Clinical Applications of Artificial Intelligence in Urologic Oncology

Sharif Hosein; Chanan R. Reitblat; Eugene B. Cone; Quoc-Dien Trinh


Curr Opin Urol. 2020;30(6):748-753. 

In This Article


In this review, we described novel applications of artificial intelligence in urologic oncology in imaging, histopathology, and genomics. Nevertheless, there are significant shortcomings in the studies mentioned above. Many of these models make use of only one data type (i.e. MRI scans). Artificial intelligence pipelines must be developed to operate with the input of the hundreds of data strings a physician classically incorporates in their trade. In practice, decisions cannot be made on treatment on the basis of a single type of study, and the same holds true for artificial intelligence. The most impactful studies presented were those that seamlessly combine multiple modalities of clinical information. Tabibu et al.'s[41] classifier, for example, most closely mimicked human intelligence by simultaneously interpreting the histology and genomic characteristics of the cancer to increase diagnostic power and improve treatment decision-making. Zhang's diagnostic machine is incredibly nuanced in that it goes through several iterations of feeding one output to a different algorithm. Their platform goes beyond analyzing an image and providing a diagnosis. They use natural language processing to provide physicians with easily interpreted characteristic features in addition to providing prognoses.[48] Natural language processing will likely be a pivotal contributor to integrating the infinite data sources in urologic oncology and has already shown promising results in providing prognoses by harvesting data from narratively written medical notes.[57]

Access to high-quality and representative data will also pose a challenge in making artificial intelligence readily accessible. In essence, an artificial intelligence machine is only as good as the data it was built on, and machines trained on stock images cannot be generalized to wider populations. Rare malignancies, like collecting duct RCC that only accounts for 1% of all renal cancers, will likely be a problematic front for artificial intelligence to cross.[58] Data used for clinical studies often do not encompass the prospective patient population. This is a poignant issue for African Americans and other minorities who are less likely to be represented in such medical data and suffer worse cancer outcomes.[59] Indeed, artificial intelligence used the wrong way may actually lead to exacerbating disparities. Future endeavors should focus on democratizing data and making intentional efforts at including a heterogeneous mix of patients.