Predicting Response to Intravesical Bacillus Calmette-Guérin in High-Risk Nonmuscle-Invasive Bladder Cancer Using an Artificial Intelligence-Powered Pathology Assay: Development and Validation in an International 12-Center Cohort

Abstract

Purpose: There are few markers to identify those likely to recur or progress after treatment with intravesical bacillus Calmette-Guérin (BCG). We developed and validated artificial intelligence (AI)-based histologic assays that extract interpretable features from transurethral resection of bladder tumor digitized pathology images to predict risk of recurrence, progression, development of BCG-unresponsive disease, and cystectomy.

Materials and methods: Pre-BCG resection-derived whole-slide images and clinical data were obtained for high-risk nonmuscle-invasive bladder cancer cases treated with BCG from 12 centers and were analyzed through a segmentation and feature extraction pipeline. Features associated with clinical outcomes were defined and tested on independent development and validation cohorts. Cases were classified into high or low risk for recurrence, progression, BCG-unresponsive disease, and cystectomy.

Results: Nine hundred forty-four cases (development: 303, validation: 641, median follow-up: 36 months) representative of the intended use population were included (high-grade Ta: 34.1%, high-grade T1: 54.8%; carcinoma in situ only: 11.1%, any carcinoma in situ: 31.4%). In the validation cohort, "high recurrence risk" cases had inferior high-grade recurrence-free survival vs "low recurrence risk" cases (HR, 2.08, P < .0001). "High progression risk" patients had poorer progression-free survival (HR, 3.87, P < .001) and higher risk of cystectomy (HR, 3.35, P < .001) than "low progression risk" patients. Cases harboring the BCG-unresponsive disease signature had a shorter time to development of BCG-unresponsive disease than cases without the signature (HR, 2.31, P < .0001). AI assays provided predictive information beyond clinicopathologic factors.

Conclusions: We developed and validated AI-based histologic assays that identify high-risk nonmuscle-invasive bladder cancer cases at higher risk of recurrence, progression, BCG-unresponsive disease, and cystectomy, potentially aiding clinical decision making.

Keywords: artificial intelligence; bladder cancer; progression; recurrence.

Figure.

A, Recurrence risk (RR) model identifies patients at high (RR-High) and low (RR-Low) risk of high-grade recurrence (HR, 2.08; 95% CI, 1.80-2.40; P < .0001). High-grade recurrence-free survival (hgRFS) is censored at 2 years. B, The progression risk (PR) model identifies patients at high (PR-High) and low (PR-Low) risk of progression to muscle invasion (HR, 3.87; 95% CI, 2.75-5.44; P < .001). Progression-free survival (PFS) was censored at 5 years. TURBT indicates transurethral resection of bladder tumor.

Figure 1.

Allocation of an international, multi-institutional study population across developmental and validation cohorts. Nine hundred forty-four patients treated for high-risk (HR) nonmuscle-invasive bladder cancer (NMIBC) with transurethral resection of bladder tumor (TURBT) and intravesical bacillus Calmette-Guérin (BCG) across 12 institutions were allocated across 2 cohorts: development cohort (5 centers, 303 patients) and an international external validation cohort (7 centers, 641 patients).

Figure 2.

Multivariable Cox proportional hazards regression analysis to determine the strength of association of clinicopathologic factors with high-grade recurrence-free survival (hgRFS) and progression-free survival (PFS) in the development cohort and the relative contribution of these factors to survival risk vs the artificial intelligence model. A, Focality, specifically the presence of multiple tumors on initial transurethral resection of bladder tumor, was significantly associated with hgRFS (HR, 1.9; 95% CI, 1.21-3.1; P = .006). B, Presence of T1 stage disease at initial transurethral resection of bladder tumor was associated with PFS (HR, 4.22; 95% CI, 1.36-13.1; P = .01).

Figure 3.

