Presentation Authors: Rohit Singla, Colin Lundeen, Connor Forbes*, David Hogarth, Christopher Nguan, Vancouver, Canada
Introduction: The efficacy of Extracorporeal Shockwave Lithotripsy (ESWL) is influenced by the time spent delivering focused energy to the stone. Manual targeting occurs at the start of the procedure, but subsequent respiration and movement significantly reduce the time that the stone is in the crosshairs. Radiographic appearance varies between stones and may change during treatment which makes targeting difficult. These effects result in increased radiation exposure and operative duration with increased shockwave exposure to surrounding structures in addition to decreased efficacy of stone fragmentation. There is a need for improved stone targeting to improve care. We propose a computer vision algorithm to locate stones during ESWL treatment.
Methods: 2413 fluoroscopic images from n=102 subjects that underwent ESWL were manually annotated followed by secondary review for annotation agreement (CL, CF, DH). A bounding box was drawn around any stone present. The algorithm RetinaNet was trained using a random split of n=90 subjects and tested on n=12. This was repeated for 10 unique splits. The mean Average Precision (AP) and stone detection time are reported.
Results: Over 10 trials, the mean (+/- stdev) AP was 0.7 +/- 0.1, indicating that in 1 of every 1.4 images the algorithm was able to locate the stone to within >= 50% of the annotation. Detection failure was attributed to small target size ( < 5% of image) or blurry image due to machine motion. The average (+/- stdev) detection time was 63 +/- 1ms.
Conclusions: An algorithm to automatically detect urinary tract stones during ESWL is presented, achieving ample precision to further develop an active targeting system. This work will be integrated into a broader artificial intelligence system for stone detection, automatic targeting and real-time in-procedure ESWL tracking for optimized outcomes.