Digital re-classification of equivocal dysplastic urothelial lesions using morphologic and immunohistologic analysis

A precise diagnostic of precursor dysplastic urothelial lesions is critical for patients but it can be a challenge for pathologists. Multiple immunohistologic markers (panel) improve ambiguous diagnostics but results are subjective, with a high degree of observational variability. Our research objective was to evaluate how a classification algorithm may help morphology diagnostic. Data coming from 45 unequivocal cases of flat urothelial lesions ("training set": 20 carcinomas in situ, 8 dysplastic and 17 reactive lesions) were used as ground truth in training a random tree classification algorithm. 50 "atypia of unknown significance" diagnostics (diagnostic set) were digitally re-classified based on morphological and immunohistochemical features as possible carcinoma in situ (20), dysplastic (17) and reactive atypia cases (13). The main sorting criterium was morphologic (nuclear area). A four-markers panel was used for a precise classification (74% correctly classified, 93% accuracy, 76% precision, averaged ROC=0.828). 3 cases were "false negative". The performance of the immunohistologic panel was evaluated based on a stain index, calculated for CD20, p53, Ki67 and observed for CD44. Within training set, the immunohistologic performance was high. In the diagnostic set both the percentage of high stain index for each marker and the percentage of cases with 2-3 strong markers were low, explaining the initial high number of equivocal cases. In conclusion, digital analysis of morphologic and immunohistologic features may bring clarification in classification of equivocal urothelial lesions. Computational pathology supports diagnostic process as it can measure features and handle data in a precise, reproducible and objective way. In our proof of concept study, a low number of cases and the (deliberate) absence of clinical data were main limitations. Validation of the method on a high number of cases, use of genomics and clinical data are essential for improving the reliability of machine learning classification