From Light to Data Using Deep Learning for Quantitative Microscopy

Abstract

Quantitative microscopy aims to convert optical measurements into reliable information about the physical properties of microscopic particles. In biological systems these particles often produce weak and variable signals, which limits the effectiveness of conventional analysis methods. This dissertation investigates how computational techniques, with a focus on deep learning guided by physical insight, can improve the extraction of quantitative information from microscopy images and enable more accurate particle characterization.

The thesis first outlines the theoretical foundations of optical microscopy, deep learning, and data acquisition used throughout this work. It then introduces a self-supervised method for particle detection and tracking that does not require annotated training data. This approach is extended to off-axis holography, enabling three-dimensional localization from raw interference patterns. A tutorial-style review of label-free microscopy follows, highlighting both the strengths and limitations of optical measurements and illustrating practical workflows for quantitative analysis.

Building on this foundation, the thesis demonstrates how hybrid imaging, implemented through dual-angle interferometric scattering microscopy, can combine complementary information to improve nanoparticle characterization. The final part presents a deep learning method for reconstructing three-dimensional refractive index distributions from images recorded at unknown orientations, providing richer structural and compositional information about individual particles.

Overall, this work presents a coherent progression from optical principles to computational tools that strengthen the connection between recorded light and quantitative particle data. The resulting methods expand the capabilities of label-free microscopy for studying microscopic particles, with particular relevance to applications in biological imaging.