Reversible Architectures for Memory-Efficient Brain Tumor Segmentation
Alnaqbi, Maitha Abdelhakeem Yousif Almushtaghil
Alnaqbi, Maitha Abdelhakeem Yousif Almushtaghil
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 addresses the challenge of high memory consumption in 3D medical image segmentation, focusing on glioma segmentation from multiparametric brain MRI. Deep learning models used in this domain, particularly for volumetric data, often face computational bottlenecks that limit their scalability. To mitigate this constraint, this work investigates the integration of reversible architectures into segmentation networks. Specifically, it proposes RevMedNeXt, a memoryefficient variant of the MedNeXt architecture, in which encoder blocks are replaced with reversible counterparts. These blocks reconstruct intermediate activations during backpropagation, enabling constant memory scaling regardless of model depth. RevMedNeXt is trained and evaluated on the BraTS 2023 Adult Glioma dataset, (Abstract full version available in the thesis document.)
Citation
Maitha Abdelhakeem Yousif Almushtaghil Alnaqbi, “Reversible Architectures for Memory-Efficient Brain Tumor Segmentation,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Glioma, Segmentation, Reversible, Efficiency, Memory