Multicenter Validation for a Deep Learning Approach for Enhanced Urine Cytological Assessment and Rapid Bladder Cancer Screening - Summary Bladder cancer is the most recurrent form of cancer, making it the most expensive cancer to treat because of the need for frequent screenings, often with cystoscopy. Urine cytology, the microscopic examination of cells from urine samples, serves as a less invasive complement to cystoscopy; however, its cost-effectiveness and impact on disease management are limited by its lower reliability, efficiency, and sensitivity. While automation has heralded a new age for Pap smear screenings over the past two decades, urine cytology has seen considerably less innovation. To address this gap, our team developed AutoParisX (APX), a semi-automated screening tool that utilizes deep learning algorithms to analyze digitized urine cytology slides. APX provides both subjective (e.g., hyperchromasia) and objective (e.g., nuclear-to-cytoplasm ratio) measurements of atypia at single-cell resolution, generating a machine learning-based Atypia Burden Score (ABS) for each patient. ABS is predictive of malignancy and future recurrence and aligned with the Paris System for Reporting Urinary Cytology (TPS). Retrospective validation has shown an AUROC of 0.9 for patient diagnoses, setting a new benchmark. However, further refinement and validation on a heterogeneous, multicenter cohort are necessary for clinical implementation. To achieve this, we have assembled a coalition of academic institutions to contribute samples from varied demographics, specimen preparations, and whole-slide scanners. This will allow us to: 1) collect a large, heterogeneous set of urine cytology slides, 2) create the largest database of urine cytology annotations/images, and 3) validate the APX-enabled user interface for AI-assisted review through a comprehensive reader study. This will be accomplished in three non-interdependent aims. In Aim 1, we will leverage large-scale self-supervised learning (SSL) to create the APX foundation model—a generalizable base model built from over 625 million cellular images to reduce the data required to adapt APX to new institutions. In Aim 2, we will compare methods for generating Slide-Level diagnoses from APX’s granular Cell-Level information, including composite statistics (e.g., abnormal cell counts) and attention mechanisms that rank the most diagnostically relevant cells. In Aim 3, we will conduct a comprehensive, multicenter reader study of traditional glass slide screening versus APX-augmented digital slide screening using a screen-washout-rescreen protocol. The reader study will leverage extensive cytopathology expertise to achieve two key objectives: A) demonstrate that APX is non-inferior to traditional glass slides in terms of diagnostic sensitivity and specificity, and B) show that APX can suggest plausible diagnoses and deliver critical cellular information in a way that expedites the review process, making it more efficient. Usability testing will gather qualitative feedback on initial impressions, diagnostic confidence, ease of use, and ergonomics, to further enhance APX’s clinical utility. These efforts will lay the groundwork for a future clinical trial to demonstrate how digital technologies can streamline high-volume bladder cancer screening and surveillance, addressing the significant burden of recurrence.
