FIBER-ML, an Open-Source Supervised Machine Learning Tool for Quantification of Fibrosis in Tissue Sections

Abstract

Pathologic fibrosis is a major hallmark of tissue insult in many chronic diseases. Although the amount of fibrosis is recognized as a direct indicator of the extent of disease, there is no consentaneous method for its quantification in tissue sections. This study tested FIBER-ML, a semi-automated, open-source freeware that uses a machine-learning approach to quantify fibrosis automatically after a short user-controlled learning phase. Fibrosis was quantified in sirius red-stained tissue sections from two fibrogenic animal models: acute stress-induced cardiomyopathy in rats (Takotsubo syndrome-like) and HIV-induced nephropathy in mice (chronic kidney disease). The quantitative results of FIBER-ML software version 1.0 were compared with those of ImageJ in Takotsubo syndrome, and with those of inForm in chronic kidney disease. Intra- and inter-operator and inter-software correlation and agreement were assessed. All correlations were excellent (>0.95) in both data sets. The values of discriminatory power between the pathologic and healthy groups were <10^-3 for data on Takotsubo syndrome and <10^-4 for data on chronic kidney disease. Intra-operator agreement, assessed by intra-class coefficient correlation, was good (>0.8), while inter-operator and inter-software agreement ranged from moderate to good (>0.7). FIBER-ML performed in a fast and user-friendly manner, with reproducible and consistent quantification of fibrosis in tissue sections. It offers an open-source alternative to currently used software, including quality control and file management.