Development of Imaging Biomarkers Utilizing Non-brain Tissue Information from Brain MRI: Deep learning-based Automated Analysis of Frank’s sign and clinical validation = 뇌 자기공명영상에서 비-뇌조직 정보를 활용한 이미징 바이오마커 개발: Deep learning 기반 Frank’s sign 자동분석 기술과 임상적 타당성 검증

다국어 초록 (Multilingual Abstract)

Frank’s sign (FS), a diagonal earlobe crease historically associated with cardiovascular disease, has long been regarded as a subjective clinical sign whose utility is confounded by heterogeneous assessment and complex risk factors. This dissertation redefines FS from a subjective clinical observation into a quantitative, automated imaging biomarker for both cerebral and systemic microangiopathy by leveraging incidental data from routine brain MRI, accomplished through three sequential studies.
First, Study 1 developed a novel 3D U-Net deep learning framework to automatically segment FS from T1 weighted images. This study addressed the foundational technical challenge by establishing a robust pipeline for detection and quantification. The model achieved high segmentation accuracy, with a mean Dice Similarity Coefficient of 0.734 in internal validation and 0.714 in external validation. It also demonstrated high classification accuracy, with an AUC of 0.942 in internal validation and 0.902 in external validation, successfully transforming FS from a subjective observation into an objective, reproducible, and quantitative biomarker.
Second, Study 2 provided the biological validation by testing this new tool in Cerebral Autosomal Dominant Arteriopathy with Subcortical Infarcts and Leukoencephalopathy (CADASIL), the archetypal pure monogenic model of small vessel disease (SVD). Results demonstrated that FS is not stochastic noise but a bona fide phenotypic marker of genetic microangiopathy. It demonstrated a 4.2-fold increased odds in CADASIL patients independent of clinical risk factors and showed a significant dose-response relationship with white matter hyperintensity (WMH) burden. Importantly, this association was found to be highly specific to diffuse WMH pathology, showing no significant correlation with focal ischemic lesions such as lacunes or microbleeds.
Third, Study 3 tested the biomarker in the complex, real-world setting of sporadic SVD and normal controls. This study uncovered that the binary marker, effective in the pure CADASIL model, failed in the sporadic cohort due to confounding by metabolic risk factors. However, quantitative deep phenotyping (radiomics) showed significantly better performance. The radiomic signature achieved perfect classification of CADASIL versus sporadic SVD (AUC 1.000). Mechanistic analysis revealed distinct data-driven signatures: a ‘Genetic Arteriopathy Signature’ (characterized by low textural complexity, termed ‘Simpler & Duller’) for CADASIL, and a ‘Metabolic Signature’ (characterized by high entropy, termed ‘Duller & Busier’) reflecting systemic metabolic burden such as hypertension, hyperlipidemia, and diabetes mellitus. This ‘Duller & Busier’ signature was shared between sporadic SVD patients and normal controls with metabolic risk factors, providing the morphological basis for the poor specificity of the conventional binary FS for SVD etiology. Furthermore, the radiomic signature provided statistically significant incremental value for screening systemic diseases, such as diabetes mellitus and hyperlipidemia, in scenarios where the binary sign was uninformative.
Collectively, this dissertation transforms FS from a confounded sign into a precise, automated biomarker. It demonstrates that the true diagnostic value lies not in its presence but in its quantitative, disease-specific morphological signature, realizing the potential of brain MRI as a ‘unified window’ for assessing both cerebral and systemic vascular health.

