Applying Deep Learning to Melanocyte Counting on Fluorescent TRP1 Labelled Images of In Vitro Skin Model

Abstract

Cell counting is an important step in many biological experiments. It can be challenging, due to the large variability in contrast and shape of the cells, especially when their density is so high that the cells are closely packed together. Automation is needed to increase the speed and quality of the detection. In this study, a cell counting method is developed for images of melanocytes obtained after fluorescent labelling with TRP1 (Tyrosinase-related protein 1) of 3D reconstructed skin samples. Following previous approaches, a strategy based on predicting the local cell density, by means of a convolutional neural network (a U-Net), was adopted. The method showed great efficiency on a test set of 76 images, with an assessed counting error close to 10% on average, which is a commonly accepted target in cytology and histology. For comparison purposes, we have made our dataset publicly available.

Keywords

• cell counting
• convolutional neural networks (cnn)
• deep learning
• fluorescence microscopy
• histology
• immunohistochemistry (ihc)
• reconstructed skin