Title: A generalizable model for histopathological spatial feature segmentation and classification by use of the SIVQ/VIPR algorithm as a spatial pre-processing stage

Biography:

David McClintock is an Associate CMIO of Michigan Medicine (Pathology Informat­ics), Director of Digital Pathology, Associate Director of Pathology Informatics and Associate Professor at the University of Michigan, USA. Ulysses G J Balis currently serves as the Director of the Division of Pathology Informatics at the University of Michigan, USA and is a member of the College of Fellows of the American Institute for Medical and Biological Engineering. Jerome Cheng is a Clinical Assistant Professor of Pathology in the Division of Pathology Informatics, in the Department of Pathology at University of Michigan, USA. He had his residency training in Anatomic and Clinical Pathology at SUNY Health Science Center in Brooklyn, New York and fellowship in Pathology Informatics at University of Michigan.

Abstract

The generation of ground truth maps from digital whole slide image subject matter remains a labour-intensive and technically challenging task, with the net effect being that the curation of large libraries of consistently and accurately tagged features remains a daunting prospect for most histology-based machine vision research teams.  To address this need, we have developed separate assisted and autonomous image segmentation computational pipelines, which both exhibit the property of converging on a highly generalizable classifier with a high area under the curve (AUC) solution, utilizing far fewer training image set tiles than typically required for deep learning schema. Initial histology feature segmentation and classification efforts with these tools have demonstrated compelling sensitivity and specificity, when compared against pixel-level ground truth maps, as generated by a panel of subject matter experts. Typical AUC values exceed 0.98 and the use of GPU-based computational clusters allows for segmentation activities to be carried out in real time, facilitating the discovery process. Interactive application demonstrations will be available as a component of the showcasing of this abstract.