Introduction:

Responses to intravesical bacillus Calmette-Guérin (BCG) therapy among patients with non-muscle invasive bladder cancer (NMIBC) remain heterogeneous. We developed and validated a histological assay using an artificial intelligence platform to analyze digitized whole slide image (WSI) histologic sections derived from pre-treatment transurethral resection of bladder tumor (TURBT) specimens to stratify risk of recurrence in high risk NMIBC patients treated with BCG. We evaluated the association of the histological assay stratification to recurrence free survival (RFS), BCG response and 12 month recurrence rates at a single institution.

Methods:

We conducted a retrospective study of BCG-treated high-risk NMIBC patients treated at a single institution (UTMB) from January 2014 to December 2021. Patients who did not receive adequate BCG therapy (5/7 doses within 12 months), could not be assessed for BCG-unresponsive disease (per FDA), and/or lack of follow-up at least 1 year following BCG completion were excluded. A board-certified genitourinary pathologist selected a representative H&E diagnostic slide for each patient obtained by TURBT. To construct the histological assay, a deep-learning algorithm first segmented nuclei from digital WSIs of the H&E specimens to extract quantitative histological features. These features were then correlated to RFS on the training set (33 patients) utilizing a multivariate Cox proportional hazards (CPH) model. RFS stratification was examined using Kaplan-Meier analysis and log-rank test on the test set (35 patients). Prognostic value of the histological assay was assessed using a multivariate CPH model with available clinical features.

Results:

A total of 68 patients were included with a median follow-up of 18 months. The “low risk” group as classified by the histological assay had superior RFS compared with the “high risk” group with a Hazard Ratio (HR) of 12.5, Confidence Interval (CI):1.56-100,log-rank: p=0.003). 10 of 20 patients classified as “high risk” had recurrence events during follow-up compared to 1 of 15 patients classified as “low risk” having recurrence events. Recurrence rates at 12 months were 43.7% in the high risk group and 0% in the low risk group. 8 of 9 BCG unresponsive patients in the test set were classified in the high risk group and one was classified in the low risk group. The histological assay was prognostic independent of CIS and T-stage of the initial TURBT (p<0.05). The sensitivity and specificity of the assay’s prediction of recurrence and BCG unresponsiveness were 90%, 58% and 88%, 45% respectively.

Conclusion:

While a small study, an artificial intelligence-based platform utilizing routine pre-treatment H&E stained histopathology specimens may further assist in the identification of patients with high-risk NMIBC most likely to recur following BCG therapy.

Funding: N/A