프로톤 자기공명영상 혈관조영술의 개발 및 임상 적용 = Development and clinical application of Proton-density magnetic resonance angiography

다국어 초록 (Multilingual Abstract)

High-resolution proton-density weighted imaging (HR-PD) offers superior vessel wall visualization compared to conventional time-of-flight magnetic resonance angiography (TOF-MRA). However, its clinical adoption is hindered by the lack of automated post-processing for three-dimensional (3D) volume rendering (VR). This study aims to develop an automated deep learning-based PD-MRA system and validate its diagnostic performance for detecting unruptured intracranial aneurysms. A semi-automated labeling strategy integrating HR-PD and TOF-MRA was developed to generate ground truth data. A 3D nnU-Net model was trained using a development cohort of 50 patients (40 training, 10 testing). For clinical validation, 125 independent patients with 368 unruptured intracranial aneurysms confirmed by digital subtraction angiography (DSA) were enrolled. The diagnostic performance of the automated PD-MRA VR was compared with TOF-MRA VR. Two readers assessed image quality and one neuroradiologist diagnostic accuracy (sensitivity, specificity, and area under the curve [AUC]). The deep learning model achieved a high Dice similarity coefficient of 0.87 ± 0.02. In the clinical validation cohort, the image quality of PD-MRA VR was comparable to TOF-MRA (mean score > 2.8/3.0, p > 0.05). For aneurysm detection, PD-MRA VR alone demonstrated an AUC of 0.764, which was comparable to TOF-MRA VR (AUC 0.768, p = 0.108). Notably, when reviewing PD-MRA VR combined with source images, the diagnostic performance significantly improved (AUC 0.819, p<0.001), showing potential superiority over the TOF-MRA protocol. We successfully developed the first automated deep learning-based PD-MRA system capable of generating high-quality 3D visualizations. The system demonstrated diagnostic performance comparable to the current standard TOF-MRA. This automated pipeline overcomes the main technical barrier of HR-PD, suggesting its feasibility and utility in routine clinical practice for aneurysm screening and evaluation. Keywords: Intracranial aneurysm; Magnetic resonance angiography; Deep learning; Proton-density weighted imaging; Vascular segmentation.

목차 (Table of Contents)

• 1. Introduction 1
• 1.1 Clinical Significance and Diagnostic Challenges 1
• 1.2 Current Imaging Modalities and Limitations 1
• 1.3 Objectives 2
• 2. Materials and Methods 3
• 2.1 Study population 3
• 2.1.1 Development Cohort 3
• 2.1.2 Clinical application Cohort 3
• 2.2 Image Acquisition 6
• 2.3 Image Preprocessing and Training Data Generation 7
• 2.3.1 Semi-automated Labeling Strategy 8
• 2.3.2 Multi-modal Image Registration 8
• 2.3.3 Fusion Image Generation 8
• 2.3.4 Automated Vessel Map Extraction 9
• 2.3.5 Expert Refinement 9
• 2.4 Deep Learning-Based Segmentation 10
• 2.5 Post-processing Pipeline for Clinical application 10
• 2.6 Imaging Analysis 10
• 2.7 Statistical Analysis 11
• 3. Results 12
• 3.1 Patient and Lesion Characteristics 12
• 3.2 Agreement and Image Quality of PD-MRA VR 13
• 3.3 Clinical Diagnostic Performance 13
• 3.3.1 Phase 1: Automated Volume Rendering Performance 13
• 3.3.2 Phase2: Multimodal Integration with Source Images 13
• 3.3.3 Location-Specific Performance 15
• 4. Discussion 17
• 4.1 Limitations 18
• 4.2 Future Directions 19
• 5. Conclusion 19
• References 20
• 국문초록 23