@thesis{Teklenburg2024, author = {Teklenburg, Linus Paul}, title = {AI-based classification of American football plays combining computer vision and historical play-by-play data}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-47451}, pages = {XII, 58,}, year = {2024}, abstract = {This bachelor's thesis explores the integration of visual and text-based models for predicting American football plays based on National Football League (NFL) games. The study focuses on automatically extracting visual features from pre-snap images of NFL plays and combining them with a text-based model trained on historical play-by-play data. The visual features are extracted using computer vision techniques and YOLOv8 model architecture, while the text-based model utilizes an XGBoost model to analyze historical play-by-play data. The research methodology involves preprocessing and analyzing a dataset of NFL pre-snap images and historical play-by-play data. Computer vision algorithms, such as Optical Character Recognition or Line Extraction are employed to extract relevant visual features, such as player positions, formations, and field dynamics, from the images. Simultaneously, an XGBoost model is trained on historical play-by-play descriptions to capture textual patterns associated with different play outcomes, such as pass or run. The extracted visual features and text-based predictions are then integrated into a unified prediction pipeline. The effectiveness of the integrated model is evaluated through extensive experimentation and performance analysis. Various metrics, including accuracy, precision, recall, and F1 score, are used to assess the predictive capabilities of the model. The results demonstrate the potential of combining visual and text-based information for accurate prediction of NFL plays. The integrated model achieves promising performance in predicting pass or run plays with a test accuracy of 74.13\% and validation accuracy of 73.78\%. The findings of this research contribute to the advancement of predictive analytics in sports and provide valuable insights for coaches, analysts, and the football and data science community.}, language = {en} } @thesis{Jacobowsky2023, author = {Jacobowsky, Emanuel}, title = {Detecting and counting wheat spike heads from UAV-based images using deep convolutional neural networks}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-36193}, pages = {XII, 49,}, year = {2023}, abstract = {Accurate detection and counting of wheat head spikes is essential for maximizing wheat yields, but this task is currently done manually by farmers, which can be time-consuming and subject to human error. This thesis demonstrates, that using a deep convolutional neural network (DCNN), specifically a Mask R-CNN, can significantly reduce the time and labor needed while also eliminating human factors such as subjectivity and exhaustion. To evaluate the performance of the Mask R-CNN, a new wheat dataset was created, containing over 500 images of size 392x392 pixels with approximately 40000 annotated wheat spike heads in total, captured by unmanned areal vehicles (UAVs). Results show that the Mask R-CNN achieved a mAP@0.50:0.95 (mean Average Precision) of 0.670 and 0.569 along with a mAR@0.50:0.95 (mean Average Recall) of 0.733 and 0.621 for bounding boxes and masks, respectively. In addition, 6338 spikes across 77 images were detected on a testset, while the annotations totaled 6200, resulting in a positive 2.14\% difference. Further analysis of 100 wheat plots using a count-based heatmap revealed potential opportunities for optimization to increase wheat yields. A gradient density map also divided plots into three different zones and showed that an average of 8.42\% of the area was occupied by wheat spike heads. Further assumptions by agricultural experts over nutrition on the border of wheat plots were disproven. Overall, this thesis provides a new, high-quality wheat dataset and new insights into the structure of wheat fields.}, language = {en} } @thesis{Ostermaier2023, author = {Ostermaier, Julian Daniel}, title = {Predicting c-Kit mutations in canine cutaneous mast cell tumors from H\&E stained whole slide images with deep learning}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-36272}, pages = {67}, year = {2023}, abstract = {Canine Cutaneous Mast Cell Tumors (CCMCTs) present one of the most common cancer subtypes in dogs. The highly variable behavior of CCMCTs makes it hard for pathologists to diagnose and treat the patients. One important prognostic factor that has been identified are mutations in the c-Kit gene, a gene which encodes the receptor tyrosine kinase and can influence cell proliferation. Recent developments in the area of computer vision have shown, that genetic mutations are reflected in the histopathological phenotype and can be accurately classified and detected from hematoxylin and eosin-stained (H\&E) whole slide images (WSIs) using deep learning. In this thesis, we apply multiple instance learning, a type of weakly-supervised learning approach, to a dataset of 457 WSIs of MCTs and show that c-Kit mutations can be detected from the image alone. We also compare different self-supervised pre-training strategies in order to learn better feature representations and improve the accuracy of our mutation prediction. Finally, we use the image regions of high diagnostic importance provided by our deep learning model, and let pathologists examine the difference in histopathological features for each c-Kit mutation.