Abstract
The ability to automatically detect and classify populations of cells in tissue sections is paramount in a wide variety of applications ranging from developmental biology to pathology. Although deep learning algorithms are widely applied to microscopy data, they typically focus on segmentation which requires extensive training and labor-intensive annotation. Here, we utilized object detection networks (neural networks) to detect and classify targets in complex microscopy images, while simplifying data annotation. To this end, we used a RetinaNet model to classify genetically labeled neurons and glia in the brains of Mosaic Analysis with Double Markers (MADM) mice. Our initial RetinaNet-based model achieved an average precision of 0.90 across six classes of cells differentiated by MADM reporter expression and their phenotype (neuron or glia). However, we found that a single RetinaNet model often failed when encountering dense and saturated glial clusters, which show high variability in their shape and fluorophore densities compared to neurons. To overcome this, we introduced a second RetinaNet model dedicated to the detection of glia clusters. Merging the predictions of the two computational models significantly improved the automated cell counting of glial clusters. The proposed cell detection workflow will be instrumental in quantitative analysis of the spatial organization of cellular populations, which is applicable not only to preparations in neuroscience studies, but also to any tissue preparation containing labeled populations of cells.