Assessing deep learning methods for the identification of kidney stones composition in endoscopic images

Current analysis of kidney stones through morphoconstitutional assessments makes it is possible to establish treatments to reduce the recurrence of kidney stones formation, but this process is seen as time-consuming and prone to errors by expert urologists. Thus, many practitioners have advocated for the introduction of automated AI-based visual identification methods to be deployed during the endoscopic exploration and stone extraction process. Such CADx tools could have a tremendous impact in the urologists workflow, providing immediate insights of the stone composition, and thus allowing timely hygienedietary advice after the operation. In this paper, we investigate the applicability of deep learning-based computer vision techniques for automatically classifying kidney stones for real-time support systems, attaining an average classification precision of 97% using Inception v3 in a challenging dataset comprised of images of four types of stones acquired in vivo.