A Deep-Learning Based Approach to Accelerate Groundtruth Generation for Biomarker Status Identification in Chromogenic Duplex Images

Satarupa Mukherjee, Qinle Ba, Jim Martin, Yao Nie

Keywords: Deep Learning, Biomarker Status, Accelerated Groundtruth Generation

TL;DR: Deep Learning Based Accelerated Groundtruth Generation

Abstract: Immunohistochemistry based companion diagnosis relies on the examination of single biomarkers for patient stratification. However, recent years have seen an increasing need to characterize the interactions among biomarkers in the tumor microenvironment. To this end, chromogenic multiplexing immunohistochemistry (mIHC) serves as a promising solution, which enables simultaneous detection of multiple biomarkers in the same tissue sections. To automate whole-slide scoring for mIHC, a crucial analysis step involves the identification of cell locations along with their biomarker staining status (presence/absence of positive staining signals), which we call biomarker status identification. However, developing algorithms for such analysis, especially deep-learning (DL) models, often requires manual labeling at the cell-level, which is time-consuming and resource-intensive. Here, we present a DL based method to accelerate groundtruth label generation for chromogenic duplex (tissue samples stained with two biomarkers) images. We first generated approximate cell labels and then developed a DL based interactive segmentation system to efficiently refine the cell labels. Our method avoided extensive manual labeling and reduced the time of label generation to 50%-25% of manual labeling, while achieving $<$5% error rate in pathologist review.