Optimizing Z-Stack Plane Selection within Urothelial Clusters in 3D Urine Cytology via AI to Improve Bladder Cancer Screening and Surveillance: Ongoing Multicenter Study

Summary: Introduction: Bladder cancer recurrence is common and requires surveillance with urine cytology. Although The Paris System (TPS) has standardized diagnostic criteria, digital image-based urine screening is not widely adopted compared to screening technologies. This is partly due to variability in morphology, cell types, and imaging practices in urine cytology. Materials and Methods: AutoParis-X (APX) employs deep-learning to assess cells in urine cytology whole slide images (WSI) for features indicating malignancy. Cells detected within clusters are assigned a ""confidence score"" which determines suitability for single-cell analysis. APX was applied to 44 WSI from Johns Hopkins with z-stacking at 11 focal planes (1 um intervals, SurePath, Ventana DP-200 scanner). Random cell clusters were hand-annotated at the optimal focal plane followed by cell detection with APX at each focal plane. Cells identified correctly by APX were detected with high confidence and overlapped with manually annotated cells. Hierarchical Poisson regression was used to calculate the 'Focal Plane Difference' (FPD), which quantifies the absolute difference in depth of focus between annotated cells and those detected with high confidence by APX. We hypothesize that cells detected with the highest confidence will be found at the optimal focal plane. Results: 97.3% (95% CI: 95.3%-98.9%) of annotated cells were correctly identified by APX. 94.3% (95% CI: 91.8%-96.6%) of annotated cells had a predicted cell within 2 focal planes. The average FPD between detected and annotated cells was 1.57 um (95% CI: 1.47-1.67 um). FPD decreased with increasing confidence score: 1.83 um, 1.43 um, 1.24 um, and 1.12 um at confidence scores of 0.5, 0.75, 0.9, and 1.0, respectively. Conclusions: These findings indicate that assessing the quality of detection via APX can guide the selection of the best focal plane for evaluating cytomorphology. Accurate cytomorphology is essential for evaluating atypia, therefore digital-image-based urine screening methods should account for focal plane for accurate diagnostic utility.