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


Sequencing tumors has shed light on the metabolic activity of several cancers, further subclassified cancers, and linked seemingly distinct cancers to each other.[49] Interpreting genetics into clinically relevant information is complicated by the number of cancer classes and subtypes, epigenetics, transcriptomics, and proteomics.[50] Artificial intelligence can be used to recognize patterns in aberrant genetic expression that is beyond the capability of any single clinician.

Prostate Cancer

A machine learning model has been used to classify different Gleason-scored samples based on biomarkers. Hamzeh et al. s[51] machine identified PIAS3 as a biomarker for Gleason 4 + 3 groups, and UBE2V2 for Gleason 6, both of which are involved in the JAK/STAT signaling pathway. Uncovering the genetic hallmarks that underlie histologic patterns could help clinicians understand how to circumvent tumor immunity to certain treatments. Marin et al. identified a subset of prostate cancer genes that are correlative with recurrence. Their deep-learning network predicted the time of recurrence within a range of 3 months, with 96.9% accuracy.[52] Small nucleotide polymorphisms (SNPs) may modify healthy tissue's sensitivity to radiation therapy and may lead to adverse outcomes following treatment. Lee et al. sought to predict genitourinary toxicity in prostate cancer patients treated with radiation by correlating SNPs with clinical symptoms associated with toxicity. Their model provided a reliable genome-wide risk signature to 'weak stream'. The genes they identified are involved in neurogenesis and ion transport and are known to be important in urinary tract infections.[53]

Kidney Cancer

Penson et al. were able to predict tumor origin, including RCC, from circulating plasma DNA at the point of care by way of machine-learning. Using data from 7791 patients with 22 different cancer types, their model aligned cancer patients circulating plasma DNA with tumor DNA. In the independent test cohort, their algorithm identified patients with RCC from circulating DNA with 70 and 74.1% sensitivity and specificity.[54] This model has varying sensitivity and specificity for cancer types but could work well as a screening tool for patients with vague or inconclusive symptoms. Histologic studies would still be necessary for patients who are found to be RCC-positive, as the machine was not able to distinguish between RCC subtypes and may contribute to the overtreatment of indolent RCCs. In the previously mentioned RCC disorder classifier, Tabibu et al.[41] sought to provide a molecular underpinning to a histologic pattern. They further correlated their classifier with a multigene assay and identified variably expressed genes in regions with nuclear atypia, emphasizing the link between genetic biomarkers and RCC prognosis. Taken together, these studies may advance the use of artificial intelligence for RCC diagnosis and stratification.

Bladder Cancer

There are several well characterized risk factors associated with bladder cancer diagnosis, including smoking, age, and sex.[55] Predicting bladder cancer survival, however, has proven difficult. Poirion et al. created a pipeline that integrated multiomics data to stratify bladder cancer prognosis. By using The Cancer Genome Atlas, they trained an algorithm to recognize commonalities between high-risk and low-risk tumors. KRT6 and PI3K-Akt pathway were found to be upregulated in aggressive subtypes of bladder cancer.[56]