Deep Learning models for retinal cell classification

Abstract Data analysis is equally important as an experimental part of the scientist’s work. Therefore any reliable automatization would accelerate research. Histology is a good example, where scientists work with different cell types. The difficulty level can be severe while trying to distinguish cell types from one another. In this paper, we focus on the retina. The retina consists of eight basic cell types, creating a layered structure. Some types of cells overlap within the layer, and some differ significantly in size. Fast and thorough manual analysis of the cross-section is impossible. Even though Deep Learning models are applied in multiple domains, we observe little effort to automatize retinal analysis. Therefore, this research aims to create a model for classifying retinal cell types based on morphology in a cross-section of retinal cell images. In this study, we propose a classification Deep Learning model for retinal cell classification. We implemented two models, each tested in three different approaches: Small dataset, Extended dataset, and One cell type vs. All cell types. Although the problem presented to the trained model was simplified, a significant data imbalance was created from multiclass to binary classification, influencing the models’ performance. Both, Sequential and Transfer Learning models performed best with the Extended dataset. The Sequential model generated the best overall results. The obtained results allow us to place prepared models within the benchmark of published models. This paper proposes the first Deep Learning tool classifying retinal cell types based on a dataset prepared from publicly available images collated from multiple sources and images obtained in our laboratory. The multiclass approach with an extended dataset showed the best results. With more effort, the model could become an excellent analytical tool.