Referenceless Proton Resonance Frequency Thermometry Using Deep Learning with Self-Attention.
Authors
Abstract
Proton resonance frequency (PRF) MR thermometry provides temperature feedback for MR-guided focused ultrasound (FUS), but conventional baseline-referenced thermometry is vulnerable to inter-scan motion and time-varying background phase, which can destabilize temperature monitoring. Referenceless approaches mitigate these issues by estimating the background phase from each frame; however, in transcranial applications they remain sensitive to susceptibility-induced phase discontinuities, often leading to reduced accuracy in heated regions and false-positive temperature elevation in surrounding tissue. This study developed and evaluated a deep learning-based referenceless PRF thermometry method that reconstructs background complex MR images to improve heated-region accuracy while maintaining stable background estimation. In this retrospective single-center study, 32 patients with essential tremor undergoing transcranial FUS (87 sonications, 1416 images) were included. Data from 28 patients were used for training and validation, and 4 patients (13 sonications, 146 images) formed the test set. A complex-valued self-attention-augmented U-Net was designed for background complex reconstruction. In heated regions, the proposed method achieved MAE = 0.64°C, Std = 0.80°C, and RMSE = 0.82°C, outperforming established referenceless techniques. The Dice coefficient of the 43°C isotherm reached 0.76, and background temperature change remained close to 0°C (MAE = 0.20°C), indicating strong suppression of spurious heating. Bland-Altman analysis showed limits of agreement from -1.37°C to +1.77°C with bias 0.20°C and R² = 0.99. These results demonstrate improved accuracy and stability of referenceless PRF thermometry for transcranial FUS monitoring, supporting more reliable intra-procedural temperature control, with potential applicability to motion-prone anatomies where baseline referenced thermometry is less reliable.