Background and Objectives

Urolithiasis is a chronic condition characterized by episodic and often debilitating symptoms, including pain, at the time of stone passage. When patients initially present with a symptomatic ureteral stone there are 2 main treatment approaches—conservative treatment with a trial of passage or early surgical intervention with either stent placement or definitive stone removal. Though prior studies have attempted to identify factors that accurately predict the likelihood of spontaneous passage, such as stone size and location, it remains difficult to predict which patients will pass their stones vs those who will ultimately require surgery.¹

A trial of passage is ideal for those patients who will eventually pass their stones because it avoids the unnecessary risks of surgery and anesthesia as well as the potential pain and morbidity associated with stone procedures and ureteral stents.² However, in the subset of those who will ultimately fail a trial of passage and require surgical intervention, observation can lead to additional emergency room visits, missed days of school and work, and an extended period of severe symptoms. Classifying individuals at the time of diagnosis into one of these 2 groups will be beneficial for optimal treatment selection.

The goal of our group, the Children’s Hospital of Philadelphia and University of Pennsylvania Center for Machine Learning in Urology, is to harness the power of deep learning to improve diagnosis, management, and clinical outcomes of common benign urologic conditions, including stone disease, in pediatric and adult patients. Prior work by our group has shown that the automated measurement and characterization of kidney stones and renal anatomy using a deep learning model is feasible, accurate, and more efficient than manual measurements.³ Our long-term aim is to create a deep learning model that incorporates both patient clinical characteristics and CT imaging features to accurately predict the likelihood of ureteral stone passage for an individual patient’s stone leading to improved, individualized patient care.

Machine Learning of Imaging Data

The initial step in pursuit of this goal is developing a novel machine learning algorithm that segments the kidneys, ureters, and bladder on a CT scan in an automated fashion, which we have termed a “Urinary Tract Atlas.” The creation of this “Atlas” is necessary because of the difficulty in identifying and visualizing the ureters on noncontrast CT scans, which are the most common imaging modality used to evaluate ureteral stones clinically. The ureters can be difficult to accurately trace on noncontrast imaging for even experienced radiologists and urologists. Therefore, to train a machine learning model that incorporates imaging characteristics of ureteral stones, the model must first be taught to accurately localize the ureter.

We initially trained this model using CT urograms obtained from adult patients. Each CT urogram was first manually segmented by labeling the kidneys, ureters, and bladder. This involves tracing and outlining the relevant renal and ureteral anatomy on each slice of the CT scan. These manual segmentation labels serve as the ground truth for training as well as assessing the accuracy of the model. Figure 1 shows examples of kidney, ureter, and bladder labels generated using our atlas along with the original images and manual segmentations that served as ground truth. Similarly, Figure 2 shows a 3D rendering of the urinary tract as segmented using our model compared with manual segmentation. The overall Dice score for this model is currently 0.854, which is considered high, with a Dice score of 1 indicating perfect accuracy.⁴

Now that we have successfully trained a model to recognize the relevant anatomy on CT urogram, we are working on a model that will be able to automatically localize the expected path of the ureter on noncontrast CT scans. We have segmented the kidneys and bladder on noncontrast CT scans in the same manner described above and are currently in the process of training a model to recognize the kidneys and bladder on these images. We expect to be able to then train the model to predict the expected path of the ureter based on the known locations of the kidneys and bladder and ultimately to recognize ureteral stones using this information.

Machine Learning of Clinical Data

To train a model that can accurately predict ureteral stone passage, we are using clinical data from both adult and pediatric patients at our institutions who have presented with ureteral stones and are known to have either spontaneously passed their stones or required surgery. Using data from approximately 250 patients, we built a random forest model to predict spontaneous stone passage using patient characteristics that are contained in the electronic health record. The most important feature is stone size (area), but additional features are relevant depending on the adult or pediatric cohort, and include prior stone episodes, stone location, hydronephrosis, age, and presenting symptoms. Currently, these models are performing with 63%-70% accuracy, which we expect to improve with additional training data.

Future Directions: Combining Clinical and Imaging Data

Once the imaging model can accurately recognize the ureters and ureteral stones, it can be used to extract imaging features that can be combined with our clinical model described above. Deep learning has the advantage of being able to identify and utilize imaging characteristics not traditionally recognized or measured by clinicians. We expect automatically extracting features from CT images and the electronic health record will create a more accurate and less laborious method to predict stone passage than manual measurements and interpretation of imaging alone or the “gestalt” obtained from clinical experience. Ultimately, this model will allow the delivery of better individualized patient care by identifying those patients most likely to benefit from early surgical intervention for ureteral stones versus those likely to have a successful trial of passage.

Acknowledgements

We thank the entire Penn Center for Machine Learning in Urology team for their work on this project, including Joanie Garrat, Joey Logan, Ester Choi, Dawud Hamden, Alice Xiang, Iqra Nadeem, Brent Cao, Ryan Mcgregor, Curran Uppaluri, and Roby Daniel.

Funding

This study was funded by NIDDK P20 CHOP/Penn Center for Machine Learning in Urology (P20DK127488) and an AUA Care Foundation and SPU Sushil Lacy Research Scholar Award (to K. M. F.).