Referenceless Proton Resonance Frequency Thermometry Using Deep Learning with Self-Attention.

Area of Science:

• Medical Imaging
• Biophysics
• Artificial Intelligence

Background:

• Proton Resonance Frequency (PRF) MR thermometry is crucial for MR-guided focused ultrasound (FUS) temperature feedback.
• Conventional methods struggle with inter-scan motion and background phase variations, impacting transcranial FUS accuracy.
• Existing referenceless techniques are sensitive to susceptibility artifacts, affecting heated regions and causing false positives.

Purpose of the Study:

• To develop and evaluate a deep learning-based referenceless PRF thermometry method for transcranial FUS.
• To enhance accuracy in heated regions and ensure stable background estimation.
• To improve intra-procedural temperature monitoring reliability.

Main Methods:

• A retrospective study included 32 patients undergoing transcranial FUS for essential tremor.
• A complex-valued, self-attention-augmented U-Net was designed for background complex MR image reconstruction.
• The model was trained and validated on data from 28 patients, with testing on 4 patients.

Main Results:

• The deep learning method achieved Mean Absolute Error (MAE) of 0.64°C in heated regions, outperforming existing referenceless techniques.
• High accuracy was demonstrated with Dice coefficient of 0.76 for the 43°C isotherm and minimal background temperature change (MAE = 0.20°C).
• Bland-Altman analysis showed good agreement (limits of agreement -1.37°C to +1.77°C, R² = 0.99).

Conclusions:

• The developed deep learning referenceless PRF thermometry method significantly improves accuracy and stability for transcranial FUS.
• This technique offers more reliable intra-procedural temperature control, especially in motion-prone applications.
• The method shows potential for broader use in scenarios where baseline-referenced thermometry is less effective.