An important parameter that affects the quality of radiofrequency ablation lesions that are produced is the contact angle and contact orientation. Both are challenging to determine in vivo, and a method to classify that information and provide feedback in real time could potentially titrate the energy dose and increase the success rates of this treatment.
In our work, a grin lens-terminated single mode fiber was integrated into a commercial RFA catheter to allow for M-mode OCT imaging at the catheter tip. Ventricular wedges were dissected from four fresh swine hearts and submerged in whole blood. M-mode imaging was performed in four different orientations: non-contact, 30 degrees, 60 degrees, and 90 degrees.
One contact classifier with two sub-classifiers was developed to classify whether the catheter is in proper contact with the tissue and the angle of the catheter when it is in contact. This classifier is based on convolutional neural networks and used keras as developing framework with tensorflow backend.
We achieved 98.51% accuracy in the "contact" or "noncontact" classifier and 91.21% in the orientation classifier with 30 degrees, 60 degrees and 90 degrees as outputs. We successfully tested the contact quality classifier in real time and achieved high accuracy in 0.0053 seconds for a group of 20 A-lines. These results support the potential of having the guidance of catheter placement during the RFA therapy using OCT image and pre-trained classifiers. Future experiments will further test M-mode OCT and our processing algorithm within a larger sample size and demonstrate the utility in vivo.
|