Introduction:

Low-grade noninvasive (LGTa) bladder cancer exhibits a heterogenous clinical course following initial diagnosis. Identification of clinical variables associated with tumor recurrence may help to risk stratify patients and develop personalized surveillance protocols. We aimed to utilize a complementary histological investigation along with artificial intelligence (AI) based histologic assay to predict LGTa tumor recurrence during follow up.

Methods:

After institutional review board approval, medical records were queried for patients with LGTa from 2015-2019, which allowed for at least 3 years of clinical follow up. Tumors were stratified by downstream recurrences (NR – no recurrences vs. R - recurrences). An experienced genitourinary pathologist (AV) reviewed H&E staining noting cytologic and architectural atypia including papillary fusion, inverted growth pattern, and percentage of non-tumor cell component of the sample including lamina propria and possible muscularis propria . Histologic features in recurrent vs. non-recurrent tumors were analyzed using Fisher’s Exact Test. To construct a morphological assay for recurrence risk stratification, a deep learning algorithm was used to extract quantitative features from segmented nuclei on high-resolution digitized whole slide images. These features were then correlated to recurrence free survival (RFS) utilizing a multivariable Cox proportional hazards (CPH) model and recurrence risk stratification was examined using Kaplan-Meier analysis and log-rank test.

Results:

Twenty-nine bladder tumors from 29 patients were identified, of which 17 (59%) subsequently recurred. The median (IQR) follow up was 59.0 (48.5-75.8) months, and the median (IQR) number of recurrences was 4 (2-10) in the R cohort. Table 1 outlines the different features identified on H&E for R vs. NR tumors. R tumors exhibited a higher frequency of inverted growth pattern (45% vs. 11%, p=0.04) and higher median stroma percentage (50% vs. 20%, p<0.01). The AI-based derived histologic assay was predictive of recurrence with AUROC of 0.81. Kaplan-Meier analysis demonstrated the model was able to risk stratify the cohort robustly with a statistically significant hazard ratio of 5.43 [95% CI 1.1, 26.76] (p=0.02) for prediction of high and low risk of recurrence (Figure 1). Patients in the high risk group had a 87.5% recurrence rate while those in the low risk group had a 28.5% recurrence rate.

Conclusion:

Utilizing a complimentary approach with histologic analysis and artificial intelligence, we were able to identify critical differences between recurrent and nonrecurrent low-grade noninvasive bladder tumors. These hypothesis-generating findings will guide future studies to identify resected tumors that are at high risk for recurrence, thereby allowing for surveillance protocol tailoring and improved patient stratification.

Funding: N/A