CNNs for electron microscopy segmentation
Realitzat a/ambÉcole Polytechnique Fédérale de Lausanne
Tipus de documentTreball Final de Grau
Condicions d'accésAccés restringit per decisió de l'autor
In the framework of Biomedicine, mitochondria are known to play an important role in neural function. Recent studies show mitochondrial morphology to be crucial to cellular physiology and synaptic function, and a link between mitochondrial defects and neuro-degenerative diseases is strongly suspected. Electron microscopy (EM), with its very high resolution in all three directions, is one of the key tools to look more closely into these tissues, but the huge amounts of data it produces make automated analysis necessary. In this work, we present an approach to perform tasks of biomedical image segmentation and visual pattern recognition by using Convolutional Neural Networks (CNNs), a pixel trainable classifier. The aim is, given a set of electron microscopy (EM) images of neural tissue, to segment mitochondria and measure the performance. Due to the fact that it is a machine learning algorithm, it previously has to learn feature extractors with large amounts of data belonging to a training dataset. Both training and test sets within the corresponding ground truth are from the striatum, a subcortical part of the forebrain. The model is composed of a series of layers, each of them containing basic computing units or neurons, with an input layer receiving raw pixel values from the images dataset and an output layer predicting the label of each pixel (mitochondrion or non-mitochondrion) in a square window centered on it. Besides, convolutional networks gather clusters of neurons in feature maps throughout a succession of convolutional and max-pooling layers that preserve 2D information and extract features with increasing levels of abstraction. The classifier is trained by stochastic gradient descent. We demonstrate that, although convolutional neural networks do not outperform state-of-theart algorithms of CVLAB in biomedical image segmentation for the same purposes, they represent an alternative and competitive approach. Despite the fact of being computationally complex and with high costs in time and CPU needs, it has the advantage of performing directly from raw pixel intensities as inputs and without ad-hoc post-processing. This characteristic makes them suitable to be applied in several other biomedical image segmentation tasks, or to use other types of inputs rather than pixels. On the other hand, neural networks assemble such an amount of degrees of freedom and unconstrained variables that it actually constitutes an interesting eld with still many unexplored paths. All these properties open interesting perspectives on further research in the building of deeper or wider neural networks, or the study of ltering and feature maps.