Self-Organizing Maps for Cellular In Silico Staining and Cell Substate Classification.

Robert West,Edwin Yuan,Sushama Varma,Korsuk Sirinukunwattana,Magdalena Matusiak,Łukasz Kidziński

doi:10.3389/fimmu.2021.765923

Abstract

Cellular composition and structural organization of cells in the tissue determine effective antitumor response and can predict patient outcome and therapy response. Here we present Seg-SOM, a method for dimensionality reduction of cell morphology in H&E-stained tissue images. Seg-SOM resolves cellular tissue heterogeneity and reveals complex tissue architecture. We leverage a self-organizing map (SOM) artificial neural network to group cells based on morphological features like shape and size. Seg-SOM allows for cell segmentation, systematic classification, and in silico cell labeling. We apply the Seg-SOM to a dataset of breast cancer progression images and find that clustering of SOM classes reveals groups of cells corresponding to fibroblasts, epithelial cells, and lymphocytes. We show that labeling the Lymphocyte SOM class on the breast tissue images accurately estimates lymphocytic infiltration. We further demonstrate how to use Seq-SOM in combination with non-negative matrix factorization to statistically describe the interaction of cell subtypes and use the interaction information as highly interpretable features for a histological classifier. Our work provides a framework for use of SOM in human pathology to resolve cellular composition of complex human tissues. We provide a python implementation and an easy-to-use docker deployment, enabling researchers to effortlessly featurize digitalized H&E-stained tissue.

Highlights

Cell organization in the tissue is deliberate [1,2,3] with specific cell types arranged into spatial structures driving tissue function both in health and disease, as well as patient outcome and therapy response in disease [2,3,4,5]
We chose a 7×7 hexagonal selforganizing map (SOM) grid composed of 49 total cell nuclei nodes, which we found to be sufficient for representing the nuclear heterogeneity present in the training set
We revealed four nuclear subtype interaction patterns associated with ductal carcinoma in situ (DCIS) lesions that are indicative of lower risk of progression from DCIS to invasive ductal carcinoma (IDC), and one epithelial/ lymphocyte interaction pattern that correlates with an increased risk of DCIS to be accompanied by IDC

Summary

Introduction

Cell organization in the tissue is deliberate [1,2,3] with specific cell types arranged into spatial structures driving tissue function both in health and disease, as well as patient outcome and therapy response in disease [2,3,4,5]. Hospitals around the world routinely digitalize histological tissue slides collecting vast amounts of image data. SOM for e-Pathology information from histological images is, the aim in the field of digital pathology. Automatic reading of digitized histological specimens is hindered by a complex nature of the images characterized by cellular and spatial heterogeneity, where both cell morphology and spatial distribution are parameterized by high dimensional space. In particular, deep learning models applied to histopathology provide the first evidence that automatic slide reading might be possible [6,7,8], the narrow focus of these models on specific diseases or cell types [9] as well as their black-box nature hinders their interpretability and widespread use [10]. Digital pathology needs more transparent models allowing for manual validation and broader application

Objectives

Methods

Results

Conclusion