Date of Award
2026
Document Type
Open Access Master's Thesis
Degree Name
Master of Science in Biomedical Engineering (MS)
Administrative Home Department
Department of Biomedical Engineering
Advisor 1
Jingfeng Jiang
Committee Member 1
Hoda Hatoum
Committee Member 2
Weihua Zhou
Abstract
Accurate characterization of intracranial aneurysm wall-related regions is important for imaging-based risk assessment, yet consistent delineation of these thin peripheral structures remains challenging in high-resolution magnetic resonance vessel wall imaging (HR-MR-VWI). This thesis establishes a framework for evaluating and improving automatic aneurysm wall segmentation in quantitative imaging analysis through two closely related lines of investigation. On the segmentation side, five deep learning models spanning three architectural paradigms—conventional CNN-based models (KiUNet, nnUNet, and ARUNet), a diffusion-based generative framework (DiffUNet), and a Mamba-enhanced variant (DiffUNet-Mamba)—were implemented and compared on 100 clinical cases. Wall mask Dice scores ranged from 0.394 to 0.599, with DiffUNet-Mamba achieving the highest performance. On the evaluation side, a synthetic Dice benchmark constructed from controlled mask perturbations revealed a non-linear relationship between geometric overlap and radiomics reliability: quantitative usability was limited below Dice 0.5 but improved rapidly within the 0.5–0.7 interval. Feature-level analysis identified 39 computationally robust radiomics features, of which 18 were independently validated for clinical risk stratification. Clinical validation demonstrated that DiffUNet-Mamba reproduced 91.4% of the 58 GT-significant features, and the overall feature–clinical association patterns showed high concordance with GT (PCC > 0.96, CCC > 0.94). These findings demonstrate that automatic aneurysm wall segmentation with moderate geometric accuracy can yield radiomics measurements that are quantitatively stable, consistent with GT, and capable of preserving clinically relevant association patterns.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Yu, Teng, "Automatic Segmentation Accuracy and Radiomics-Based Quantitative Reliability in High-Resolution Magnetic Resonance Vessel Wall Imaging of Intracranial Aneurysms", Open Access Master's Thesis, Michigan Technological University, 2026.
Included in
Artificial Intelligence and Robotics Commons, Bioimaging and Biomedical Optics Commons, Cardiovascular Diseases Commons, Diagnosis Commons, Radiology Commons