Kidney stones are characterized by the episodic occurrence of debilitating stone events, which lead to painful passage, emergent visits, and surgery. Proper selection of medical and surgical treatments depends on accurate assessment of stone characteristics, including size and location. Current methods for quantifying these characteristics depend on manual measurement by humans, which introduces unnecessary variation, is laborious, and makes analyzing the large number of imaging studies performed for clinical trials very difficult. Existing automated measurements are proprietary, only segment (partition) the stone from the surrounding structures without considering other clinically important features such as hydronephrosis, and are slow. A critical barrier to effectively implementing individualized therapies that decrease the burden of nephrolithiasis is the lack of automated analyses of diagnostic imaging that could accurately measure stone and kidney characteristics, and predict, in real time, an individual?s risk of stone events, such as spontaneous stone passage. In this Research Project, the Children?s Hospital of Philadelphia (CHOP) and the University of Pennsylvania (Penn) Center for Machine Learning in Urology (CMLU) forges a collaboration among experts in machine learning of diagnostic imaging, clinical epidemiology, and benign urologic disease. We build upon our recent discoveries that machine learning (particularly deep learning) of diagnostic images accurately, reliably, and rapidly predicts disease risk strata and outcomes. This project uses machine learning of CT to automate measurement of conventional characteristics of stones (e.g. size, location, and shape) and renal anatomy (e.g. hydronephrosis, ureteral dilation). We then apply this method to predict spontaneous passage of ureteral stones for individuals across the lifespan. In doing so, the proposed studies will develop clinically useful open-access prediction tools that will transform the standard of quantifying urinary stones and, in a fully automated way, accurately, reliably, and rapidly identify patients with ureteral stones most likely to benefit from early surgical intervention.
In Aim 1, we will use deep learning to automatically segment and measure conventional features of urinary stones (e.g. size, density) and adjacent renal and ureteral anatomy (e.g. degree of hydronephrosis) in CT images of 2,000 children and adults evaluated at CHOP and Penn, respectively.
In Aim 2, we will use deep learning to extract informative features from CT images that predict ureteral stone passage for 723 unique children and adults. The features include conventional features, engineered features, and deep-learning features that may neither be appreciated by nor be able to be measured by humans. These results would transform clinical care and research and provide insights into those who would be most likely to benefit from early elective surgery to remove stones to prevent future pain and emergent visits.