}, language = {en} } @thesis{Willweber2023, author = {Willweber, Julian Thomas Dieter Rudolf}, title = {Comparison of deep learning and word embedding approaches for German fake news detection}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-38934}, pages = {viii, 50}, year = {2023}, abstract = {The ever-growing amount of fake news that is spread via social media and other digital platforms poses an increasingly dangerous risk for society. The mass of news has become too large to check for fake news manually. Therefore, automatic fake news detection systems have to be developed. Because most of the current research on fake news detection is done in English language, the goal of this thesis is to reach an overview of multiple approaches for German fake news detection. The overview is gained by testing and comparing various machine learning algorithms, namely, BERT, DistilBERT and the Support Vector Machine combined with multiple word embedding algorithms. The algorithms were trained and tested on the FANG-COVID dataset. While the different models and word embedding algorithms achieved varying results in performance, all models were capable of achieving accuracy scores over 90\%. This promises potential for their application and future research in German fake news detection.}, language = {en} } @thesis{Winkler2023, author = {Winkler, Jakob Matthias}, title = {Entwicklung eines modellgest{\"u}tzten Deep Learning Ansatzes zur Rekonstruktion multiparametrischer Magnetresonanz-Korrelationsspektren}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-36441}, pages = {80}, year = {2023}, abstract = {In der Arbeit werden Neuronale Netze verwendet, um T1/T2-Spektren aus MRT-Messdaten zu rekonstruieren. Anschließend werden die Neuronalen Netze mit einem in-vivo-Datensatz evaluiert.}, language = {de} } @thesis{Beerschwinger2023, author = {Beerschwinger, Lukas Bastian}, title = {Patient or tumor-reidentification from histopathology images}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-37770}, pages = {65}, year = {2023}, abstract = {In this bachelor thesis, deep learning methods are used to try to re-identify Patients or Tumors using histopathology images (Whole Slide Images, Slides). Since medical data contains a lot of sensitive personal information, it is very important to anonymize it before publication. If, however, Re-identification is possible, it would be a serious invasion of the privacy of the persons who once agreed that their tissue sections could be published under the condition of confidentiality. According to the current status, it has not yet been checked whether a Re-identification of Patients or Tumors is only possible using histopathological images (without further meta data). It is therefore important to check potential possibilities with the help of deep learning methods in order to be able to take appropriate measures in the future. The peculiarity of such Slides is that they are very large files, which cannot be sent in their entirety through a neural network or pipeline. They are divided into patches, which are sections of a Slide. In previously published Re-identification / Verification tasks (e.g. face recognition or x-ray images) mainly structures (face, thorax) were learned for recognition by the neural networks. This is not possible with Whole Slide Images. Rather, an attempt must be made to extract the Patient's feature information in order to be able to recognize it in another Slide of the Patient. The thesis shall investigate this question on a Dataset of around 1000 microscopy images showing meningioma Tumors from around 300 Patients that is available from a major German university hospital. The data was completely anonymized beforehand and a random ID was assigned to a Patient so that all Slides of a Patient can be traced back to them. Two multi-class classification approaches were developed and evaluated. Approximately 80\% of the Patients could be re-identified. Using an evaluation adapted to the Dataset, this work highlights possibilities and limitations of Patient Re-identification based on meningioma slides.