Representation learning and applications in neuronal imaging

Confocal fluorescence microscopy is a microscopic technique that provides true three-dimensional (3D) optical resolution and that allows the visualization of molecular expression patterns and morphological structures. This technique has therefore become increasingly more important in neuroscience, due to its applications in image-based screening and profiling of neurons. However, in the last two decades, many approaches have been introduced to segment the neurons automatically. With the more recent advances in the field of neural networks and Deep Learning, multiple methods have been implemented with focus on the segmentation and delineation of the neuronal trees and somas. Deep Learning methods, such as the Convolutional Neural Networks (CNN), have recently become one of the new trends in the Computer Vision area. Their ability to find strong spatially local correlations in the data at different abstraction levels allows them to learn a set of filters that are useful to correctly segment the data, when given a labeled training set. The overall aim of this thesis was to develop a new algorithm for automated segmentation of confocal neuronal images based on Deep Learning techniques. In order to realize this goal, we implemented a U-Net-based CNN and realized the dataset necessary to train the Neural Network. The results show how satisfactory segmentations are achieved for all the test images given in input to our algorithm, by obtaining a Dice coefficient, as average of all the images of the test dataset, greater than 0.9.