Machine Learning and Multi-Modal Image Analysis for Image-Guided Therapy and Clinical Decision Support
|PhD_Thesis_JNitschPDFA.pdf||Dissertation Jennifer Nitsch||51.87 MB||Adobe PDF||View/Open|
|Authors:||Nitsch, Jennifer||Supervisor:||Kikinis, Ron||1. Expert:||Kikinis, Ron||2. Expert:||Preim, Bernhard||Abstract:||
This PhD thesis makes two different contributions to the field of medical image analysis.
The first is a novel approach for using segmentation-based image registration in the
field of real-time image-guided glioma surgery. This was the first approach demonstrating that a segmentation-based registration of preoperative magnetic resonance imaging (MRI) with intraoperative ultrasound (iUS) imaging is more accurate, more robust, and faster than purely intensity-based registration. For the registration, central cerebral structures were segmented and served as a guiding-frame. Moreover, the intraoperative applicability of MRI-iUS registration within a strict time frame during glioma resection could be demonstrated with an overall computation time of about 91 seconds, which also included computing the segmentations.
The second contribution is a proof of concept of a novel automatic image-based severity assessment of liver cirrhosis for end-stage liver disease patients. Currently, the degree of severity is determined by a blood test from which the so-called MELD score is calculated. This score is also used for ranking patients on the liver transplant list. For an image-based assessment, quantitative hepato-splenic radiomic features are extracted and analyzed to determine their importance for classifying the severity of liver cirrhosis. Different clinically established metrics for assessing the disease severity were employed and compared with extracted radiomic features. Furthermore, a control group could be identified that was scanned with the exact same MR imaging protocol. This allowed a direct comparison of radiomic features extracted from cirrhotic patients with those from non-cirrhotic patients. Apart from automatic cirrhosis detection, important findings can be reported from our studies that are novel in this field. Some findings are also in line with perceptions of radiologists but were not quantified before. To the best of my knowledge, no other research group has previously investigated the specific radiomic features for risk-stratification of cirrhotic patients and for automatic cirrhosis classification.
|Keywords:||Medical Image Computing; Deep Learning, Invertible Neural Networks, Adversarial Examples, Imaging Mass Spectrometry, Normalizing Flows; Radiomics; Segmentation; Registration; Image-Guided Neurosurgery; Hepatology; Cirrhosis; Biomarker||Issue Date:||27-Feb-2020||DOI:||10.26092/elib/36||URN:||urn:nbn:de:gbv:46-elib42521||Institution:||Universität Bremen||Faculty:||FB3 Mathematik/Informatik|
|Appears in Collections:||Dissertationen|
checked on Sep 30, 2020
checked on Sep 30, 2020
This item is licensed under a Creative Commons License