Kaplan-Meier curves demonstrating stratification of patients in the validation cohort into high-risk and low-risk groups for key clinical outcomes by clinicopathologic risk scoring model or by artificial intelligence–enabled model. A, European Organization for Research and Treatment of Cancer (EORTC) 2016 risk does not stratify patients for recurrence risk (RR; baseline: EORTC 3; EORTC 4 HR 1.66, 95% CI 1.09-2.54; EORTC 5 HR 1.35, 95% CI 1.16-1.57; EORTC 6 HR 2.02, 95% CI 1.53-2.65; likelihood ratio P = .058). Model instability is suggested by the survival plot for EORTC group 4 crossing over that for group 5. The HRs for recurrence-free survival (RFS) for the ordinal EORTC groups were not monotonically increasing, suggesting a lack of trend. B, The RR model identifies patients at high (RR-high) and low (RR-low) risk of high-grade (hg) recurrence (HR 2.08, 95% CI 1.80-2.40; P < .0001). C, The European Association of Urology (EAU) nonmuscle-invasive bladder cancer 2021 risk scoring tool identifies patients at very high, high, and intermediate risk of progression (reference: EAU “high”; EAU “intermediate” HR 0.17, 95% CI 0.06-0.49; EAU “very high” HR 1.51, 95% CI 1.13-2.01; likelihood ratio P = .01). D, Comparatively, the progression risk (PR) model identifies patients at high (PR-high) and low (PR-low) risk of progression to muscle invasion with greater separation of risk groups (HR 3.87, 95% CI 2.75-5.44; P < .001). E, The PR model was able to identify patients at high and low risk of cystectomy (HR 3.35, 95% CI 2.51-4.47; P < .001). F, Presence of the bacillus Calmette-Guérin–unresponsive disease (BUD) histologic signature (BUD+) portends a shorter time to development of BUD than absence of the signature (BUD−) among cases receiving adequate bacillus Calmette-Guérin (HR 2.31, 95% CI 1.89-2.82; P < .0001). CFS indicates cystectomy-free survival; PFS, progression-free survival; TURBT,  transurethral resection of bladder tumor.

Figure 4.

Artificial intelligence model stratifies outcomes across patient subgroups in the validation cohort. A, Recurrence risk (RR) artificial intelligence model performance is demonstrated across different population subgroups with HRs of 1.73 to 3.16 between RR-high and RR-low groups. B, Progression risk (PR) artificial intelligence model performance is shown across population subgroups with HRs of 2.35 to 11.82. C, Risk of cystectomy as determined by the PR model maintains performance across subgroups with HRs of 1.99 to 12.22. D, Time to bacillus Calmette-Guérin–unresponsive disease (BUD) model performance is sustained across subgroups with HRs of 1.67 to 3.46. E, Application of the RR artificial intelligence model to the subgroup of patients treated with Food and Drug Administration–defined adequate bacillus Calmette-Guérin demonstrates continued discriminatory performance between RR-high and RR-low cases (HR 2.23, 95% CI 1.89-2.64; P < .0001). F, In the subgroup of patients treated with adequate bacillus Calmette-Guérin, PR-high cases showed significantly poorer progression-free survival (PFS) than PR-low cases (HR 3.24, 95% CI 2.19-4.79; P = .008). Note that the sample size reported for the adequate bacillus Calmette-Guérin cohort exceeds the number of cases reported in the external validation cohort for Table because cases that received adequate bacillus Calmette-Guérin and those that had Food and Drug Administration–defined BUD were included. cis indicates carcinoma in situ; hgRFS, high-grade recurrence-free survival; TURBT, transurethral resection of bladder tumor.

Figure 5.

Artificial intelligence models demonstrate predictive performance beyond and independently of clinicopathologic factors in the validation cohort. The recurrence risk (RR) artificial intelligence model identified cases at high risk of high-grade recurrence-free survival (hgRFS) independently of individual clinicopathologic features (HR 1.95, 95% CI 1.40-2.70; P < .001; A) and of composite measures including the European Organization for Research and Treatment of Cancer (EORTC) 2016 recurrence score (HR 2.0, 95% CI 1.51-2.70; P < .001; B). The progression risk (PR) artificial intelligence model identified patients at high risk of progression-free survival (PFS) independently of individual clinicopathologic risk factors (HR 3.24, 95% CI 1.62-6.50; P < .001; C) and composite measures including the European Association of Urology (EAU) 2021 risk scoring model for progression (HR 2.89, 95% CI 1.53-5.40; P < .001; D). cis indicates carcinoma in situ.

Figure 6.

Artificial intelligence–derived segmentation accurately identifies specific cell and tissue types to generate clinically relevant morphologic features. A, Area under the receiver operating characteristic curve of the cell segmentation model for different cell types (using pathologist annotation as the ground truth) derived from the development cohort. B, Correlations of different tissue area predictions by the model when a set of 23 slides derived from single-center data included in the validation cohort were scanned with 8 different scanner models. Whole-slide images were obtained using Leica Aperio AT2, CS2, and GT450 models (Leica, Weltzlar, Germany); 3D Histech Pannoramic 250 and 1000 models (3D Histech, Budapest, Hungary); and Hamamatsu S20 and S360 models (Hamamatsu Photonics, Shizuoka, Japan). C, Representative histologic image from a case in the validation cohort illustrating the tissue segmentation model on a transurethral resection of bladder tumor sample.