Accurate focal‐plane selection is crucial for artificial intelligence assessment of three‐dimensional urine cytology specimens for bladder cancer screening and surveillance
Yoseph Sayegh, Brody McNutt, Minh‐Khang Le, I‐Chuang Liao, Keluo Yao, Camille Ng, Ahmad Kohsar, Daniel Shou, Michael Yu, YinXian Jin, Louis J. Vaickus, Xiaoying Liu, Christopher J. VandenBussche, Samuel E. Harvey, Joshua J. LevyAbstract
Background
Bladder cancer is a common and highly recurrent malignancy requiring lifelong surveillance. Urine cytology serves as a noninvasive triage tool to guide cystoscopy but is limited by variable sensitivity and manual review. Although deep learning enables quantitative cell‐level assessment, limited work has examined three‐dimensional urine cytology preparations (e.g., SurePath) containing residual cellular fragments, where restricting analysis to a single nominal focal plane may obscure diagnostically relevant features. This study aimed to quantify focal‐plane heterogeneity, measure degradation of nuclear‐to‐cytoplasmic (NC) ratio and nuclear area estimates off plane, and evaluate focal‐plane selection algorithms for performance recovery.
Methods
A total of 325 SurePath whole‐slide images scanned as 11‐plane Z‐stacks were analyzed that spanned negative through high‐grade urothelial carcinoma cases. A detection model identified cells and clusters across planes, and 343 clusters (2435 urothelial cells) were reannotated at the optimal nuclear and cytoplasmic focal depths. Classical focus metrics and vision‐transformer models were evaluated for focal‐plane prediction. A U‐Net segmentation model generated NC ratios and nuclear areas, and Spearman correlations compared annotated and predicted measurements across optimal, off‐plane, and algorithm‐selected conditions.
Results
Focal‐plane prediction accuracy ranged from 42% to 88%, with classical focus metrics outperforming deep‐learning approaches. NC ratio correlation was 0.774 at optimal focus and declined progressively off plane (∼0.50 at ±5 planes). Algorithm‐selected planes partially recovered performance (up to 0.748). Similar trends were observed for nuclear area estimation.
Conclusions
Accurate focal‐plane selection is critical for artificial intelligence–based assessment of three‐dimensional urine cytology. Future work will extend this analysis to cluster‐ and patient‐level outcomes in multi‐institutional validation studies.