TY  - GEN
N2  - This thesis shows that data-driven approaches have the potential to solve many of the challenges of achieving quantitative photoacoustic imaging. To this end, rigorous in silico evaluation of machine learning algorithms for the inverse problems associated with photoacoustic imaging is conducted, a data-driven approach for blood oxygenation estimation from multispectral photoacoustic measurements is applied in vitro and in vivo, and methods for uncertainty estimation for the developed algorithms are analyzed.

Photoacoustic imaging is an emerging imaging modality in healthcare. It promises noninvasive and radiation-free imaging of optical tissue properties. In contrast to commonly used optical imaging techniques, it can visualize optical tissue properties up to several centimeters deep in tissue. Photoacoustic imaging is based on the photoacoustic effect, which enables spatially resolved imaging of optically absorbing chromophores. When pulsed laser light is sent into tissue and is absorbed by chromophores, sound waves emerge at the location of the absorption event. These can be measured with ultrasound transducers and are reconstructed into a spatial image of absorbed energy. When using multiple measurements of the absorbed energy at different wavelengths of light, knowledge on clinically relevant functional tissue parameters, such as blood oxygenation, can be derived.

However, one of the critical challenges in photoacoustic imaging remains unsolved. This challenge is accurate and reliable quantification of the underlying optical tissue properties. Especially the estimation of the optical absorption coefficient of the tissue is essential for the derivation of functional tissue parameters. However, the absorbed energy is not only proportional to the optical absorption coefficients, but instead, it is also proportional to the light fluence. The fluence describes the distribution of light in tissue, which is predominantly determined by the optical absorption and scattering properties. Due to this, there is a non-linear interaction of optical absorption and fluence with respect to the absorbed energy and quantification of the signal is an ill-posed inverse problem, to which no general and easy-to-compute solution yet exists.

In the field, iterative model-based approaches have been proposed and thoroughly investigated. In numerous in silico investigations, these techniques have shown great theoretical potential. However, they have not successfully been applied to real images acquired in clinically relevant freehand imaging settings. 

In this thesis, the feasibility of developing and applying data-driven methods to fill this gap is investigated. Data-driven methods refer to machine learning algorithms that learn an optimal inference model for a particular application based on training data that are relevant to the problem. For optical imaging applications generally, no ground truth information on the underlying tissue properties is available, and in this thesis, it is attempted to train the data-driven algorithms on computer-simulated data with the hope to gain the ability to infer tissue properties also in real scenarios.

The capability of data-driven methods to estimate the light fluence or the optical absorption coefficients was examined in two rigorous in silico studies. Each of these was conducted on several different data sets of different difficulty levels. To this end, an algorithm was developed that can encode the entire three-dimensional signal context in a voxel-specific low-dimensional feature vector. Furthermore, state-of-the-art deep learning algorithms were used to estimate the optical absorption distribution directly on two-dimensional photoacoustic images. The results of these experiments show the general feasibility of data-driven approaches in the photoacoustic imaging context and also reveal the current limitations of these methods.

Another data-driven approach was developed that was trained on simulated data and enabled the estimation of functional tissue properties in both in vitro and in vivo settings. Here, the capability of the method to predict plausible results for blood oxygenation in various contexts was demonstrated. The method continuously outperformed linear unmixing techniques in terms of the estimated dynamic range and agreement of the estimates with the expected values.

Finally, four different techniques for uncertainty estimation were examined towards their applicability to photoacoustic imaging. The conducted experiments revealed that the integration of uncertainty estimates during the calculation of results on a bigger region of interest could potentially be of great benefit.

For sure, more work is required to achieve a successful and robust application of data-driven quantification of photoacoustic signals. This work, however, revealed the potential of data-driven methods in this context and outlined several possible ways of applying them to the associated inverse problems. It can be assumed that the combination of model-based and data-driven approaches will form the foundation for a successful clinical translation of quantitative
photoacoustic imaging into clinical practice.
TI  - Data-driven Quantitative Photoacoustic Imaging
A1  - Gröhl, Janek Matthias
AV  - public
KW  - Photoakustische Bildgebung
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/29676/
Y1  - 2021///
ID  - heidok29676
CY  - Heidelberg
ER  -