visualizing edema and inflammation but require dedicated scanning time.
• Generating STIR images from existing multi-contrast sequences could save time and resources.

Purpose of the Study:

• To develop and validate a deep neural network (DNN) for synthesizing STIR MRI images from standard T1-weighted, T2-weighted, and Gradient Recalled Echo (GRE) images.
• To eliminate the need for additional scanning time to acquire STIR sequences.
• To provide a potential alternative for STIR imaging when acquisition is limited or artifacts are present.

Main Methods:

• A contrast-conversion deep neural network (CC-DNN) was designed as an end-to-end architecture.
• The CC-DNN was trained using multi-contrast simulation and in-vivo knee MRI datasets from 12 healthy volunteers.
• A novel loss function was implemented to address intensity differences, misregistration, and local variations. Quantitative metrics (MSE, PSNR, SSIM, MS-SSIM) and subjective radiologist evaluations were used.

Main Results:

• The CC-DNN successfully generated STIR images from T1-w, T2-w, and GRE sequences.
• The synthesized STIR images demonstrated superior performance in all quantitative evaluation metrics compared to existing methods.
• Musculoskeletal radiologists rated the CC-DNN-generated STIR images highest in subjective evaluations.

Conclusions:

• The proposed deep learning method is feasible for generating STIR MRI sequences without additional scans.
• This approach offers a practical alternative to conventional STIR pulse sequences, especially in time-constrained situations or when STIR artifacts are problematic.
• The study highlights the potential of AI in optimizing MRI workflows and enhancing diagnostic capabilities.