Show simple item record

dc.rights.licenseCC-BY-NC-ND
dc.contributor.advisorKuijf, Hugo
dc.contributor.authorBos, Lonneke
dc.date.accessioned2022-09-09T04:03:12Z
dc.date.available2022-09-09T04:03:12Z
dc.date.issued2022
dc.identifier.urihttps://studenttheses.uu.nl/handle/20.500.12932/42747
dc.description.abstractMultiple Sclerosis (MS), a neuroinflammatory disease of the central nervous system, is characterized by lesions in the brain. Lesions are inflammations in the brain and are visible in Magnetic Resonance images as void shaped hyperintensities after administration of a Gadolinium-based contrast agent (GBCA’s). Artificial intelligence (AI) methods have been used among other researchers, in MS lesion segmentations; however, the existing methods have limitations making them difficult to be applicable to different MR scanners. Likewise, the goal of this research is to segment lesions on MR images using deep learning. Deep learning is an artificial intelligence method where data processing is are used to extract features from data. Features of lesions are for example their shape, intensity, and location in the brain. The deep learning method used in this research is called a V-net, which is a 3D structure that is suitable for volumetric image segmentations. A deep learning model is first trained with a dataset, here it learns all the features needed for lesion segmentation. Subsequently, the trained model is tested with a similar dataset to test how well it performs. To acquire the dataset two annotators manually segmented lesions in MR data and fed it to the model. The goal of the deep learning model is to predict an accurate lesion segmentation. Analysis of the segmentation quality before and after discussion between the two annotators concluded that this increased the quality of the delineation and therefore the model performance. A model can be trained with different number of epochs, this is defined as the number of times that the algorithm will work through the dataset. The model is analyzed with 50 epochs (shallow training) and 150 epochs (extensive training). The accuracy of the lesion segmentation is increased when the model is shallow trained, compared to extensive training. The quality of the model when used on a completely new, unseen dataset was tested and it showed that the model is not good enough to predict lesions when data with other MR protocols are used. Additionally, based on the findings of this work, there are preliminary findings that using different MR scanners can influence the model performance. The method can be improved by using a larger dataset. Summarizing, this deep learning model stands a proof of concept for lesion segmentation using limited data.
dc.description.sponsorshipUtrecht University
dc.language.isoEN
dc.subjectI did research about a new method to segment enhancing lesions in MS in T1w data using deep learning.
dc.titleEnhancing lesion segmentation on contrast MR images in Multiple Sclerosis using deep learning
dc.type.contentMaster Thesis
dc.rights.accessrightsOpen Access
dc.subject.courseuuMedical Imaging
dc.thesis.id10374


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record