Date of Award

Summer 8-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Program/Concentration

Computer Science

Committee Director

Jiangwen Sun

Committee Member

Yaohang Li

Committee Member

Sampath Jayarathna

Committee Member

Mahadev Satyanarayanan

Abstract

Accurate glioma segmentation from multi-modal Magnetic Resonance Imaging (MRI) is essential for diagnosis, treatment planning, surgical guidance, radiation targeting, and longitudinal monitoring. MRI modalities such as T1, T1Gd, T2, and FLAIR provide complementary information for identifying clinically important tumor sub-regions, including enhancing tumor (ET), tumor core (TC), and whole tumor (WT). However, many existing segmentation methods do not fully preserve modality-specific information, process all modalities and slices uniformly, and often rely on a single shared fusion strategy for all sub-regions. These limitations can reduce segmentation accuracy, increase computational cost, and limit clinical interpretability.

This dissertation addresses these challenges through three adaptive frameworks for multimodal glioma segmentation. First, AIMS introduces a hybrid CNN–Transformer architecture with modality-specific feature extraction, adaptive self-attention, and late fusion to improve segmentation of ET, TC, and WT. Second, AIMS-Select extends this framework by learning budget-aware modality–slice gates that select informative inputs before expensive 3D CNN and Transformer encoding, reducing redundant computation while maintaining strong segmentation performance. Third, a region-specific modality gating framework reduces task interference by learning separate modality fusion strategies for ET, TC, and WT.

Experiments on BraTS 2019 and BraTS 2021 demonstrate that the proposed frameworks improve segmentation accuracy, computational efficiency, robustness, and interpretability. Attention maps, Grad-CAM visualizations, modality-importance weights, and gating patterns provide insight into how the models use MRI modalities and focus on tumor-relevant regions. Together, these contributions advance adaptive and interpretable multi-modal glioma segmentation and support the development of reliable AI-assisted tools for neuro-oncology.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

DOI

10.25777/hrnp-s758

ISBN

9798193214366

Download

Share

COinS