Nuclear membrane irregularity in high‐grade urothelial carcinoma cells can be measured by using circularity and solidity as morphometric shape definitions in digital image analysis of urinary tract cytology specimens

Abstract

Background The Paris System for Reporting Urine Cytology defines objective (elevated nuclear/cytoplasmic ratio ≥0.7) and subjective (nuclear membrane irregularity, hyperchromicity, and coarse chromatin) cytomorphologic criteria to identify conventional high‐grade urothelial carcinoma (HGUC) cells. Digital image analysis allows quantitative and objective measurement of these subjective criteria. In this study, digital image analysis was used to quantitate nuclear membrane irregularity in HGUC cells. Methods Whole‐slide images of HGUC urine specimens were acquired, and HGUC nuclei were manually annotated using the open‐source bioimage analysis software QuPath. Custom scripts were used to calculate nuclear morphometrics and perform downstream analysis. Results In total, 1395 HGUC cell nuclei were annotated across 24 HGUC specimens (48.1 ± 6.0 nuclei per case) using both pixel‐level and smooth annotation approaches. Nuclear membrane irregularity was estimated by calculating nuclear circularity and solidity. Annotating at pixel‐level resolution artifactually increases nuclear membrane perimeter, thus smoothing is necessary to better approximate a pathologist's assessment of nuclear membrane irregularity. After smoothing, nuclear circularity and solidity discriminate between HGUC cell nuclei with visually apparent differences in nuclear membrane irregularity. Conclusions Nuclear membrane irregularity defined by The Paris System for Reporting Urine Cytology is inherently subjective. This study identifies nuclear morphometrics that visually correlate with nuclear membrane irregularity. HGUC specimens show intercase variation in nuclear morphometrics, with some nuclei appearing remarkably regular while others show marked irregularity. A small population of irregular nuclei generates most of the intracase variation in nuclear morphometrics. These results highlight nuclear membrane irregularity as an important, but not definitive, cytomorphologic criterion in HGUC diagnosis.

Large-Scale Validation Study of an Improved Semi-Autonomous Urine Cytology Assessment Tool: AutoParis-X

“…The algorithm achieved an area under the receiver operating characteristic curve of 0.917 on the internal validation set, indicating a strong ability to distinguish between atypical and normal cells. Model interpretation using integrated gradients revealed that the algorithm placed a high emphasis on irregularities in the nuclear membrane as a key feature in determining cytological atypia ( Figure 2B ) 56 .…”

“…Statistical and machine learning models were implemented in Python and R 52–54 . A graphical overview is provided in Figure 1 : Slide processing– Connected components analysis to isolate individual cells and cell clusters, sped up through parallel processing 55 . Cell border detection (BorderDet)– Isolates cells within urothelial clusters with overlapping cytoplasmic borders through neural network detection model 44 . Cell-Level Measures: Morphometric measures– Additional morphological features to improve cell-type classification and atypia estimation (e.g., size / area). Urothelial Classifier (UroNet)– Used to filter urothelial cells from potentially conflated cell types through a convolutional neural network, which operates on images of cells and their morphometric measures 56 – trained on an expanded dataset with more cell classes. NC ratio estimation (UroSeg)– Estimates the NC ratio by neural network pixel-wise segmentation of background, nucleus and cytoplasm. Used as objective marker of atypia. Atypia score (AtyNet)– For predicted urothelial cells at a particular cutoff threshold, a subjective score which incorporates multiple screening criteria (e.g., hyperchromasia, etc.)…”

“…Cell-Level Measures: Morphometric measures– Additional morphological features to improve cell-type classification and atypia estimation (e.g., size / area). Urothelial Classifier (UroNet)– Used to filter urothelial cells from potentially conflated cell types through a convolutional neural network, which operates on images of cells and their morphometric measures 56 – trained on an expanded dataset with more cell classes. NC ratio estimation (UroSeg)– Estimates the NC ratio by neural network pixel-wise segmentation of background, nucleus and cytoplasm. Used as objective marker of atypia. Atypia score (AtyNet)– For predicted urothelial cells at a particular cutoff threshold, a subjective score which incorporates multiple screening criteria (e.g., hyperchromasia, etc.)…”

“…Urothelial Classifier (UroNet)– Used to filter urothelial cells from potentially conflated cell types through a convolutional neural network, which operates on images of cells and their morphometric measures 56 – trained on an expanded dataset with more cell classes.…”

“…Various morphometric features were estimated from individual candidate cells, including: 1) area; 2) convex area; 3) eccentricity; 4) equivalent diameter; 5) extent; 6) Feret’s diameter; 7) maximum diameter; 8) filled area; 9) major axis length; 10) minor axis length; 11) perimeter; and 12) solidity, extracted using scikit-image (Python v3.8) 56, 71 . These morphometric features were primarily used to help demarcate urothelial cells.…”