A Coordinate-Regression-Based Deep Learning Model for Catheter Detection during Structural Heart Interventions

With a growing geriatric population estimated to triple by 2050, minimally invasive procedures that are image-guided becoming both more popular and necessary for treating variety of diseases. To lower the learning curve new procedures, it is develop better guidance systems methods analyze procedure performance. Since fluoroscopy remains primary mode visualizations, ability perform catheter tracking from fluoroscopic images an important part this endeavor. This paper explores use deep landmark detection in 3D-printed heart models. We show two-stage deep-convolutional-neural-network-based model architecture can provide improved performance initially locating region interest before determining coordinates tip within image. has average error less than 2% image resolution be performed 4 milliseconds, allowing its potential real-time intraprocedural tracking. Coordinate regression models have advantage directly outputting values used quantitative future applications easier create ground truth (~50× faster), as compared semantic segmentation require entire masks made. Therefore, we believe work long-term broader class cardiac devices, catheters, guidewires.