A deep learning approach to neurite prediction in high throughput fluorescence imaging

Changes to neuronal morphology and loss of neurites and synaptic connections can be an important early indicator of neurological diseases, and a pathognomonic sign of neurodevelopmental disorders. These changes are typically detectable by microscopy in cell culture or histological samples, but quantification can be challenging. The neurites extending from cell soma can be quite thin, dim, overlapping and complex, making them laborious to trace manually and difficult to annotate and quantify computationally or automatically. Moreover, the tools available to aid this aim are limited in their capacity to generalize to high throughput image acquisition such as time-lapse or longitudinal imaging, where imaging conditions can change dramatically over the course of the experiment. In order to facilitate neurite quantification, we developed a deep learning (DL) neurite annotation prediction algorithm (NAPA) to predict the structure and length of neurites. NAPA overcomes experimental variation inherent to fluorescence imaging by learning more broader features that are important for neurite recognition. Based on a dataset with partial annotation, NAPA generated predictions on several unannotated datasets, and was able to capture differences between disease and control conditions. We also defined a sequence of steps to generate custom models with a small number of annotation inputs, and extended the predictions to a 3D tissue sample and longitudinal imaging. With this algorithm we developed an approach to quantify neurites with an accuracy that nears and sometimes exceeds human curation, in 1/100th of the time. This approach makes accurate analysis of large or longitudinal datasets feasible across a broad range of datasets.