Histological phenotyping and neural networks

Manuscript in preparation

With the advent and increased accessibility of deep neural networks (DNNs), complex properties of histological images can be rigorously and reproducibly quantified. We used DNN-based transfer learning to analyze histological 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 histological image processing tasks. The neural network extracted quantitative image features and used them as classifiers to look for differences between mice of different genotypes. We are able to show excellent performance at segmenting glomeruli from non-glomerular structure and subsequently predict the genotype of the animal based on glomerular quantitative image features. The DNN-based genotype classifications highly correlate with mesangial matrix expansion scored by a pathologist, which differed in these animals.  Additionally, by analyzing non-glomeruli images, the neural network identified novel histological features that differed by genotype, including the presence of vacuoles, nuclear count, and proximal tubule brush boarder integrity, which was validated with immune-histological staining. These features were not identified in systematic pathological 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.