Deep Learning Architectures for Precise Segmentation of Brain Tumors in MRI Images.
Loading...
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
University of Msila
Abstract
This master’s thesis addresses the critical challenge of precise brain tumor segmentation
in multi-modal MRI images. Manual segmentation, the current clinical standard,
is hindered by its time-consuming nature and inter-observer variability. To overcome
these limitations, this research investigates the application of advanced deep learning
architectures for automated and accurate tumor delineation.
The study systematically evaluates and compares a range of models, from a baseline
U-Net to state-of-the-art architectures such as FPN, DeepLabV3+, and U-Net++, enhanced
with pre-trained encoders like SE-ResNet50 and EfficientNet to leverage transfer
learning. Furthermore, ensemble methods, specifically Voting and Stacking, were
implemented to boost segmentation performance.
The results demonstrate that the Feature Pyramid Network (FPN) with an EfficientNet-
B4 backbone achieved the best performance among single models, with a Dice Score
of 0.8235, offering an excellent balance between accuracy and efficiency. However, the
Stacking Ensemble model proved superior to all individual models, attaining the highest
Dice Score of 0.8263 by effectively combining predictions from diverse base models.
This research concludes that while the Stacking Ensemble provides the optimal solution
for maximizing accuracy, the FPN architecture stands as a highly competitive and
practical alternative for clinical applications
Description
Keywords
Brain Tumor Segmentation, Magnetic Resonance Imaging (MRI), Deep Learning, Convolutional Neural Networks (CNN), U-Net, FPN, Ensemble Methods, Stacking.