
Takes place 10:15 - 11:15.
« Physics-informed machine learning for microscopy image analysis, today and in the future »
Takes place 12:15 - 15:00.
« Analysing Mitochondrial Morphology in Fluorescence Microscopy Images using Simulation-Supervised Learning, Electron Microscopy Data, and Learned Shape Priors »
Quantitative analysis of mitochondrial morphology offers valuable insights into cellular state, yet it remains technically challenging due to the small size, structural diversity, and dense packing of mitochondria. Although fluorescence microscopy enables selective and relatively accessible visualisation of mitochondria, its finite spatial resolution and incomplete axial information complicate reliable segmentation and three-dimensional morphology analysis. Deep learning has addressed related image-analysis problems in other domains, but its application to microscopy is constrained by the need for large, accurately annotated datasets that are difficult and time-consuming to produce. This thesis explores how simulation-supervised learning, high-resolution structural data, and learned shape priors can reduce reliance on manual annotations and support quantitative mitochondrial morphology analysis. It first introduces a simulation-supervised learning pipeline for segmenting mitochondria in fluorescence microscopy images and quantifying experimentally induced morphological changes. Secondly, it constructs a large and morphologically diverse dataset of mitochondrial 3D structures along with corresponding simulated microscopy observations by leveraging high-resolution electron microscopy data, providing a foundation for training and evaluating image-analysis models. Finally, it proposes a data-driven shape prior that reconstructs a plausible three-dimensional mitochondrial surface from single two-dimensional microscopy images or short z-stacks. The proposed methods are evaluated using complementary measures of spatial overlap, volumetric and surface geometry, topological metrics, and downstream morphological utility. The findings demonstrate that simulation can reduce reliance on manually annotated training data, support the detection of experimentally induced differences in mitochondrial morphology, and provide paired supervision for learning plausible three-dimensional shapes when corresponding experimental references are scarce. Nevertheless, broader validation on experimental data remains necessary, and single-view reconstructions should be interpreted as plausible prior-constrained estimates rather than unique recoveries of the true 3D structure. Overall, the thesis establishes a computational framework that connects microscopy simulation, high-resolution structural data, and learned shape priors for quantitative mitochondrial analysis.
The Trial Lecture and Thesis Defense will be streamed via Panopto:
Trial Lecture (10:15 - 11:15)
Watch the Trial Lecture
Thesis Defense (12:15 - 15:00)
Watch the Thesis Defense