Entwicklung eines Algorithmus namens Hanny-Score zur Verbesserung der Kategorisierung von Tiles im Preprocessing, welche aus Whole Slide Images von neuronalen Gewebeproben der Entität Gangliogliom generiert wurden

1.1 Background and objectives The threshold filters Otsu and Hysteresis are a frequently used method for digital image processing or in the preprocessing of neuronal networks to recognize tissue within a whole slide image (Dimitriou et al., 2019). These filters are not very computationally intensive and were originally developed for the abstraction of the subject and background on a digital image. But digital images from tissue samples in medicine differ from everyday subjects. Tissue that is leaky, torn, or washed out during the fixation process can create large gaps in which the background becomes visible. However, this type of background should be assessed differently than the background next to or away from a depicted subject. This difference pushes Otsu and Hysteresis to their limits and requires a different approach, as the Hanny-Score in this dissertation points out. This score not only analyzes the different colors of the background compared to the subject, but also its spatial distribution which is registered using lines instead of individual pixels. In addition, the filter can be adapted to tissue with gaps of different sizes and color wise, which leads to more flexibility and additional possible uses. 1.2 Design and methods This dissertation included 167 histological tissue sections in hematoxylin-eosin staining with neuronal tissue of the entity ganglioglioma from the Erlangen Neuropathological Institute and other institutes. These were digitized into Whole Slide Images (WSIs) using a Hamamatsu S60 scanner. After analyzing the WSIs with the program NDP.view2 version 2.7.52 (Hamamatsu, 2019) and defining Regions Of Interests (ROIs), a total of 3.510.882 image sections (tiles) with a size of 272x272 pixels were created using the program QuPath version 0.2.3 (Bankhead et al., 2017). After collecting a sample of 100 tiles per WSI using a selfwritten program in the Python programming language version 3.11.3 (Van Rossum and Drake, 2009), 5 WSIs including their tiles were excluded because the background was too colorful. Of the remaining 16.200 tiles, 14.755 tiles could be clearly assigned manually to the categories “No Background” and “Background” with 12.264 and 2.941 tiles respectively, thus serving as Ground Truth for evaluating the filters. 1.445 tiles were excluded from the evaluation due to a lack of clear assignment. The selfwritten programs required for this, including their results, were saved in Jupyternotebooks version 6.5.4 (Kluyver et al., 2016). 1.3 Observations and results When comparing the filters Otsu, Hysteresis and Hanny-Score using another selfwritten program in Python with the additional libraries Matplotlib (Hunter, 2007), Scikit-image (van der Walt et al., 2014) and Numpy (Harris et al., 2020), it was shown in fictitious examples and the real data set with the previously determined Ground Truth that the Hanny-Score has a higher probability of assigning the correct category “No Background” or “Background”. Overall, the assignment on the real data set improved to 99,66% with the Hanny score compared to Otsu with 85,28% and Hysteresis with 90,21%. It should be noted that the best individual Cut-Off for separating the two categories was calculated and used for the Otsu and Hysteresis filter. The standard Cut-Off is 30% and with this the results would have been 16,75% instead of 85,28% and 74,04% instead of 90,21%. Thereby the improvement would have been significantly greater. 1.4 Conclusions When selecting suitable data in preprocessing, in this dissertation tissue sections, it is important to analyze them precisely and then apply suitable filters to ensure the most representative and stable data basis possible for further processing operations, such as the use of neural networks. This means more time is spent, but inadequately selected or unsuitable data inevitably leads to unsatisfactory results in the final output. The Hanny-Score requires increased computational effort, but significantly facilitates and improves the selection of suitable digital image sections of tissue samples.