Machine Learning Basics

The advent of artificial intelligence (AI) has transformed how we live and work, including in health care. The applications of AI in medicine are myriad and have the potential to greatly enhance patient care. One form of AI that has shown particular promise is machine learning (ML). ML uses deep learning technology to detect and learn patterns in data and create models for inferring outcomes on its own. Thereby, ML algorithms are capable of mimicking human cognition and could be utilized by clinicians to enhance data interpretation and improve clinical decision making.

Generally, common goals of an AI study are to 1) validate the accuracy of the automated method compared to the “gold standard,” which usually relies on manual input by humans, and 2) assess the impact and clinical relevance of these outputs. When designing an AI validation study, it is important to denote the expected risks and benefits of the anticipated use of this AI method, and whether it is replacing or augmenting the human method. First, the model is trained on a large data set, with multiple iterations, to recognize patterns of interest (fig. 1). For example, in order to train a ML model to recognize different breeds of dogs, one might provide an input of 100 labeled images of Labradors, German Shepherds, and Poodles. Once the machine learns the spatial and geographic patterns associated with each of these breeds, it should be able to correctly identify a unique, unlabeled image of any of these dogs.

Figure 1. Illustration of study design process and major steps.

Potential Value of ML Application to Renal Stone Disease

Since 2009, successful studies of deep artificial neural networks have exploded, initially in the use of brain magnetic resonance imaging image analysis.¹ These advances have pushed both academic and corporate interest in the development of such algorithms in urology, such as the detection of urinary stones and kidney segmentation on computerized tomography scan.^1,2 Currently, manual assessment by radiologists and urologists is the gold standard for the identification and measurement of clinically important stone features and anatomy on noncontrast computerized tomography. However, human measurements are flawed and exhibit substantial interrater variability,^2,3 calling into question the validity of these dimensions in clinical decision making. This point is critical, as even small differences in stone size may alter a patient’s treatment course and may result in an inaccurate ascertainment of stone burden, inappropriate management strategies, and sub-optimal outcomes. Further complicating this picture is the fact that radiology reports commonly only include the largest or most clinically salient stone in a kidney and rarely report important features such as increased density of renal papillae that are associated with stone activity but are laborious to manually measure.⁴ Teaching a ML algorithm to recognize urinary stones and renal anatomy could reduce the uncertainty and inefficiency of human measurements while providing clinicians with standardized, complete information to apply toward medical decision making and patient counseling.

Current Applications of ML to Renal Stone Diseases

Toward that end and more, the NIDDK (National Institute of Diabetes and Digestive and Kidney Diseases)-funded Children’s Hospital of Philadelphia and University of Pennsylvania Center for Machine Learning in Urology (CMLU) was founded to apply ML to improve understanding of the pathophysiology, diagnosis, risk stratification, and prediction of treatment responses of benign urological disease among children and adults. Initial studies from the CMLU have demonstrated that these algorithms outperform humans in several ways.^2,3 ML utilized for imaging assessment carries out its analysis in terms of voxels, a computerized unit much like a 3-dimensional pixel. Early results from a ML study of adult and pediatric patients with symptomatic kidney and ureteral stones demonstrated that the use of a ML algorithm accurately detected and characterized kidney stones. Automated measurement of urinary stone size on noncontrast computerized tomography images was faster and more accurate than measurements by human raters in this study. As seen in figure 2, due to its ability to discern subtle changes in intensity at the minute voxel level, the ML algorithm captured stone borders with little to no error compared to humans.³

Figure 2. Manual segmentation results of 3 humans with red, green, and blue representing each individual and zoomed-in image as example of 1 stone (inset). Reprinted with permission of authors.³

Figure 3. Deep learning segmentation process for training data set in order to teach algorithm to identify left and right kidneys. Reprinted with permission of authors.³

Future Applications of ML to Urological Diseases

The use of ML in urology is rapidly evolving and is certainly not limited to applications within stone disease (fig. 3). Much work has been done already exploring applications of this technology within urologic oncology. For example, ML algorithms have been developed to detect upper urinary tract cancers,⁵ automate Gleason grading of prostate cancer,⁶ predict early cancer recurrence after prostatectomy,⁷ and predict 5-year survival after radical cystectomy.⁸ Within pediatric urology, ML has been successfully used to identify congenital abnormalities of the kidney and urinary tract via postnatal ultrasound imaging.⁹ Because ML models are able to learn and “think” without human input and provide informative, highly accurate outputs, the potential clinical applications of these programs may yet be much broader. There is likely still a long way to go before the true benefit of this technology is fully appreciated and integrated into clinical care.

One of the advantages of AI is that machines are capable of quickly processing very large amounts of data rapidly and without fatigue. The current state of the technology certainly allows for utilization of ML algorithms for large-scale research projects with sample sizes in the thousands, something that has previously required years to analyze by human labor with great interrater variability and error. In the case of stone disease, it may be possible to leverage this benefit to create models that integrate individual characteristics with anatomical features (eg stone size in 3 dimensions, volume, density, exact location, renal pelvis width, ureter diameter etc) within seconds, to predict clinically important questions such as “will the stone pass?”

ML is a very powerful and promising tool for the future of urological care. This technology is meant to augment and empower rather than replace human decision making. Furthermore, future prospective studies comparing these models to the gold standard are necessary.