Supervised and Unsupervised Learning Methods for the Automation of Glomerular Scoring
John D. Bukowy, Louise Christine Evans, Elizabeth Broadway, Alex Dayton, Allen W. Cowley · The FASEB Journal · 2015
A single rat kidney can contain in excess of 30,000 glomeruli whose function is to act as a filter at the initial interface between the bloodstream and tubular fluid. Sufficient damage to glomeruli and their surrounding Bowman's space leads to altered fluid handling within the kidney and disease states. Scoring of glomeruli from histological preparations with Gomori's One‐Step Trichrome staining can quantify glomerular damage but requires a tedious and time consuming process of manual scoring individually selected glomeruli. In order to decrease inter/intra‐observer variability, analysis time, and resource dedication we propose a method of feature selection for the automation of glomerular scoring. The dataset was comprised of over 1,300 previously human‐scored bright‐field 40x histological images of rat glomeruli obtained under varying experimental conditions. The images consisted of glomeruli surrounded by tubular segments. Using an approach that considers the center of mass of salient features from the images, regions containing glomeruli were located and digitally “masked”, or isolated, within individual images. Unsupervised k‐ means clustering was then implemented for color segmentation of the masked glomerular region to select features emphasized from trichrome staining. By training with the described feature set, the scores predicted by the classifier were found to be not significantly different than the human observer, thereby automating a subjective workflow while maintaining precision of scores. While the specific feature set construction described here may only be applicable to the scoring of glomerular damage assayed through trichrome staining, this approach may be applied to a variety of histological assays and stains. (HL116264)