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


As a proportion of total US doctors, pathologists have decreased from 2.03 to 1.43% between 2007 and 2017, while an aging population has caused a steady rise in cancer incidence.[29,30] As such, the potential of artificial intelligence to act as a force multiplier for the existing pathology workforce could prove vital to ensuring adequate pathologic service availability in the future.

Prostate Cancer

To diagnose prostate cancer, pathologists assess prostate needle core biopsies for parameters including the number of positive cores and the tumor extent in biopsies.[31] However, some biopsies harbor well differentiated or low-grade tumors that may require expert second opinions or advanced diagnostic measures like immunohistochemistry to be detected.[32,33] The recently developed artificial intelligence system, Paige Prostate Alpha, is a detection platform that predicts prostate cancer probability for whole slide images (WSI) based on feature analysis.[34] Paige Prostate Alpha can analyze WSIs and detect cancer independently, or if utilized by a pathologist, can mark areas with high probability for cancer. Raciti et al. recently assessed the clinical utility of Paige Prostate Alpha by observing pathologists' changes in sensitivity and specificity when assessing slides with and without artificial intelligence assistance. With artificial intelligence assistance, pathologist sensitivity significantly increased from 74 to 90% with no significant decrease in specificity.[34] For risk stratification of prostate cancer, pathologists assign Gleason scores to prostate biopsies based on microscopic morphology, which are the best indicator for progression.[35] Bulten et al.[36] developed a deep-learning system capable of grading prostate biopsies according to the Gleason standards to address the shortage of pathologist expertise. Their system was trained, tuned, and validated on 5759 biopsies from one center, and further validated on biopsies from an external center. The deep-learning model achieved a high agreement with reference standard Gleason scores set by three expert pathologists and outperformed 10 of 15 pathologist observers on the same internal validation set. The algorithm similarly achieved a high agreement on the external validation set with standards set by two independent pathologists. With extensive tuning on more comprehensive data sets, these systems would work well clinically to screen out benign biopsies and prioritize high-risk biopsies to the shrinking pool of expert pathologists.

Kidney Cancer

Percutaneous renal biopsies are recommended prior to ablating renal tumors to determine tumor histotype and tumor stage.[37] RCC tumor features and nuclear morphology can elucidate phenotypic information and can be of diagnostic value.[38,39] Holdbrook et al. designed an automated classifier that quantifies pleomorphic nuclear patterns for histopathologic slides from patients with suspected CCRCC. Their classifier successfully distinguished between high-grade and low-grade CCRCC tumors and correlated well with the prognosis of a retrospective patient cohort.[40] Tabibu et al. developed a pipeline capable of classifying RCC subtypes and identifying features that predict survival outcomes. Their algorithm distinguished CCRCC and chromophobe RCC from normal tissue with 93.39 and 87.34% accuracy. Tumor and nuclear morphological features were significantly associated with survival outcomes.[41] Artificial intelligence has also been used extensively to detect the extent of kidney disease in kidney samples. de Bel et al.,[42] for example, created a CNN capable of segmenting healthy and damaged glomeruli, tubules, and capillaries in WSIs. As algorithms can be tuned with the addition of datasets, this system could feasibly be altered to detect malignancies at a lower power. The slides used to assess for kidney disease could also be used to create more robust and representative training datasets.

Bladder Cancer

Urine cytology is a widely used noninvasive diagnostic test for detecting high-grade bladder tumors.[24] Urothelial cells that have been sloughed off may harbor abnormal features that may be suggestive of the underlying disorder. However, urine cytology has poor sensitivity for low-grade tumors (11.9%) and may be complicated by low cellular yield, comorbidities, and user interpretation, making it an unreliable screening test.[43,44] Sanghvi et al. integrated artificial intelligence into cytology by designing a deep-learning CNN to predict bladder cancer diagnosis from WSIs. They trained their system with images from 1464 patients and tested it on 790 retrospective patients. The CNN's overall sensitivity and specificity were 79.5 and 84.5%, and it detected 3 of 3 low grade urothelial neoplastic cells, all of which were missed by the pathologist.[45] Although these results are promising, recent studies have shown that other components to a urinalysis may be more effective at cancer diagnostics than cytology like a genomic profile.[46] This model also does not provide any insight into patient prognosis. Harmon et al. s deep-learning model was designed to predict lymph node involvement from primary bladder tumors. Their training and validation cohort combined 294 patients from The Cancer Genome Atlas and 13 patients from their center. A traditional multivariable logistic regression model for predicting lymph node status achieved an AUC of 0.678 in their test cohort, but AUC increased to 0.784 when cross-validated with the artificial intelligence score.[47] Zhang et al.[48] developed a groundbreaking artificial intelligence diagnostic model capable of providing pathologist-level interpretable diagnoses from WSIs. The model provided rich descriptions like 'severe crowding of nuclei', which pathologists can then use to provide a diagnosis of bladder cancer.