Institutional Repository of Vilnius University
Smegenų auglių segmentavimas su konvoliuciniais neuroniniais tinklais /
Abstract
dc:descriptionMachine learning is a methodology that is being applied in an increasing number of fields. One of those fields is medicine, where one of many applications is the automatic segmentation of various objects. Manually annotating the area of a tumor is a challenging task. Due to imaging noise and resolution limitations, the manual segmenting of the tumor is shown to vary between experts. The purpose of creating automatic models is to improve both the time and the accuracy of the process, serving as a valuable tool for doctors. Deep convolution neural networks (CNNs) have shown great promise in image segmentation, with the symmetric U-Net architecture being a major breakthrough. During supervised learning epochs, CNNs utilize hierarchical convolutional kernels to extract spatial patterns across multiple scales. The U-Net architecture became state-of-the-art in segmentation for its encoding and decoding paths, which allow the model to focus both on fine-grain details and general trends. Skip connections implemented in the U-net prevent these details from being overlooked, ensuring high levels of automatic segmentation. In this work, the U-Net framework is extended by adding Attention Gates and Residual blocks to further improve the focus on tumor boundaries and mitigate vanishing gradients. This thesis specifically applies this architecture to the segmentation and classification of meningiomas, gliomas, and pituitary tumors in 2D MRI images.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Širov, Aleksandr,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308327115&prefLang=en_US