Does Artificial Intelligence Meaningfully Enhance Cystoscopy?

Cystoscopy, a procedure developed in the 19th century to visualize the inside of the urinary bladder, is a critical step in the management of patients suspected of having bladder cancer (1). Over the years, advances in optics and instrumentation have improved the fidelity of visualization of the urothelial lining. More recently, enhanced cystoscopic techniques, such as narrow band imaging and blue light cystoscopy, have increased our ability to differentiate bladder tumors from adjacent normal tissue (2,3). Consequent prospective studies of these techniques have demonstrated improvements in cancer detection. Still, the experienced urologist’s eye, honed over thousands of procedures, is critical to interpret cystoscopic findings: multiple studies have shown that experience is independently associated with improved outcomes.

In this issue of the Journal, Wu and colleagues (4) present their study in which they used nearly 70 000 cystoscopic images from more than 10 000 consecutive patients to develop and train an artificial intelligence system to interpret cystoscopic images, emulating an experienced urologist. For study purposes, images were included from patients being evaluated for possible bladder cancer (eg, gross hematuria, imaging findings suspicious for tumor) or for surveillance after known bladder cancer. Each image of each lesion was classified as either “tumor” or “normal” depending on the final pathology from biopsy and confirmed with at least 6 months of follow-up. To interpret each lesion, the Cystoscopy Artificial Intelligence Diagnostic System (CAIDS) first normalizes the cystoscopic image to standardize the extraction of various features. Next, these features are analyzed using a pyramid pooling system that captures unique multiple levels, as well as a global assessment of the image. The images are processed within 50 milliseconds, and the system can be integrated with standard cystoscopic and endoscopic resection equipment used for diagnosis and treatment, respectively.

CAIDS performed exceptionally well in both the internal and external validation datasets. In the internal validation database, the sensitivity, specificity, and negative predictive value were higher than 0.975, and the positive predictive value was 0.819 (4). The study also compared CAIDS assessment with urologists’ assessment of 260 cases of carcinoma in situ and small tumors (defined as <40 × 40 pixels) compared with inflammatory changes. Urologists were categorized as trainees (2 years of experience), competent urologists (7 years of experience), or expert urologists (at least 20 years of experience). Despite the experience of the urologists in the study, CAIDS outperformed them in terms of sensitivity to detect carcinoma in situ and small tumors regardless of the urologists’ experience. Further, CAIDS enhanced the urologists’ ability to differentiate tumor from inflammatory changes.

Despite impressive diagnostic performance of CAIDS, several limitations are notable. First, the cohort was derived from 6 centers in China; therefore, the findings may not be generalizable to more ethnically and racially diverse populations. Second, more than 12% of the cohort was excluded because of inadequate cystoscopic images or positive follow-up but without suspicious lesions on cystoscopy. These types of “unideal” scenarios that commonly occur in the management of patients with bladder cancer are precisely when the experience of the urologist becomes critical to perform the procedure and interpret the results. Further, it is not specified whether comparisons of diagnostic accuracy between CAIDS and urologists used white light only or enhanced cystoscopy, which is now guideline recommended by the American Urological Association for use in certain situations if available (5).

Most important is the question of the clinical impact of CAIDS on the outcomes of patients with bladder cancer. Although the study clearly demonstrates an improvement in the accuracy and diagnostic capabilities of CAIDS over urologists of varying experience, this was not linked to the important endpoints in bladder cancer, namely, recurrence and progression. To assess the clinical utility of CAIDS, one must understand the various forms and the natural history of the disease. In truly low-risk, nonmuscle invasive bladder cancer, the focus is on minimizing recurrences, as these tumors generally lack the genomic drivers of more aggressive biology and rarely progress to life-threatening muscle-invasive disease. Therefore, CAIDS may identify a small low-risk lesion several months earlier than standard cystoscopy; however, this is unlikely to impact the patient’s ultimate quantity or quality of life. In fact, as bladder cancer is one of the most expensive cancers to treat per capita largely because of such invasive and lifelong follow-up, using enhanced cystoscopic techniques such as CAIDS may increase financial toxicity without clinical benefit (6). On the other hand, intermediate- and high-risk, nonmuscle invasive tumors carry substantial risk of progression to invasive disease, and identification of additional tumors may well capture those who need more aggressive therapy. Further, intermediate- and high-risk papillary tumors often coexist with carcinoma in situ, which is notoriously difficult to detect on white light cystoscopy. Similar to blue light cystoscopy, the cost of CAIDS may be most beneficial in these high-risk scenarios. Finally, identification of additional tumors may be warranted in some cases of muscle-invasive disease in patients who are planning to undergo so-called trimodal therapy that includes maximal tumor resection followed by chemotherapy and radiation.

In summary, the authors demonstrate robust diagnostic performance of CAIDS and the ability to potentially integrate into the workflow of urologists experienced in treating patients with bladder cancer. These results provide the impetus for further study, which should focus on identifying those subpopulations of patients with bladder cancer who may benefit the most from identification of additional lesions. Ultimately, prospective studies will determine the clinical utility of this novel technology.

Funding

No funding was used for this study.

Notes

Role of the funder: Not applicable.

Disclosures: The authors have no disclosures.

Author contributions: Writing, original draft: ATL, MSL. Writing, reviewing and editing: ATL, MSL.

Data Availability

Not applicable.