Refine
Document Type
- Master's Thesis (2)
- Bachelor Thesis (1)
Has Fulltext
- yes (3)
Is part of the Bibliography
- no (3) (remove)
Keywords
- Alzheimer (1)
- Bildgebung (1)
- Hippocampus (1)
- Maschinelles Lernen (1)
- Neuronales Netz (1)
- Software (1)
- Strahlentherapie (1)
- VR-Anwendung (1)
- VR-Applikation (1)
- bildgebende Verfahren (1)
Institute
- Informatik (3)
Im Mittelpunkt dieser Arbeit steht die Entwicklung eines Verfahrens für das Erzeugen
von möglichst realitätsnahen Telepräsenzsimulationen für die Information von Patienten
in der Strahlentherapie sowie die Anwendung des Verfahrens in der Entwicklung
einer VR-Applikation auf Basis eines Demonstrators.
Nach einer Literaturrecherche bezüglich des aktuellen Stands der Aufklärung von Patienten
wurden die Grundlagen von Virtual und Augmented Reality hinsichtlich der
gegebenen Hardware ermittelt. Hierauf folgte die Auswahl von Software für das Scannen
von realen Objekten mit dem gegebenen Tablet in einem Bestrahlungsraum des
DKFZ sowie die Auswahl einer Game Engine für die Entwicklung des Demonstrators.
Daraufhin wurde ein Rekonstruktionsalgorithmus ausgewählt. Anschließend wurden
verschiedene Objekte im Bestrahlungsraum gescannt, sodass die Parameter des Algorithmus
iterativ hinsichtlich der Qualität der erzeugten Objekte für den Einsatz in
einer VR-Anwendung optimiert werden konnten. Daraufhin erfolgte eine Texturierung
der Oberfläche mit Kamerafotos. Nach einer Aufbereitung der Modelle wurden diese
in ein Virtual Environment importiert. Parallel dazu wurde nach der Auswahl der Unreal
Engine als Game Engine, der Demonstrator entwickelt, in welchen die gescannten
Modelle integriert wurden.
Das Verfahren liefert ausreichend genaue Ergebnisse, um Konzepte in der Strahlentherapieaufklärung
vermitteln zu können. Der Effekt und die Akzeptanz der Technik
spielen eine weitere wichtige Rolle für den Einsatz der Methodik und müssen durch eine
Evaluation im klinischen Alltag validiert werden, wofür die Entwicklung einer klinisch
anwendbaren Software auf Basis der gewonnenen Erkenntnisse notwendig wird.
Medical imaging produces many images every day in clinical routine. Keeping up with the
daily image analysis task and this vast amount of data is quite a challenge for radiologists.
However, these analysis tasks can be automated with well-proven automatic segmentation
methods. Segmentation reviewing of an expert is necessary because learningbased
automatic segmentation methods may not perform well on exceptional image
data. Creating valid segmentations by reviewing them also improve the learning-based
methods.
Combining established standards with modern technologies creates a flexible environment
to efficiently evaluate multiple segmentation algorithm outputs based on different metrics
and visualizations and report these analysis results back to a clinical system environment.
The presented software system can inspect such quantitative results in a fast and intuitive
way, potentially improving the daily repetitive segmentation review and rework of a
research radiologist. The presented system is designed to be integrated into a virtual
distributed computing environment with other systems and analysis methods. Critical
factors for this particular environment are the handling of many patient data and routine
automated analysis with state of the art technology.
First experiments show that the time to review automatic segmentation results can be
roughly divided in half while the confidence of the radiologist is enhanced. The system
is also able to highlight individual slices which are essential for the expert’s review
decision. For this highlighting, different metric scores are compared and evaluated.
Alzheimer’s Disease affects millions of people worldwide, but till today, the gold standard
for definitive diagnosis of this disease is a biopsy. Nevertheless, with the progress
of the disease, a volume loss in the Hippocampus can be observed. Therefore, good
segmentation methods are crucial to facilitate quantification of this loss.
The focus of this work is on the development of a Machine Learning algorithm, more
precisely a Generative Adversarial Network, for the automated segmentation of the
human Hippocampus and its substructures in Magnetic Resonance Images. In particular,
the task is to determine if the integration of a pre-trained network that generates
segmentations into a Generative Adversarial Network scheme can improve generated
segmentations. In this context, a segmentation network in form of a U-net corresponds
to the generator. The discriminator is developed separately and merged in a second
step with the generator for combined training.
With a literature review regarding the automated segmentation of the Hippocampus,
current methods in this field and their medical and technological basics were identified.
The datasets were preprocessed to make them suitable for the use in a neural
network. In the training process, the generator was trained first until convergence.
Then, the Generative Adversarial Network including the pre-trained generator was
trained. The outcomes were evaluated via cross-validation in two different datasets
(Kulaga-Yoskovitz and Winterburn). The Generative Adversarial Network scheme
was tested regarding different architectural and training aspects, including the usage
of skip-connections and a combined loss function.
The best results were achieved in the Kulaga-Yoskovitz dataset with a Dice coefficient
of 90.84 % after the combined training of generator and discriminator with a joined
loss function. This improves the current state of the art method in the same task and
dataset with a Dice index of 88.79 % by Romero [Rom17]. Except of two cases in the
Winterburn dataset, the proposed combined method could always improve the Dice
results after the training of only the generator, even though only by a small amount.