Abstract
With the advent and increased accessibility of deep neural networks (DNNs), complex properties of histologic images can be rigorously and reproducibly quantified. We used DNN-based transfer learning to analyze histologic images of periodic acid-Schiff-stained renal sections from a cohort of mice with different genotypes. We demonstrate that DNN-based machine learning has strong generalization performance on multiple histologic image processing tasks. The neural network extracted quantitative image features and used them as classifiers to look for differences between mice of different genotypes. Excellent performance was observed at segmenting glomeruli from non-glomerular structure and subsequently predicting the genotype of the animal on the basis of glomerular quantitative image features. The DNN-based genotype classifications highly correlate with mesangial matrix expansion scored by a pathologist (R.E.C.), which differed in these animals. In addition, by analyzing non-glomeruli images, the neural network identified novel histologic features that differed by genotype, including the presence of vacuoles, nuclear count, and proximal tubule brush border integrity, which was validated with immunohistologic staining. These features were not identified in systematic pathologic examination. Our study demonstrates the power of DNNs to extract biologically relevant phenotypes and serve as a platform for discovering novel phenotypes. These results highlight the synergistic possibilities for pathologists and DNNs to radically scale up our ability to generate novel mechanistic hypotheses in disease.