Automated Segmentation of 3D Electron Microscopy Images

Mitochondria and endoplasmic reticulum (ER) are of great interest to biologists because of their influence on diseases such as cancer, dementia, Alzheimer, etc. To visualise and analyse these structures, 3D electron microscopy (EM) is typically used due to its high resolution imaging capability. As a result, biologists are confronted with large amounts of complex ultrastructural data that has to be analysed.

In this work, techniques are studied to automate image analysis (particularly, segmentation) in volume EM. Restoration techniques are developed that improve the quality of EM data by removing image degradations such as noise and blur. Furthermore, machine learning domains such as unsupervised learning, ensembles and domain adaptation are studied to allow for a more ergonomic and automated segmentation pipeline.