PhD Oral Exam - Dalia Alzu’bi, Information and Systems Engineering

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

Accurate delineation of anatomical structures, organs, and pathological regions is essential for computer-aided diagnosis across diverse medical imaging modalities. Despite significant advances, existing segmentation models often face challenges in balancing local feature representation, global contextual understanding, and computational efficiency. This thesis presents effective approaches for medical image segmentation that progressively enhance feature representation and contextual understanding to improve segmentation accuracy and generalization while maintaining computational efficiency. The thesis contributions are threefold. First, we introduce U-CycleMLP, a novel U-shaped encoder-decoder architecture that integrates dense atrous convolutions, attention-weight excitation mechanisms, and channel CycleMLP modules to bridge spatial awareness and global contextual modeling. It achieves an effective balance between segmentation performance and computational efficiency. Second, building on this foundation, we develop AdaKAN, an Adaptive Kolmogorov-Arnold Network framework that introduces a dual-branch AdaptKAN module and an efficient attention mechanism. AdaKAN outperforms strong baselines across several benchmarks, while reducing the number of learnable parameters by 43.2% compared with U-CycleMLP. Third, to further enhance nonlinear representation learning, we propose KAAN, a Kolmogorov-Arnold Attention Network integrating global-local attention and dual width-wise and height-wise attention pathways. Compared with AdaKAN, KAAN achieves superior performance while reducing learnable parameters by 51.2% and computational overhead by 64%. Collectively, these contributions establish a progressive evolution of medical image segmentation architectures, advancing from efficient contextual modeling to adaptive and attention-guided learning frameworks. The results demonstrate that accurate segmentation, enhanced nonlinear representation learning, and computational efficiency can be achieved simultaneously, providing a strong foundation for medical image segmentation.