Deep neural network based histological scoring of lung fibrosis and inflammation in the mouse model system

• DOI: 10.1371/journal.pone.0202708
• Publication Date: 2018-08-23T13:48:30Z
• Authors: Fabian Heinemann; Gerald Birk; Tanja Schoenberger; Birgit Stierstorfer
• Keywords: Nicotiana; 0301 basic medicine; Science; Pulmonary Fibrosis; Q; R; Pneumonia; Severity of Illness Index; Bleomycin; Disease Models, Animal; Mice; 03 medical and health sciences; Deep Learning; Smoke; Image Interpretation, Computer-Assisted; Medicine; Animals; Neural Networks, Computer; Research Article

Preclinical studies of novel compounds rely on quantitative readouts from animal models. Frequently employed histopathological tissue scoring are time consuming, require highly specialized staff and subject to inherent variability. Recent advances in deep convolutional neural networks (CNN) now allow automating such tasks. Here, we demonstrate this for the case Ashcroft fibrosis score a newly developed inflammation characterize fibrotic inflammatory lung diseases. Sections mice exhibiting wide range states were stained with Masson trichrome. Whole slide scans using 20x objective acquired cut into smaller tiles 512x512 pixels. The subsequently classified by CNNs, either an "Ashcroft CNN" or "inflammation CNN". For CNN was fine-tuned 14000 labelled tiles. trained 3500 After training, achieved accuracy 79.5% 80.0%. An error analysis revealed that misclassifications almost exclusively neighboring scores, which reflects ambiguity parts data. variability between two experts found be larger than classifications ground truth. generated very good agreement pathologist (r2 = 0.92). Our results costly consuming tasks can automated standardized learning. New scores as easily approach presented here.