Preview

PDF, English - main document Download (35MB) | Terms of use

Abstract

Methodological progress in the age of Artificial Intelligence (AI) is increasingly driven by large-scale, high-quality datasets. Medical imaging is no exception. As the amount of imaging data grows exponentially, the development and validation of data-driven al- gorithms critically depend on realistic, well-annotated datasets. Photoacoustic imaging (PAI), a hybrid modality combining optical contrast with ultrasound resolution, holds great promise for non-invasive, functional biomedical imaging with the potential for quantitative measurement of blood oxygen saturation. Prospective clinical applications range from cardiovascular imaging to early diagnosis and treatment monitoring in oncology. However, as a comparatively young modality, PAI faces a fundamental challenge: the absence of sufficiently large, annotated datasets for the development and validation of quantitative methods. This thesis addresses the data scarcity problem in quantitative PAI through a comprehensive pipeline for the generation and validation of synthetic spectral imaging data. First, the thesis reviews the current role of deep learning in PAI, identifying the generation of realistic simulated data and domain adaptation as key techniques to address data scarcity. Consequently, SIMPA is introduced, the open-source simulation and image processing for photonics and acoustics framework enabling large-scale and reproducible PA image simulation. To overcome the remaining domain gap between simulated and experimental data, a conditional invertible neural network-based method is proposed for unsupervised domain transfer. For empirical validation of data-driven PA methods, anatomically realistic tissue-mimicking phantoms are fabricated and characterized, enabling the creation of a comprehensive dataset, including experimental and simulated multispectral images. This resource allows, for the first time, rigorous benchmarking of quantitative methods such as oximetry. This thesis contributes not only technically through novel methods and validation strate- gies, but also promotes reproducibility and open science by making all software tools and datasets publicly available. In doing so, it lays the foundation for systematic and transpar- ent development and validation of data-driven methods in PAI and supports the broader translation of PAI into preclinical and clinical applications.