A Deep Learning Model for Tip Force Estimation on Steerable Catheters Via Learning-From-Simulation

Abstract

Atrial Fibrillation (AFib) is the most common arrhyth- mia among the elderly population, where electrical activity becomes chaotic, leading to blood clots and strokes. During Radio Frequency Ablation (RFA), the arrhythmogenic sites within the cardiac tissue are burned off to reduce the undesired pulsation. Manual catheters are used for most atrial ablations, however, robotic catheter intervention systems provide more precise map- ping. Several studies showed excessive contact forces (> 0.45 N) increase the incidence of tissue perforation, while inadequate force (< 0.1 N) results in ineffective ablation. Fig.1 shows a schematic of a cardiac RFA catheter used for AFib treatment. For robot-assisted RFA to be safe and effective, real-time force estimation of catheter’s tip is required. As a solution, finite element (FE) analysis can provide a useful tool to estimate the real-time tip contact force. In this work, a nonlinear planar FE model of a steerable catheter was first developed with parametric material properties in ANSYS software. After that, a series of simulations based on each mechanical property was performed, and the deformed shape of the catheter was recorded. Next, validation was conducted by comparing the results of the simulation with experimental results between the range of 0-0.45 N to determine the material properties. Despite the previous work, which was a study to estimate the tip contact force of a catheter using a deep convolutional neural network [1], [2], the main contribution of this study was proposing a synthetic data generation, so as to train a light deep learning (DL) architecture for tip force estimation according to the FE simulations. Due to the availability of real-time X-ray images during RFA procedures (fluoroscopy), the shape of the catheter is available intraoperatively. The proposed solution not only feeds the data-hungry methods based on DL with a sufficient amount of data, but also shows the feasibility of replacing the fast, accurate, and light-weight learning-based methods with slow simulations.