Deep learning-based image analysis methods for brightfield-acquired multiplex immunohistochemistry images

• DOI: 10.1186/s13000-020-01003-0
• Publication Date: 2020-07-28T07:38:13Z
• Authors: Danielle J. Fassler; Shahira Abousamra; Rajarsi Gupta; Chao Chen; Maozheng Zhao; David Paredes; Syeda Areeha Batool; Beatrice S. Knudsen; Luisa Escobar-Hoyos; Kenneth R. Shroyer; Dimitris Samaras; Tahsin Kurc; Joel Saltz
• Keywords: Digital Pathology; Multiplex; Stain

Multiplex immunohistochemistry (mIHC) permits the labeling of six or more distinct cell types within a single histologic tissue section. The classification each type requires detection unique colored chromogens localized to cells expressing biomarkers interest. most comprehensive and reproducible method evaluate such slides is employ digital pathology image analysis pipelines whole-slide images (WSIs). Our suite deep learning tools quantitatively evaluates expression in mIHC WSIs. These methods address current lack readily available than four circumvent need for specialized instrumentation spectrally separate different colors. use case application our study that investigates tumor immune interactions pancreatic ductal adenocarcinoma (PDAC) with customized panel.Six were utilized label T-cells (CD3, CD4, CD8), B-cells (CD20), macrophages (CD16), (K17) formalin-fixed paraffin-embedded (FFPE) PDAC sections. We leveraged pathologist annotations develop complementary learning-based methods: (1) ColorAE autoencoder which segments stained objects based on color; (2) U-Net convolutional neural network (CNN) trained segment color, texture shape; ensemble both U-Net, collectively referred as (3) ColorAE:U-Net. assessed performance using: structural similarity DICE score segmentation results against traditional color deconvolution; F1 score, sensitivity, positive predictive value, predictions from ColorAE, ColorAE:U-Net pathologist-generated ground truth. then used prediction spatial (nearest neighbor).We observed comparable deconvolution single-stain IHC (note: cannot be mIHC); are detect 6 classes performance; combinations into outperform using either alone; (4) can employed detailed microenvironment (TME). developed scalable analyze distinctly labeled populations evaluated found they reliably detected classified microenvironment. also present case, wherein we apply across 3 WSIs quantify all perform nearest neighbor analysis. Thus, provide proof concept these describe distribution deployable clinical research studies.