Automatic specific absorption rate (SAR) prediction for hyperthermia treatment planning using deep learning method

Area of Science:

• Biomedical Engineering
• Medical Physics
• Artificial Intelligence

Background:

• Accurate Specific Absorption Rate (SAR) prediction is crucial for effective hyperthermia treatment planning (HTP).
• Current simulation methods for SAR distribution are computationally intensive, limiting real-time treatment adjustments.
• Deep learning offers a potential solution for rapid and accurate SAR prediction.

Purpose of the Study:

• To develop a deep learning method for fast and accurate prediction of SAR distributions in the human head.
• To support real-time hyperthermia treatment planning (HTP) for brain cancer patients.

Main Methods:

• An encoder-decoder neural network with cross-attention blocks was proposed.
• The model predicts SAR maps using brain electrical properties, tumor coordinates, and antenna phase settings.
• A dataset of 201 finite-element modeling simulations was used for training and evaluation.

Main Results:

• The model achieved a mean RMSE of 3.3 W/kg and MAE of 1.6 W/kg across the brain.
• In target regions, RMSE and MAE were 4.8 W/kg and 2.5 W/kg, respectively.
• Computation time was reduced from 10 minutes to 4 seconds, with a mean SSIM of 0.90.

Conclusions:

• The deep learning method enables accurate and efficient SAR prediction for HTP.
• This approach has the potential to support real-time HTP, optimizing tumor temperature and improving clinical outcomes.
• This work introduces a novel deep learning approach that significantly accelerates SAR calculation for adaptive hyperthermia therapy strategies.