CirnetamorNet: An ultrasonic temperature measurement network for microwave hyperthermia based on deep learning

Area of Science:

• Biomedical Engineering
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• Accurate noninvasive temperature monitoring is crucial for effective microwave thermotherapy in cancer treatment.
• Current thermometry techniques face challenges in precision and reliability.
• Integrating multi-feature data offers a promising avenue for enhanced temperature prediction.

Purpose of the Study:

• To develop an advanced noninvasive thermometry technique for microwave thermotherapy.
• To achieve accurate temperature prediction by efficiently integrating multi-feature data.
• To enhance the reliability of temperature monitoring during cancer treatment.

Main Methods:

• Proposed an enhanced recurrent neural network architecture, CirnetamorNet.
• Utilized a simulated human tissue model for experimental data acquisition.
• Extracted 5 temperature-correlated parameters from ultrasonic image data using gray scale covariance matrix and Homodyned-K distribution.
• Employed a multi-head attention mechanism for temperature prediction based on multi-feature inputs.

Main Results:

• CirnetamorNet demonstrated superior performance compared to common models.
• Achieved low training losses (1.4589) and mean square error (0.1856).
• Attained a temperature prediction accuracy of 0.3 °C, outperforming many advanced models.
• Ablation experiments confirmed the critical contribution of each model module to overall performance.

Conclusions:

• CirnetamorNet shows exceptional performance for noninvasive thermometry in microwave thermotherapy.
• Presents a novel approach for multi-feature data fusion in medical applications.
• Holds significant practical value for improving cancer treatment outcomes.