TY - THES A1 - Stolz, Roland T1 - Classification Of 3D MRI Head Scans using pretrained CNNS – comparison, optimization and evaluation of approaches N2 - This work focuses on providing a convolutional neural network (CNN) architecture for the classification of 3D MRI head scans. In order to offer a comprehensible recommendation for an architecture, this work has four main objectives: 1. Implementation of two CNN architectures based on transfer learning, for the classification of 3D MRI head scans 2. Optimization of both architectures to provide a fair comparison 3. Comparison of both architectures with relevant criteria for the usage of the CNNs in medical diagnosis 4. Evaluation of the effects of transfer learning on both architectures KW - Artificial Intelligence KW - medical imaging KW - MRI KW - Künstliche Intelligenz KW - Kernspintomografie KW - Bildgebendes Verfahren Y1 - 2020 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:860-opus4-2005 ER - TY - THES A1 - Greipel, Julian T1 - Deep learning registration of non-rigid objects in medical AR applications - optimization of methods and data N2 - The goal of this thesis is to identify and evaluate parameters within a medical registration dataset that impact registration accuracy in a medical augmented reality application. To achieve this goal, • a realistic synthetic medical dataset is generated using physics-based simulations, • a state of the art deep-learning non-rigid point cloud registration algorithm is implemented, • relevant factors in the dataset are evaluated with regards to their impact on registration accuracy. KW - Registration KW - Augmented Reality KW - Artificial Intelligence KW - Liver Surgery KW - Point clouds Y1 - 2024 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:860-opus4-3956 ER - TY - THES A1 - Reinfurt, Guido T1 - Deep Learning methods for partial to whole object registration in mixed reality guided liver surgery - preliminary experiments N2 - This work primarily focuses on improving registration procedures within augmented reality (AR) applications, by analyzing the reliability of point cloud reconstruction as a preprocessing step. Therefore, a state-of-the-art, transformer base network is utilized. For the training process a comprehensive dataset specifically designed for training deep learning models is created. This dataset will be tailored to the task of reconstructing partial liver views for augmented reality guidance in medical applications. The validation process for this dataset will be two-fold: • Deep Learning Network Training: The created dataset will be used to train a state-of-the-art deep learning network, specifically a transformer-based model, the “PoinTr”-network. This training will evaluate the dataset's effectiveness in facilitating accurate point cloud reconstruction. • Testing with Real-World Scenarios: A separate testing dataset, meticulously designed to reflect various real-world conditions encountered in AR guidance, such as variations in point cloud density, occlusions, and sensor noise, will be used to assess the performance of the trained model. This evaluation will analyze the model's ability to generate accurate reconstructions under diverse and potentially challenging scenarios. KW - Registration KW - Augmented Reality KW - Artificial Intelligence KW - Liver Surgery KW - Point clouds Y1 - 2024 U6 - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:860-opus4-3943 ER -