Computational tools for characterizing extracellular structures in H&E whole slide images.
Introduction
State-of-the-art hematoxylin and eosin (H&E)-stained tumor tissue biomarker quantifications primarily rely on detecting cell nuclei through manual annotation or artificial intelligence. Typical nucleus characterizations include metrics, such as mitotic index and nuclear density, and are usually computed only within a few representative regions of interest (ROI). We propose unique methods for quantifying biomarkers that characterize non-nuclear and multi-scale tissues and fully utilize the whole slide image (WSI).
Methods
We focus on developing characterizing factors in WSIs, especially those related to the non-nuclear tumor microenvironment (TME), towards enhancing diagnostic and prognostic values for clinical histopathologists. We have established a pipeline for segmenting nuclei in H&E-stained WSIs and injecting a nuclear background derived from stain deconvolution and local color space analysis. The resulting nuclei-free WSIs can be further analyzed by kernel-based computer vision and deep learning techniques to derive unstudied TME metrics.
Results
We present a high-throughput histopathology image processing pipeline, based on classical computer vision and artificial intelligence, that strives to maximize the utilization of WSIs. Results have been tested on WSI data from TCGA-LUAD and BRACS datasets. Non-reliance on ROI-based analysis has the potential to ease the burden on the histopathologist, while revealing clinically relevant information that might otherwise be missed.
Conclusion
A software toolkit is proposed for generating non-standard, multi-scale, and robust characterizations of H&E-stained WSIs, especially concerning non nucleus-derived metrics, is proposed. The diagnostic and prognostic value of such metrics pertaining to correlations with tumor proliferation rate and predicted patient outcome will continue to be studied.