Digitally derived Ki-67 proliferation index for GastroEnteroPancreatic neuroendocrine neoplasms

Ki-67 proliferation indices (PIs) define the grading of GastroEnteroPancreatic NeuroEndocrine Neoplasms (GEPNENs) and are crucial for therapeutic decisions. The precise Ki-67 assessment relies on manual counting, which is time-consuming, hardly accessible during routine pathological signout and thus usually replaced by the easier eye-estimation/balling method prone to interobserver variability and differences originating from the hot-spot size, localisation and tumor heterogeneity. These discrepancies can significantly affect the final PI resulting in misgrading of GEPNENs with potential adverse patient outcomes. In the era of digital pathology more and more applications are available to overcome this problem. In our retrospective study of 60 surgically resected GEPNEN cases, we tested the equivalence of traditional clinical (C) grading, manual counting with a MarkerCounter (MC) application and automatic grading with tumor recognition PatternQuant application with subsequent NuclearQuant (NQ) PI-assessment within 3DHistechs digital pathology platform. We found almost perfect agreement between the various grading methods (Spearman rank-order correlations: C vs. MC: ρ = 0.912, C vs. NQ: ρ = 0.883, MC vs NQ: ρ = 0.953) without clinically significant misgradings. Also the numerical values of the PIs derived with the various methods showed close correlations (Linear regression: C vs. MC: r = 0.952, C vs. NQ: r = 0.925, MC vs NQ: r = 0.978). The automated PI-assessment involved a mean 5-fold more tumor cells, better approximating the global/total Ki-67 PI, which was earlier shown to deliver more robust prognostic power and decreased interobserver variability. Furthermore, G3 tumors differed from G2 and G1 tumors in their cytomorphological parameterers: high grade tumors had significantly larger and more polymorphic, less regular tumor cell nuclei, which parameters could be also utilized for grading and/or prognostication purposes. Our study applied a simple, quick, easy-to-use, Machine Learning-based method that could be incorporated into routine digital pathology signout alleviating pathologists' workload and increasing precision and recall rate.