Investigating Relationships Between Quantitative Nuclear Grading Features and Prognosis in Non-Muscle Invasive Bladder Cancer

Abstract

Histopathologic grading in non-muscle invasive bladder cancer (NMIBC) aims to predict how quickly the cancer will grow and spread, guiding disease management through its role in risk scoring. However, the qualitative nature of the current NMIBC grading system limits the ability to optimize its prognostic value, compromising the potential for data-driven care decision making. Using machine learning-based image analysis, we have quantified grading features such as the size, shape, and texture of tumour nuclei as well as number of cells undergoing division. We hypothesize that quantified nuclear features add significant prognostic value during the NMIBC grading and risk stratification process. Small (1.0 mm diameter) histopathology image samples and clinical event timelines were obtained for 163 patients with stage Ta NMIBC. Nuclear measurements of 22 grade-based histologic features were extracted from images using Visiopharm image analysis software. Cox Proportional Hazards (CPH) and Random Survival Forest (RSF) models were constructed to determine the added value of histologic features compared to clinical standard NMIBC grading and American Urological Association (AUA) risk scoring algorithm on survival. Data were split into train and test sets with cross validation on the training set. Time-dependent model performance was evaluated using concordance index (C-index) on the holdout test set. Mitotic index, mean lesser diameter (size, shape) and mean variance HEM (texture) carried the most prognostic value in both the CPH and the RSF models. Upon validation, the CPH model constructed using these QNFs achieved a C-index of 0.73 (95% CI: 0.56-0.88) compared to 0.55 (95% CI: 0.40-0.69) using diagnostic grade alone. The RSF model using these quantitative nuclear grading features achieved a C-index of 0.70 (95% CI: 0.66-0.70). Quantified nuclear grading features including size, shape, texture and mitotic index have the potential to add value and nuance to the standard AUA risk-scoring procedure. Our goal is to produce data-driven, pathology-informed decision support for NMIBC. These prognostic models will lay the groundwork for proper allocation of treatment and monitoring needed for each individual’s level of disease risk, while mitigating the current economic and resource burden on patient quality of life and the healthcare system.