Re-classification of archival Ovarian Carcinoma diagnostics using immunohistologic digital quantification and algorithmic prognosis

Twenty years of research improved the classification of ovarian carcinoma, making the diagnostic relevant from a scientific and clinical perspective. Our research question was to find out if old studies are still pertinent under new diagnostic criteria and how we can use machine learning techniques for re-classification purposes. The same main investigator re-classified 60 cases of ovarian carcinoma after 15 years, using 2014 WHO diagnostic criteria. Selected pathology data only (macro, micro information and immunohistochemistry images coming from a seven-stain panel) were provided for digital analysis. Biomarker images were digitalized and quantified using open source software and a validated methodology. 1080 attributes were classified using a random forest (open source) algorithm, using a supervised learning technique (the training dataset used 180 attributes). Human results were considered "ground truth" for the digital analysis. The human analysis maintained the initial histopathologic diagnostic in 61.5% of cases. The digital prediction shows 80% accuracy and 73% precision when compared with human reclassified data. Based on results, we concluded that "recycling" of old studies is possible. Limitation of the study are the low number of cases analyzed, the total absence of clinical, treatment and prognostic data and a possible human criteria selection bias. Even if technical difficulties related to biomarker selection and histological analysis exist, digital investigation of existing, large archival registries is feasible, reliable and it can be done at a low cost.