Mathematical and Computational Convolutional Neural Network Methods for MRI Segmentations

Simple item page
AuthorAlsaeed, Huda Jaffer
metadata.dc.contributor.departmentUNIVERSITY OF BAHRAIN COLLEGE OF SCIENCE DEPARTMENT OF MATHEMATICS
metadata.dc.date.accessioned2026-09-20T15:51:23Z
Publication date2025
Degree levelMaster
AbstractAbstract : Segmenting brain tumors using magnetic resonance imaging (MRI) is essential for effective clinical diagnosis and treatment planning. Conventional segmentation techniques, such as thresholding, region-growing algorithms, and active contour models, face considerable challenges. These methods struggle with brain tumor segmentation because they require manual input and struggle with variations in MRI intensity. They also fail to capture the tumor's irregular shapes and indistinct edges due to limited morphological feature extraction. Additionally, traditional methods often handle 2D slices separately, overlooking vital 3D contextual information crucial for precise volumetric analysis of brain tumors. Deep learning has transformed medical image segmentation by overcoming various constraints. Convolutional Neural Networks (CNNs), especially fully convolutional architectures such as U-Net and V-Net, have shown impressive achievements in automating the segmentation of brain tumors. These deep learning methods are adept at learning hierarchical features directly from data, managing 3D volumetric information, and ensuring spatial consistency across slices—all crucial skills for tackling the complexities of brain tumor segmentation. This study provides an in-depth comparison of two deep learning architectures for brain tumor segmentation: 3D U-Net and V-Net. Using the BraTS 2020 dataset for evaluation, our 3D U-Net implementation achieved a segmentation accuracy of 98.26 percent with an IoU of 0.6792. In contrast, V-Net outperformed with 98.65 percent accuracy and an IoU of 0.7330. The findings highlight V-Net's particular strengths in addressing tumor boundary uncertainties through its residual connections and advanced feature learning abilities. Compared to traditional methods, both deep learning models offer notable improvements in processing speed, consistency, and the ability to capture complex tumor characteristics without manual intervention. The findings highlight the significant impact of deep learning in neuro-oncology imaging, providing more dependable and efficient tools for clinical decision-making. Keywords: Brain tumor segmentation, MRI, traditional methods, deep learning, 3D U-Net, V-Net, convolutional neural networks
metadata.dc.identifier.otherDrupal node 4642
metadata.dc.identifier.otherhttps://digitalrepository.uob.edu.bh/en/dar/mathematical-and-computational-convolutional-neural-network-methods-mri-segmentations
metadata.dc.identifier.urihttp://165.227.56.122/handle/123456789/7149
Languageen
Linkhttps://digitalrepository.uob.edu.bh/en/_flysystem/fedora/2026-02/002462_mathematical_and_computational_convolutional_neural_network_methods_for_mri_segmentations.pdf
TitleMathematical and Computational Convolutional Neural Network Methods for MRI Segmentations
TypeMaster's Thesis
CollegeCollege of Science
DepartmentMathematics

Files

Original bundle

Now showing 1 - 1 of 1
Thumbnail Image
Name:
node-4642-1-002462_mathematical_and_computational_convolutional_neural_network_methods_for_mri_segmentations.pdf
Size:
8.09 MB
Format:
Adobe Portable Document Format
Download

Collections

Master's Theses