Medical Image Segmentation: A Computationally Efficient Adaptation of ShaSpec for Brain Tumors
Alblooshi, Khalil Ibrahim Hasan Mohammed
Alblooshi, Khalil Ibrahim Hasan Mohammed
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This thesis proposes a computationally efficient adaptation of the ShaSpec architecture for brain tumor segmentation in multimodal MRI. By merging four separate modality-specific encoders (T1, T1ce, T2, FLAIR) into two unified encoders using a "model soups" weight-averaging approach, and removing the dedicated domain classifier, the modified model achieves good reductions in computational complexity (~7% fewer GFLOPs and parameters). Evaluated on the BraTS18 dataset, segmentation accuracy remains comparable to the original model but dropped in one category (Dice scores: WT 0.908 vs. 0.909 baseline; TC improved to 0.897 vs. 0.855 baseline; ET 0.68 vs. 0.78 baseline), this ensure a simplified design that facilitates faster inference and easier clinical deployment without sacrificing accuracy.
Citation
Khalil Ibrahim Hasan Mohammed Alblooshi, “Medical Image Segmentation: A Computationally Efficient Adaptation of ShaSpec for Brain Tumors,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Medical Image Segmentation, ShaSpec Architecture, Brain Tumors, Computational Efficiency, Multi-modal MRI, Model Soups