Analyzing Neuronal Dendritic Trees with Convolutional Neural Networks

In the biological sciences, image analysis software are used to detect, segment or classify a variety of features encountered in living matter. However, the algorithms that accomplish these tasks are often designed for a specific dataset, making them hardly portable to accomplish the same tasks on images of different biological structures. Recently, convolutional neural networks have been used to perform complex image analysis on a multitude of datasets. While applications of these networks abound in the technology industry and computer science, use cases are not as common in the academic sciences. Motivated by the generalizability of neural networks, we aim to develop a machine learning algorithm to detect morphological features in the dendritic trees of Drosophila Melanogaster class IV neurons. Our approach is based on the Single Shot Multibox Detector (Liu et. al.) and our training dataset is synthesized from simulations of dendritic trees that we previously developed. Our preliminary results show that the network performs well on the training set. However, on the test set, it sometimes misses objects of interest, which calls for further improvements.