Master's Thesis Defense by Niels Liljedahl Christensen – Københavns Universitet

Master's Thesis Defense by Niels Liljedahl Christensen


Automatic Detection of Vesicles in FIB-SEM Generated Images


This paper addresses the issue of automatically segmenting neuronal elec- tron microscopy images by neuronal structure, with the main goal of detecting vesicles. Detection of vesicles is necessary in order to properly understand their distribution, both in relation to each other, and in relation to other neuronal structures. To find the optimal vesicle detection method, different machine learning methods were compared by their ability to detect cell membranes on an anistropic dataset, although vesicle detection was done an isotropic dataset. A deep neural network was found to be the best performing of the methods investigated, beating random forests, support vector machines, and logistic re- gression by a large margin. The vesicle detection system created classifies vox- els based on intensity values in their cubic neighbourhoods. Model predictions are compared with annotations labelled by a human observer, and accuracy is measured using the F1-score. The best F1-score obtained is 0.5576, which is an acceptable score considering that ground truth may contain some errors.


Jon Sporring


Jens Damgaard Andersen