Accelerated proton resonance frequency-based magnetic resonance thermometry by optimized deep learning method.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Therapeutic Ultrasound

Background:

• Proton resonance frequency (PRF)-based magnetic resonance (MR) thermometry is vital for focused ultrasound (FUS) thermal ablation therapies.
• Accurate and rapid temperature feedback is essential for ensuring safety and effectiveness in clinical FUS treatments.

Purpose of the Study:

• To enhance temporal resolution in dynamic MR temperature map reconstructions using an advanced deep-learning method.
• To support real-time monitoring capabilities for effective FUS treatments.

Main Methods:

• Applied five neural network architectures (cascade net, complex-valued U-Net, shift window transformer, real-valued U-Net, ResUNet) for temperature map reconstruction from undersampled k-space data.
• Incorporated training optimizations: data augmentation, knowledge distillation, and a novel amplitude-phase decoupling loss function.
• Validated the approach using phantom, ex vivo tissue, and clinical uterine fibroid datasets.

Main Results:

• Achieved acceleration factors of 1.9 and 3.7 for 2× and 4× undersampling, respectively.
• The optimized ResUNet demonstrated superior performance with low Root Mean Square Error (RMSE) for temperature map reconstruction (e.g., 0.89°C for phantom data at 2× acceleration).
• High Dice coefficients (e.g., 0.81 for 43°C isotherm regions at 2× acceleration) and favorable Bland-Altman analysis confirmed accuracy.

Conclusions:

• Deep learning-based reconstruction significantly improves MR thermometry accuracy and efficiency for FUS treatments.
• The method shows clinical applicability, particularly for uterine fibroid treatments.
• Potential for extension to other MRI-guided FUS applications like essential tremor and prostate cancer treatment.