국문 초록 (Abstract)

Frank 징후(Frank’s sign, FS)는 귓불의 대각선 주름으로 역사적으로 심혈관 질환과의 연관성이 제시되어 왔으나 주관적 평가 방식과 복잡한 교란 변수로 인해 그 임상적 효용성이 불명확했습니다. 본 학위논문은 뇌 MRI 영상에서 부수적으로 얻어지는 정보를 활용하여 FS를 뇌혈관 및 전신 미세혈관병증의 정량적이고 자동화된 바이오마커로 재정의하고자 하였으며 이는 세 단계의 순차적 연구를 통해 수행되었습니다.
첫째, 연구 1에서는 T1 강조 영상에서 FS를 자동으로 분할하는 3차원 U Net 기반의 딥러닝 프레임워크를 개발하였습니다. 이 연구는 FS 평가의 기술적 난제를 해결하여 탐지 및 정량화를 위한 견고한 파이프라인을 구축하였습니다. 개발된 모델은 내부 검증에서 0.734, 외부 검증에서 0.714의 높은 평균 Dice Similarity Coefficient (DSC)를 달성하며 우수한 분할 정확도를 보였습니다. 또한 내부 검증 AUC 0.942 및 외부 검증 AUC 0.902의 높은 분류 정확도를 기록하여 FS를 주관적 관찰에서 객관적이고 재현 가능한 정량적 바이오마커로 전환시켰습니다.
둘째, 연구 2에서는 순수 유전성 소혈관질환 (SVD) 모델인 CADASIL 환자군에 이 도구를 적용하여 생물학적 타당성을 검증하였습니다. 연구 결과 FS의 존재는 단순한 잡음이 아니라 유전성 미세혈관병증의 진정한 표지자임이 밝혀졌습니다. FS는 임상적 위험 인자와 무관하게 CADASIL 발병 위험과 4.2배 높은 연관성을 보였으며 백질고신호 (WMH) 부하와 뚜렷한 용량 반응 관계를 나타냈습니다. 특히 이러한 연관성은 미만성 백질 병리에 매우 특이적이었으며 열공성 경색이나 미세출혈과 같은 국소적 허혈 병변과는 유의한 상관관계를 보이지 않았습니다.
셋째, 연구 3에서는 산발성 SVD 및 건강한 대조군이 포함된 실제 임상 환경에서 이 바이오마커를 검증하였습니다. 그 결과 순수 CADASIL 모델에서 유효했던 이진 분류 (binary) 표지자가 산발성 환자군에서는 대사성 위험 인자의 교란으로 인해 효용성이 떨어짐을 확인하였습니다. 그러나 정량적 심층 표현형 분석 (Radiomics)은 탁월한 우수성을 입증하였습니다. 영상체학적 시그니처는 CADASIL과 산발성 SVD를 구분하는 데 있어 AUC 1.000 의 뛰어난 분류 성능을 달성하였습니다. 기전 분석 결과 이는 CADASIL의 더 단순하고 어두운 (Simpler and Duller) 특성과 산발성 SVD의 더 어둡고 복잡한 (Duller and Busier) 특성이라는 데이터 기반의 고유한 시그니처 차이에 기인한 것으로 밝혀졌습니다. 또한 이 시그니처는 이진 징후가 정보를 제공하지 못하는 고혈압, 당뇨병이나 고지혈증과 같은 전신 질환 스크리닝에서도 통계적으로 유의한 증분 가치를 제공하였습니다.
결론적으로 본 학위논문은 FS를 교란 변수가 많은 단순 징후에서 정밀하고 자동화된 바이오마커로 변화시켰습니다. FS의 진정한 진단적 가치는 단순한 존재 여부가 아니라 정량적이고 질병 특이적인 형태학적 시그니처에 있음을 증명하였으며 뇌 MRI가 뇌혈관과 전신 혈관 건강을 동시에 평가하는 통합된 창(Unified Window)으로 활용될 수 있는 가능성을 실현하였습니다.

목차 (Table of Contents)

