Real-Time Radiofrequency Ablation Lesion Depth Estimation Using Multi-frequency Impedance With a Deep Neural Network...

Area of Science:

• Medical Engineering
• Biomedical Signal Processing
• Machine Learning in Medicine

Background:

• Radiofrequency ablation (RFA) is a minimally invasive procedure requiring precise control of lesion depth.
• Accurate real-time estimation of RFA lesion depth is crucial for effective treatment and minimizing collateral damage.
• Current methods for RFA depth monitoring may lack the necessary precision or real-time capabilities for optimal clinical application.

Purpose of the Study:

• To design and optimize statistical models for estimating radiofrequency ablation (RFA) lesion depths with soft real-time performance.
• To evaluate the efficacy of deep neural networks (NN) and tree-based ensembles (TEs) for RFA lesion depth prediction using multi-frequency electrical impedance data.
• To enhance the accuracy of RFA lesion depth estimation by incorporating additional data features.

Main Methods:

• Utilized multi-frequency complex electrical impedance data from a low-cost embedded system.
• Trained deep neural network (NN) and tree-based ensemble (TE) models for regression-based RFA lesion depth estimation.
• Incorporated frequency sweep data, previous depth data, and previous RF power state data to improve model accuracy.

Main Results:

• The developed statistical models achieved significantly improved accuracy, with root mean square errors as low as 0.04 mm for TEs and 0.4 mm for NN.
• Simulation ablation performance showed a mean difference of 0.5 ±0.2 mm for the NN method and 0.7 ±0.4 mm for the TE method against physical measurements.
• Multi-frequency data was shown to significantly enhance the depth estimation performance of the statistical models.

Conclusions:

• Multi-frequency electrical impedance data substantially improves the accuracy of RFA lesion depth estimation models.
• The proposed RFA lesion depth estimation methods achieve millimeter-resolution accuracy.
• These methods demonstrate soft real-time performance on an embedded system, indicating potential for clinical RFA technology translation.