Author, Institution: Modupe Olufunke Odusami, Kaunas University of Technology
Science area, field of science: Technological Sciences, Informatics Engineering, T007
Scientific Supervisor: Prof. Dr. Rytis Maskeliūnas (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)
Dissertation Defense Board of Informatics Engineering Science Field:
Prof. Dr. Nikolaj Goranin (Vilnius Gediminas Technical University, Technological Sciences, Informatics Engineering, T007) – chairperson
Prof. Dr. Hab. Romualdas Baušys (Vilnius Gediminas Technical University, Technological Sciences, Informatics Engineering, T007)
Prof. Dr. Vytautas Galvanauskas (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)
Assoc. Prof. Dr. Antonio Martinez-Millana (Polytechnic University of Valencia, Spain, Technological Sciences, Informatics Engineering, T007)
Assoc. Prof. Dr. Tomas Tamošuitis (Lithuanian University of Health Sciences, Medical and Health Sciences, Medicine, M001)
Dissertation defense meeting will be at Rectorate Hall of Kaunas University of Technology (K. Donelaičio 73–402, Kaunas)
The doctoral dissertation is available at the library of Kaunas University of Technology (Gedimino g. 50, Kaunas)
Annotation: This thesis develops a novel methodology based on deep learning to improve the neuroimaging fusion to classify Alzheimer’s disease stages. To address the issue of inconsistence specificity and sensitivity across binary classification of single imaging techniques, such as structural MRI (sMRI) and fluorodeoxyglucose positron-emission tomography (FDG-PET), which respectively emphasize structural and metabolic change. These modalities are combined for greater coverage of AD’s structural-functional and structural-metabolic progression. This thesis advances the fusion of neuroimaging modalities by adopting a feature-level fusion approach, preferred for its ability to combine essential details from each modality. It applies multiscale transforms, to MRI and FDG-PET images to decompose them into multiple frequency components. While multiscale transformations are crucial for isolating high- and low-frequency details, they risk degrading feature quality. To mitigate this, feature map clarity is enhanced with transposition convolution based on the VGG19 network and harmonizes cross-modal data using instance normalization. A key innovation in this thesis is the use of colorization with a cosine mapping method, where grayscale fused images are then converted into colorization forms emphasizing on distinctive pattern in multimodal data. In addition, a lightweight Mobile Vision Transformer with a Swish activation function was used as the classification model, optimized for handling the intricacies of colorized fused images. To validate the model’s robustness, it has been tested on multiple datasets, demonstrating strong generalizability and performance across varied patient demographics and imaging protocols.
December 9 d. 10:00
Rectorate Hall at Kaunas University of Technology (K. Donelaičio 73-402, Kaunas)
Įtraukti į iCal
