Real-time monitoring radiofrequency ablation using tree-based ensemble learning models.

Area of Science:

• Medical Engineering
• Oncology
• Artificial Intelligence

Background:

• Radiofrequency ablation (RFA) is a minimally-invasive cancer treatment that destroys malignant tumors using heat generated by alternating current.
• Real-time monitoring of RFA is crucial for treatment reliability but current methods are time-consuming and expensive.
• There is a need for faster, more cost-effective RFA monitoring techniques in clinical settings.

Purpose of the Study:

• To develop a machine learning (ML) approach to reduce RFA monitoring time while maintaining accuracy.
• To evaluate different ML algorithms and hardware setups for predicting ablation depth in three dimensions.

Main Methods:

• Two distinct hardware setups were employed to conduct RFA and simultaneously collect impedance data.
• Various ML algorithms, including random forest, adaptive boosting (Adaboost), and neural networks, were tested.
• The ML models predicted the 3D ablation depth based on the collected impedance data.

Main Results:

• Random forest and Adaboost models achieved over 98% R-squared accuracy on data from an embedded system-based hardware setup.
• These ML models outperformed neural network-based approaches in predicting ablation depth.
• The optimal combination of hardware and ML algorithm (Adaboost) demonstrated high accuracy.

Conclusions:

• An optimal hardware setup paired with the Adaboost ML algorithm can effectively control RFA by estimating lesion depth.
• The system achieved an average estimation error of 0.3mm for lesion depth.
• The estimation process was completed within 10ms, significantly reducing monitoring time for clinical application.