Open Access

Tim Tobias Arndt

• In the ever-evolving landscape of healthcare, the demand for sophisticated and efficient diagnostic tools has intensified. Medical imaging, serving as a cornerstone in disease detection and characterization, has witnessed a transformative shift with the advent of ar- tificial intelligence. In this thesis we explore the usage of both traditional image processing methods and advanced deep learning techniques to address medical questions. This thesis starts with a short overview of mathematical concepts, definitions, and proofs which we will use in later chapters. In Chapter 3, three image processing techniques are introduced, namely color decon- volution, background subtraction, and thresholding. Within Section 3.3, we propose a novel algorithm designed to efficiently compute multilevel thresholds. Depending on the input parameters, this algorithm is multiple orders of magnitude faster than currently used implementations. We achieve this by using an improving moves algorithm to find aIn the ever-evolving landscape of healthcare, the demand for sophisticated and efficient diagnostic tools has intensified. Medical imaging, serving as a cornerstone in disease detection and characterization, has witnessed a transformative shift with the advent of ar- tificial intelligence. In this thesis we explore the usage of both traditional image processing methods and advanced deep learning techniques to address medical questions. This thesis starts with a short overview of mathematical concepts, definitions, and proofs which we will use in later chapters. In Chapter 3, three image processing techniques are introduced, namely color decon- volution, background subtraction, and thresholding. Within Section 3.3, we propose a novel algorithm designed to efficiently compute multilevel thresholds. Depending on the input parameters, this algorithm is multiple orders of magnitude faster than currently used implementations. We achieve this by using an improving moves algorithm to find a local maximum and then repeat this process with different initial values to increase the likelihood of terminating in the global maximum. We show that this approach will find the best threshold in almost all test images, for both medical and images from the ImageNet (Deng et al. (2009)) dataset. In Chapter 4 we use a deep neural network to predict whether patients will develop a distant metastasis in five years after treatment based on images of the invasion front of their tumor. The early identification of metastatic risk can be a valuable tool for guiding treatment decisions, allowing for the implementation of more aggressive protocols when deemed necessary. The model takes binary images of tumors as input to emphasize the architecture of the tumor. These images were prepared from stained histological images for the invasion front of the tumor, which were subsequently transformed into binary black and white images. Based on the output of the neural network each patient is assigned to a high or low risk group. A drawback associated with the utilization of deep learning models is their tendency to function as black boxes, often making it challenging to decipher the factors influencing a particular prediction. Following the fitting and evaluation of the neural network model in Section 4.1, we employed two methodologies to delve into the reasoning behind the model’s predictions in Section 4.2. Firstly, we generated heatmaps to visually represent the areas of the input image exerting the most influence on the prediction. Subsequently, we crafted synthetic images with specific features to examine how these features, along with their intensity, impacted the prediction of the model. Our initial approach to create synthetic images involved using noise to generate random images. This enabled us to observe the influence of noise parameters on the model’s predictions. Further exploration involved modifying tumor borders of images previously used in the evaluation of the model and analyzing the effect of these alterations. In Chapter 5 we use deterministic image manipulation algorithms, as introduced in Chapter 3, to calculate a score for the expression of focal adhesion kinase (FAK) in tissue using microscope images. FAK expression is often used as a diagnostic marker, but is mostly evaluated subjectively by the individual pathologist. Our approach involves approximating the proportional stain concentration within lesions in comparison to the surrounding tissue, leveraging the Beer–Lambert Law (Beer (1852)). In Chapter 6, we present four supplementary studies, each of which culminated in publications resulting from collaborative efforts with research groups at the Augsburg University Hospital. Here the majority of the work was performed by the project partners at the Hospital, who performed the studies and aggregated the data. My involvement was planning and performing the statistical evaluation of said data. In Section 6.1, we present two studies scrutinizing the disparities in lymphocyte cell counts between distinct patient groups and a healthy control cohort. The initial study focuses on patients afflicted with COVID-19, while the subsequent investigation centers on patients diagnosed with colorectal cancer. In Section 6.2 two studies are outlined that investigate the potential advantages that patients can derive from the implementation of virtual reality-based interventions during their hospitalization and recovery periods.…