Leveraging Machine Learning for Renal Tubular Cast Identification in Dahl Salt-Sensitive Rats

Lauren E. Yunker, Megan C. Harwig, and Alison J. Kriegel

The Dahl salt-sensitive (SS) rat is well established model of SS hypertension. When fed high salt, SS rats experience hypertension and progressive renal injury. One of the hallmark features of renal pathology seen in this model is the presence of proteinaceous tubular casts. It is thought that tubular casts contain a mix of precipitated proteins and thus, can present as various colors when stained for histological analysis. Due to the heterogenous nature of staining, tubular cast quantification can be challenging and time-consuming. The goal of this study was to utilize machine learning to develop a classification method for tubular cast quantification. Kidneys used in this study were obtained from male and female SS rats (n=4-6/group) fed either a low (0.4% NaCl) (LS) or high (4% NaCl) (HS) salt diet for two weeks and were stained with Masson’s trichrome. Training and segmentation were performed using QuPath, an open source software designed for histopathology. Briefly, we trained on four representative kidney scans with varying degrees of pathology. Annotations of different regions were classified depending on tissue type, cast color, and slide background. After visual validation, we built the pixel classifier using the Random Trees model with the following identification features: Gaussian (color/intensity), Laplacian of Gaussian (blob shapes), and weighted deviation (textured/smooth). The pixel classifier was applied across the sample population and tubular casting was measured as % total kidney area. We found that male SS rats on HS had significantly more casting than the LS controls (6.89% ± 0.59 vs 3.98% ± 0.73; mean ± SEM; t-test *p<0.05). There were no significant differences in the female SS rats. To further validate our results, we manually quantified casts in representative samples and saw similar results, suggesting this new method of tubular cast identification is time-effective and accurate and can be implemented in future studies.