Color Deconvolution for Color-Agnostic and Cross-Modality Analysis of Immunohistochemistry Whole-Slide Images with Deep Learning

In digital pathology, deep learning algorithms are emerging for the automated analysis of bright-field immunohistochemistry (IHC) whole-slide images (WSIs). However, their performance falters when applied to images stained differently than the training set or acquired through multiplex immunofluorescence (mIF). Therefore, we propose an approach to build color-agnostic and cross-modality deep learning for IHC histopathology WSIs. Our approach leverages color deconvolution to extract single-chromogen intensity signals from RGB images and uses them as inputs for downstream deep learning models. We evaluate our method using lymphocyte detection in IHC as a use case and based our work on a training set of DAB-stained IHC slides. We trained a modified version of a published U-Net with intensity signals and compared it with an RGB-trained baseline on four datasets. First, we report comparable performance of our method versus using RGB data on the public LYON19 benchmark. Second, we show the method’s color-agnostic properties by applying it ‘out of the box’ to a colorectal cancer multiplex IHC dataset with red-stained lymphocytes. Finally, we report on the effectiveness of our method, trained with bright-field IHC data, on an in-house lung cancer mIF dataset and a publicly available multi-cancer mIF dataset.