Quantitative histomorphometry of digital pathology as a companion diagnostic: Predicting outcome for ER+ breast cancers.

摘要

This work involves the creation of an image-based companion diagnostic framework that employs quantitative features extracted from whole-slide, H & E stained digital pathology (DP) images to distinguish patients based on disease outcome, with a clinical application aimed at distinguishing estrogen receptor-positive (ER+) breast cancer (BCa) patients with good and poor outcomes. Quantitative histomorphometry (QH) -- the conversion of a digitized histopathology slide into a series of quantitative measurements of tumor morphology -- is a rapidly growing field aimed at introducing advanced image analytics into the histopathological workflow. The thrust towards personalized medicine has led to the development of companion diagnostic tools that measure gene expression, yielding quantitative outcome predictions for improved disease stratification and customized therapies, e.g. Oncotype DX (Genomic Health, Inc.) for ER+ BCa. Yet, tumor morphology is often correlated with genomic assays, suggesting that genotypic variations in biologically distinct classes of tumors lead to distinct patterns of tumor cell morphology and tissue architecture in histopathology.;The application of this work to ER+ BCa is highly relevant to current clinical needs. Current treatment guidelines recommend that the majority of women with ER+ BCa receive chemotherapy in addition to hormonal therapy; yet, approximately half will not benefit from chemotherapy while still enduring its harmful side effects. Hence, there is a clear need for the development of automated prognostic tools to identify women with poorer outcomes who will likely benefit from chemotherapy.;The primary novel contributions of this work are (1) a color standardization system for improving the consistency in appearance of tissue structures across images, (2) the identification of tissue structures and corresponding QH signatures with prognostic value in ER+ BCa, (3) a multi-field-of-view framework for robust integration of prognostic information across whole-slide DP images, and (4) a method for predicting classifier performance for a large data cohort based on the availability of limited training data. This work will pave the way for the development of novel companion diagnostic systems capable of producing quantitative and reproducible image-based risk scores. These risk scores will play a vital role in decision support by helping clinicians predict patient outcome and prescribing appropriate therapies.