Unsupervised Deep Learning Tools for Microscopy Image Analysis: From 2D to 3D

Abstract

Microscopy can generate detailed and information-rich data, but converting these data into reproducible quantitative measurements is often limited by the need for dense manual annotation. This is especially restrictive when the image content is heterogeneous, weakly defined, dynamic, or volumetric. The work presented in this thesis addresses this challenge by developing annotation-light deep learning workflows for microscopy image analysis across different imaging modalities and biological scales. The thesis brings together three studies. In bright-field microscopy of Bacillus subtilis in microfluidic droplets, an unsupervised variational autoencoder (VAE)-based workflow is used for segmentation and time-resolved quantification of biofilm development. In live-cell atomic force microscopy, the same latent-space workflow is adapted to quantify fenestration-associated regions and their local remodelling dynamics in liver sinusoidal endothelial cells. In volumetric label-free microscopy, LodeSTAR3D extends self supervised object detection to three-dimensional data, enabling label-free localisation and quantitative analysis of intracellular lipid droplets in prostate epithelial cell models. Across these studies, a common conclusion emerges: the main value of annotation-light learning does not lie in the model output alone, but in how that output is translated into operationally defined biological descriptors. This thesis shows that unsupervised and self-supervised learning can provide a practical route from complex microscopy data to quantitative, reproducible, and biologically interpretable readouts in both 2D and 3D imaging. These capabilities establish a foundation for investigating microbial adaptation processes, including antibiotic resistance, as well as cellular remodelling in disease-relevant contexts such as metabolic disorders and cancer.