Cell compartments can vary in morphology. Mitochondria within eukaryotic cells, the focus of this thesis, can be branched or touch each other. Our objective is the classification of mitochondrial morphology.
Since it is difficult to obtain a representation of the raw shape that would describe its morphology well, we turned to shape abstraction. Abstraction describes a shape with a small set of geometric primitives. We evaluated three shape abstraction methods that utilize deep learning. These differ in the type of input, the method of evaluating abstraction quality and in how the number of primitives is predicted. By modifying the best performing method, we achieved good abstraction quality.
We then performed morphology classification based on the distance between vectors of abstraction parameters. The classification was not satisfactory. We also showed that these distances poorly separate mitochondria with different morphologies by embedding the distances and plotting the embeddings on a scatter plot. We were unable to perform classification with more powerful machine learning methods due to a lack of mitochondria with labelled morphology.
|
