Immunohistochemistry-Free Enhanced Histopathology of the Rat Spleen Using Deep Learning.

Abstract

Enhanced histopathology of the immune system uses a precise, compartment-specific, and semi-quantitative evaluation of lymphoid organs in toxicology studies. The assessment of lymphocyte populations in tissues is subject to sampling variability and limited distinctive cytologic features of lymphocyte subpopulations as seen with hematoxylin and eosin (H&E) staining. Although immunohistochemistry is necessary for definitive characterization of T- and B-cell compartments, routine toxicologic assessments are based solely on H&E slides. Here, a deep learning (DL) model was developed using normal rats to quantify relevant compartments of the spleen, including periarteriolar lymphoid sheaths, follicles, germinal centers, and marginal zones from H&E slides. Slides were scanned, destained, dual labeled with CD3 and CD79a chromogenic immunohistochemistry, and rescanned to generate exact co-registered images that served as the ground truth for training and validation. The DL model identified individual splenic compartments with high accuracy (97.8% Dice similarity coefficient) directly from H&E-stained tissue. The DL model was utilized to study the normal range of lymphoid compartment area and cellularity. Future implementation of our DL model and expanding this approach to other lymphoid tissues have the potential to improve accuracy and precision in enhanced histopathology evaluation of the immune system with concurrent gains in time efficiency for the pathologist.