}, language = {en} } @thesis{Klose2023, author = {Klose, Laura Manuela}, title = {Prediction of histopathological markers for computer-aided grading of mammary carcinoma}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-37690}, pages = {47}, year = {2023}, abstract = {An important task of pathologists is to assign a grade during a tumor diagnosis. The grade correlates with a prognosis, that indicates the chances of survival. This prognosis plays a central role in therapy selection and prospects. Breast carcinomas have many histological subtypes, which are quite similar among themselves. Nevertheless, the pathologists need to perform the grading as accurately as possible. There are multiple grading schemes available. All schemes evaluate the 3 criteria: tubule formation, nuclear pleomorphism and mitotic count. During the evaluation, a label between 1-3 is assigned to each criterion. Afterward, the sum of all labels determines the tumor grade. Whereas the mitotic count is a rather objective evaluation, the tubule formation and nuclear pleomorphism labels are subjective decisions and therefore a strong inter-observer variability exists. Having an algorithm that could provide a deterministic evaluation, would improve reliability and could also improve diagnostic accuracy. The goal of this bachelor thesis is to develop a machine learning-based framework, that on the one hand classifies the grade of the subjective criteria (i.e., tubule formation and nuclear pleomorphism) and on the other hand semantically segments important tissue regions used for the prediction. The overall goal of this framework is to support pathologists in clinical decision-making by providing not only a score for each category per image, but also segmentation information and thus providing a more detailed and interpretable decision support. The dataset provided for this thesis contains mammary tumors in dogs. Due to the similarity between canine and human mammary tissue, successes achieved with canine tissue could be directly transferred to humans. Since annotating many images to train a supervised model is time-consuming, a relatively small dataset is provided. According to that limitation, the first research topic is how to overcome this problem by using transfer learning, image augmentation and image preprocessing. Secondly, it will be evaluated whether the outputs of the segmentation and classification tasks are precise enough to be used by a pathologist.}, language = {en} } @thesis{Pfaller2023, author = {Pfaller, Thomas}, title = {Semantic segmentation of infill development potentials on aerial images via neural networks and the fusion of image- and geodata}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-37092}, pages = {55}, year = {2023}, language = {en} } @thesis{Hecker2023, author = {Hecker, Moritz Ernst}, title = {Volumenbasierte Berechnung der Mitoseaktivit{\"a}t von Mammakarzinomen mit Hilfe neuronaler Netze auf H{\"a}matoxylin-Eosin-Bildern}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-37984}, pages = {60}, year = {2023}, abstract = {H{\"a}ufig wird in der Pathologie zur Beurteilung und Einteilung der unterschiedlichen Tumorarten die Mitoseaktivit{\"a}t in den Gewebestrukturen gez{\"a}hlt (engl. mitotic count). Zusammen mit anderen Gewebemerkmalen wird daraus ein histologischer Grad bestimmt, wodurch sich Prognosen und Behandlungsempfehlungen {\"u}ber das Krankheitsbild ableiten lassen. Ein Problem dieses Ansatzes ist, dass h{\"a}ufig nur ungenaue und nicht reproduzierbare Aussagen {\"u}ber den Tumor getroffen werden k{\"o}nnen. Dies liegt unter anderem daran, dass im Untersuchungsbereich auch gewebsfreie Bereiche oder nicht-tumor{\"o}ses Gewebe enthalten sein kann. Dies f{\"u}hrt, insbesondere bei epithelialen Tumoren (Karzinomen) zu einer Verzerrung der Ergebnisse und verf{\"a}lscht dadurch die Prognose. Ein verbesserter und aussagekr{\"a}ftigerer Ansatz ist die volumenbasierte Berechnung der Mitoseaktivit{\"a}t (engl. volume corrected mitotic). Hierbei werden gewebefreie Bereiche ausgeschlossen und die Berechnung erfolgt in Abh{\"a}ngigkeit von dem relevanten Epithelgewebe. Dadurch k{\"o}nnen pr{\"a}zisere Vorhersagen getroffen und umfassendere Diagnosen gestellt werden. Dennoch hat sich dieser Ansatz in der Praxis der Pathologie derzeit noch nicht durchgesetzt, da neben der Standardf{\"a}rbung mit H{\"a}matoxylin und Eosin (H\&E) eine weitere Immunhistochemische Spezialf{\"a}rbung mit Cytokeratin-Antik{\"o}rper durchgef{\"u}hrt werden muss, wodurch ein gr{\"o}ßerer Zeitaufwand f{\"u}r die Pathologin oder den Pathologen entsteht und damit h{\"o}here Kosten verursacht. Ziel dieser Bachelorarbeit ist es, ein automatisiertes Verfahren zu entwickeln, das auf H\&E gef{\"a}rbten Gewebeproben von Hunden zun{\"a}chst die Epithelzellen in den Gewebestrukturen detektiert. In einem weiteren Schritt soll nun die volumenbasierte Mitoseaktivit{\"a}t automatisiert berechnet und damit Reproduzierbarkeit und Vorhersagegenauigkeit erh{\"o}ht werden. Dies f{\"u}hrt zu Kosten- und Zeitersparnissen, wodurch eine h{\"o}here Akzeptanz der Methode im Praxiseinsatz zu erwarten ist und damit eine h{\"o}here Genauigkeit bei der Bestimmung dieses wichtigen diagnostisch relevanten Kriteriums erreicht werden kann.