Open Access

Felix Herter

• Aim of this thesis was to evaluate the performance of three popular machine learning methods – decision trees, support vector machines, and neural networks – on a supervised image classification task from the domain of cell biology. Specifically, the task was to classify microtubule ends in electron tomography images as open or closed. Microtubules are filamentous macromolecules of the cytoskeleton. Distribution of their end types is of interest to cell biologists as it allows to analyze microtubule nucleation sites. Currently classification is done manually by domain experts, which is a difficult task due to the low signal-to-noise ratio and the abundance of microtubules in a single cell. Automating this tedious and error prone task would be beneficial to both efficiency and consistency. Images of microtubule ends were obtained from electron tomography reconstructions of mitotic spindles. As ground truth data for training and testing four independent expert classifications for the same samples from different tomograms were used. Image information around microtubule ends was extracted in various formats for further processing. For all classifiers we considered how the performance varies when different preprocessing techniques (per-feature and per-image standardization) are applied. or decision trees and support vector machines we also evaluated the effect of training on a) imbalanced versus under- and over-sampled data and b) image-based vs feature-based input for specifically designed features. The results show that for decision trees and support vector machines classification on features outperforms classification on images. Both methods give most equalized per-class accuracies when the training data was undersampled and when preprocessed with per-image standardization prior to features extraction. Neural networks gave the best results when no preprocessing was applied. The final decision tree, support vector machine, and neural network obtained accuracies on the test set for (open,closed ) samples of (62%, 72%), (66%, 70%), and (61%, 78%) respectively, when considering all samples where at least one expert assigned a label. Restricting the test set to samples with at least three agreeing expert labels raised these to (78%, 84%), (74%, 92%), and (82%, 88%). It can be observed that many samples misclassified by the algorithms were also difficult to classify for the experts.