PhD Oral Exam - Somayeh Davar, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Brain tumors are among the most challenging neurological diseases due to their heterogeneous appearance, irregular morphology, and potentially life-threatening consequences. Magnetic Resonance Imaging (MRI) is widely used for brain tumor assessment because of its excellent soft-tissue contrast and non-invasive nature. Accurate tumor segmentation from MRI is essential for computer-aided diagnosis, treatment planning, and disease monitoring. However, manual delineation is labor-intensive, time-consuming, and subject to inter-observer variability. Although deep learning has significantly advanced automated brain tumor segmentation, conventional full-image approaches require segmentation networks to identify tumor regions while simultaneously delineating their boundaries, increasing the segmentation search space, class imbalance, and background interference, particularly for relatively small tumor regions. This dissertation investigates the hypothesis that improving tumor localization and localization-derived guidance can facilitate more accurate and efficient brain tumor segmentation. A series of detection-guided deep learning frameworks is developed in which each successive model addresses limitations identified in the preceding study while introducing increasingly informative localization guidance. The research therefore evolves systematically from coarse localization to region-of-interest (ROI) processing, detector-guided spatial attention, and confidence-aware spatial guidance.

The proposed research begins with an R-CNN-based localization framework integrated with a modified U-Net for tumor localization and segmentation. The framework subsequently replaces region-proposal-based detection with more efficient YOLO-based localization, progressing through YOLOv3 and YOLOv8 while enhancing the segmentation network through architectural refinements, improved feature representation, and attention mechanisms. The next stage introduces YOLOv11-based localization and automatic ROI extraction to restrict segmentation to the detected tumor region. Detector-derived binary bounding-box attention is then incorporated into an Attention U-Net to provide explicit spatial guidance. Building on this progression, the dissertation culminates in CASA-Net, which introduces a Confidence-Aware Spatial Attention (CASA) module that transforms detector confidence into a continuous spatial prior, enabling the segmentation network to exploit both tumor location and the confidence associated with the detector prediction. Collectively, the proposed frameworks demonstrate that progressively strengthening localization guidance can support more focused and accurate brain tumor segmentation. Beyond the specific frameworks investigated in this dissertation, the proposed detection-guided paradigm provides a foundation for future research on confidence-aware segmentation, multimodal medical image analysis, medical foundation models, and related medical imaging applications.