}, language = {de} } @thesis{Rinder2023, author = {Rinder, Marcel}, title = {Towards training spiking neural networks using local learning for speech recognition}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-39266}, pages = {55}, year = {2023}, abstract = {Neuromorphic hardware and Spiking Neural Networks have proven to be a beneficial alternative to traditional Deep Neural Network architectures. They leverage significant benefits in terms of energy efficiency and processing speed when compared to the execution on graphics cards. However, conventional backpropagation-based training techniques are suboptimal within the constraints of neuromorphic devices and are considered biologically implausible. The brain uses mostly local processes to update weights, yet the integration in spiking neural architectures remains unclear. Within the context of Speech Command Recognition, a Spiking Neural Network architecture with local learning is proposed. It uses output-class-specific dopamine for modulated Spike-Timing Dependent Plasticity, which allows for supervised learning without a backwards pass. Additional brain-inspired characteristics are incorporated, such as spiking neurons, Dale's Law, complex connectivity and homeostatic plasticity. The model achieves good accuracy on a simplified speech classification task and shows signs of generalization on a complex task.}, language = {en} } @thesis{Stroehlein2023, author = {Str{\"o}hlein, Dominik Ludwig Josef}, title = {Comparison of deep learning VSLAM approach vs. classic lidar camera fusion with a Unitree A1 robot.}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-43209}, pages = {55}, year = {2023}, abstract = {SLAM or Visual SLAM (Simultaneous Localization and Mapping) is a topic that is be- coming increasingly important in the digitalized world. In addition to the automotive industry and robotics, it can also be found in other areas such as argumented reality and even in the household, for example vacuum cleaner robots. VSLAM is therefore a topic that we often encounter unconsciously in everydays life. It is often not even possible to see whether and how exactly such algorithms work and what other components are required. In this thesis, this process, VSLAM, is examined more closely in terms of how it works. It also looks at how the individual components work and explains where they are currently used in the environment and where these systems have their strengths and weaknesses. Using a robot equipped with a stereo camera and a LiDAR sensor, data was recorded, evaluated and discussed in various locations of a university once complex. The recorded data is compared qualitatively and technically. The data from the VSLAM is primarily considered and the data from the LiDAR is used as ground truth and a basis for comparison. Another part of the thesis deals with the work with the robot, the sensors and the development of the code. Among other things, the problems that arose during the research are explained and their solutions and/or alternatives are shown. Finally, the results are compared with the current state of the art and an outlook for further research in relation to this work is given.}, language = {en} } @thesis{Mathyssek2023, author = {Mathyssek, Corinna}, title = {Cyberwar}, subtitle = {Angriffsszenarien und daraus resultierende technische Maßnahmen f{\"u}r Unternehmen und deren kritische Infrastruktur}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-41681}, pages = {iii, 89}, year = {2023}, abstract = {In dieser Abschlussarbeit wird auf die Existenz von Cyberwar eingegangen sowie die aktuelle Lage der IT-Sicherheit in Deutschland. Die Forschung bezieht sich auf die Maßnahmen und Strategien, die Unternehmen ergreifen k{\"o}nnen, um ihre IT-Infrastruktur bestm{\"o}glich vor Cyberangriffe zu sch{\"u}tzen.}, language = {de} } @thesis{Rampas2024, author = {Rampas, Dominic Antonio}, title = {On neural compression using diffusion autoencoders}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-44286}, pages = {65}, year = {2024}, abstract = {This work presents an examination of using diffusion models to achieve efficient data compression. As diffusion models have become more dominant in areas such as generative modelling, this work shows that they achieve a strong performance on the task of image-, and video-compression and outperform classical algorithms, as well as established neural compression algorithms. This type of model, referred to as Diffusion Autoencoders, is able to achieve spatial compression factors of 32x, while achieving high quality reconstructions that even maintain fine details of the input. The bachelor thesis presents a thorough analysis of individual components of the diffusion autoencoder through extensive experiments, while undermining the solidness of the approach by numerous comparisons and ablation studies. Additionally, an in-depth overview of related work for classical-, and neural-compression, diffusion models and evaluation methods is given.