Automatic Detection of Immunolabeled Cells Using Convolutional Neural Networks

Abstract

Transsynaptic viral tracing is an important technique used to studyconnections between distant areas of the brain. Currently, the state ofthe art methods for automatically detecting the cells marked by this techniqueinvolve sliding predetermined, spherical filters over the image andapplying a watershed transform on the result. While potentially effectivethey have several drawbacks, such as the need to manually tune severalhyper-parameters and in particular are not robust to sample-to-samplevariation common in a dataset. This thesis proposes a supervised learningapproach to solve the problem of cell detection. By using ConvolutionalNeural Networks and Filtered Local Max post-processing, I proposea method to detect fluorescently-labeled virus-infected cells in lightsheetimaged mouse brains that removes the need for manual tuning of hyperparameters.Additionally, I examine the process with which I had determinedthe appropriate network architecture, in the hopes of streamliningfuture similar projects for people without extensive knowledge of NeuralNetworks.