High-throughput quantitative histology in systemic sclerosis skin disease using computer vision

摘要

Deep neural network (DNN) processing of trichrome-stained skin sections. A) Trichrome-stained skin biopsy sections from patients with SSc and healthy controls were photomicrographed at 40x resolution. To sample variability in tissue structure, we randomly selected 100 from the dermis (red box) corresponding to ~ 0.16 mm . B) Each image patch was used as input to the AlexNet DNN. AlexNet maps the raw pixel values of the input image to a series of more complex image features. The final output is a 4096-dimensional signature of abstract Quantitative Image Features that were used for subsequent multivariate statistical analyses. C) Principal components analysis and multivariate analyses using QIF as the predictor variables were conducted in order to develop, D) a Biopsy Score, E) a Diagnostic Score, F) a Fibrosis Score that was compared to mRSS and skin gene expression biomarkers