}, language = {en} } @thesis{Lachermeier2024, author = {Lachermeier, Tobias Sebastian}, title = {Development of a deep learning approach to reconstruct multiparametric magnetic resonance correlation spectra}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-45223}, pages = {40}, year = {2024}, abstract = {Magnetic resonance imaging (MRI) has assumed a central role in modern medical imaging thanks to its ability to provide detailed images of the inside of the body. This thesis aims to push the boundaries of conventional MRI by demonstrating two deep learning approaches to reconstruct multparametic correlation spectra. The multi-model approach determines each peak in the spectrum with a separate model, but does not generalise and requires knowledge of the number of spectra being searched for. The residual approach is an iterative process. The strongest signal is determined from the current residual and subtracted from the residual. The spectrum is created using the information obtained. This process is repeated until the residual no longer contains any information. A proof of concept was carried out with the restriction that the number of iterations was determined in advance. This thesis shows that the residual approach works and proves to be promising for future work.}, language = {en} } @thesis{Achter2024, author = {Achter, Lea}, title = {KI-basiert versus klassisch - File Carving in der digitalen Fahrzeugforensik}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-45885}, pages = {X, 45, X}, year = {2024}, abstract = {Durch die voranschreitende Entwicklung der Technologie im Bereich der Automobilindustrie k{\"o}nnen immer mehr digitale Spuren in Fahrzeugen festgestellt werden. Das f{\"u}hrt zu einem immer wichtiger werdenden Gebiet, der digitalen Fahrzeugforensik. Dieses besch{\"a}ftigt sich mit dem Auslesen der Fahrzeugspeicher, die meist propriet{\"a}re Dateitypen der Automobilhersteller enthalten. Somit ist der Aufbau dieser Dateien meist unbekannt, was einen Unterschied zur digitalen Forensik darstellt. Hier k{\"o}nnen durch sogenannte File Carver Dateitypen, anhand bekannter Byte Sequenzen, wie Header oder Footer, erkannt werden. Unbekannte propriet{\"a}re Dateien, wie die der Automobilindustrie k{\"o}nnen somit meist nicht gefunden werden. Das Ziel dieser Arbeit ist es zu untersuchen, in wieweit klassische File Carver die spezifischen Dateitypen der Automobilbranche erkennen, und ob KI-basierte Ans{\"a}tze hier m{\"o}glicherweise einen Vorteil bieten k{\"o}nnen. Hierzu wird ein synthetischer Datensatz erstellt, um eine Basis mit relevanten Dateitypen zu schaffen. Die Tests der Softwares auf dem erstellten Datensatz zeigen, dass sich der File Carver Autopsy am Besten f{\"u}r eine Untersuchung in der digitalen Fahrzeugforensik eignet. Jedoch lassen die erzielten Ergebnisse der KI-basierten Methoden auf ein deutliches Entwicklungspotential schließen.}, language = {de} } @thesis{Schlund2024, author = {Schlund, Johannes Reinhold}, title = {KI-gest{\"u}tztes Fibertracking in der MR-Bildgebung von Multipler Sklerose}, publisher = {Technische Hochschule Ingolstadt}, address = {Ingolstadt}, url = {http://nbn-resolving.de/urn:nbn:de:bvb:573-45671}, pages = {53}, year = {2024}, abstract = {In der medizinischen Bildgebung ist die Aufnahme durch Magnetresonanztomographie (MRT) wohl eine der g{\"a}ngigsten Methoden, wenn es um die Informationsgewinnung von Gewebestrukturen geht. Vor allem im Gehirn, wo jeder Eingriff ein hohes gesundheitliches Risiko birgt, bietet das MRT eine sichere Methode, dieses zu analysieren. Als Erweiterung davon kann mittels Diffusions-Bildgebung auch die Bewegung von Protonen im Gewebe gemessen und daraus der Verlauf von Nervenfasern rekonstruiert werden. Diese Methode ist als Traktographie oder auch Fiber Tracking bekannt und kann {\"u}ber verschiedene Algorithmen umgesetzt werden. In dieser Arbeit soll der Fokus auf Fibertracking mit KI-gest{\"u}tzten Methoden gesetzt werden, um eine Pipeline zu schaffen, die auf beliebige rohe Diffusionsdaten anwendbar ist. Der Aufbau setzt sich dabei aus verschiedenen Schritten zusammen. Zun{\"a}chst werden die Daten vorverarbeitet, um einen Ground Truth zu erzeugen, der die ben{\"o}tigten Koeffizienten zur Traktographie beinhaltet. Diese k{\"o}nnen dann im zweiten Schritt von einem Transformer-basierten KI-Modell trainiert werden. Damit soll im letzten Schritt dann aus den gesch{\"a}tzten Werten die eigentliche Traktographie umgesetzt werden. Mit der Verwendung einer gut trainierten KI-Architektur soll so Zeit und Rechenleistung im gesamten Prozess eingespart werden k{\"o}nnen, da das Modell aufwendige Algorithmen zur Berechnung der Koeffizienten ersetzen kann. Es soll dabei gezeigt, werden wie die komplette Anwendung im Trainingsablauf aussieht, und welche Anpassungen und Optimierungen m{\"o}glich sind. Die resultierende Code-Pipeline soll den Grundstein f{\"u}r weitere Forschung an Daten von MS-Patienten legen.}, language = {de} }