• Abstract i
• Table of Contents iv
• List of Tables vii
• List of Figures viii
• CHAPTER 1. INTRODUCTION １
• 1.1 DEEP LEARNING REVOLUTION IN MEDICAL IMAGE SEGMENTATION AND MORPHOMETRIC ANALYSIS １
• 1.2 FRANK’S SIGN: AN UNEXPECTED TERRITORY FOR COMPUTATIONAL ANALYSIS ４
• 1.3 FROM CARDIOVASCULAR TO CEREBROVASCULAR: EXPANDING CLINICAL RELEVANCE ９
• 1.4 TECHNICAL FRAMEWORK: BRIDGING SEGMENTATION AND MORPHOMETRY １１
• 1.5 VALIDATION STRATEGY: FROM CADASIL AS A PURE MODEL TO THE SPORADIC SVD SPECTRUM １３
• 1.6 RESEARCH OBJECTIVES AND INNOVATION １７
• CHAPTER 2. METHODS ２１
• 2.1 STUDY 1: DEVELOPMENT AND VALIDATION OF DEEP LEARNING FRAMEWORK FOR AUTOMATED FRANK’S SIGN DETECTION FROM BRAIN MRI ２１
• 2.1.1 Study Population and Data Collection ２３
• 2.1.2 Image Acquisition Parameters ２４
• 2.1.3 Image Preprocessing ２４
• 2.1.4 Manual Annotation Protocol ２６
• 2.1.5 Deep Learning Architecture ２７
• 2.1.6 Loss Function and Optimization ２９
• 2.1.7 Model Evaluation and Validation ２９
• 2.2 STUDY 2: FRANK’S SIGN AS A DOSE-DEPENDENT MARKER IN CADASIL ３０
• 2.2.1 Study Population ３０
• 2.2.2 Clinical and Neuropsychological Assessment ３１
• 2.2.3 Neuroimaging Acquisition and Analysis ３２
• 2.2.4 Automated Frank’s sign Detection ３３
• 2.2.5 Statistical Analysis ３３
• 2.3 STUDY 3: MORPHOMETRIC SIGNATURES OF FRANK’S SIGN: LINKING PERIPHERAL MORPHOLOGY TO CEREBRAL AND SYSTEMIC VASCULAR PATHOLOGY ３４
• 2.3.1 Study Population ３６
• 2.3.2 Clinical and Demographic Data ３７
• 2.3.3 FS Radiomic Feature Extraction ３７
• 2.3.4 Statistical Analysis ４０
• CHAPTER 3. RESULTS ４５
• 3.1 STUDY 1: DEEP LEARNING FRAMEWORK RESULTS ４５
• 3.1.1 Dataset Characteristics ４５
• 3.1.2 Architecture Comparison ４９
• 3.1.3 Segmentation Performance ５３
• 3.1.4 Binary Classification Performance ５３
• 3.1.5 Correlation with Manual Measurements ５７
• 3.2 STUDY 2: CADASIL VALIDATION RESULTS ６０
• 3.2.1 Case-Control Comparison ６０
• 3.2.2 Frank’s sign and Disease Characteristics in CADASIL ６３
• 3.2.3 Dose-Response Relationship ６６
• 3.3 STUDY 3: MORPHOMETRIC ANALYSIS RESULTS ６９
• 3.3.1 Population characteristics and Frank’s sign prevalence ６９
• 3.3.2 Limitations of Conventional Binary Assessment ７１
• 3.3.3 Diagnostic Performance of Morphological Signatures ７５
• 3.3.4 Deep Phenotyping: Morphological Signatures in the FS-positive Cohort ７９
• 3.3.5 Linking Morphological Signature for Screening Vascular Risk Factors ８６
• CHAPTER 4. DISCUSSION ９３
• 4.1 STUDY 1: TECHNICAL INNOVATION AND VALIDATION ９３
• 4.1.1 Technical Innovations and Contributions ９３
• 4.1.2 Clinical Applicability and Limitations ９４
• 4.1.3 Implications for Multi-Modal Biomarker Development ９６
• 4.2 STUDY 2: BIOLOGICAL VALIDATION IN CADASIL ９７
• 4.2.1 Frank’s sign in the Context of Genetic Arteriopathy ９７
• 4.2.2 Dose-Response Characteristics and Clinical Implications ９９
• 4.2.3 Specificity to White Matter Pathology ９９
• 4.2.4 Limitations and Future Directions １００
• 4.3 STUDY 3: MORPHOLOGICAL SIGNATURES AND CLINICAL TRANSLATION １００
• 4.3.1 Limitations of Binary Assessment: A Confounded and Non-Specific Marker １００
• 4.3.2 Disease-Specific Morphological Signature: The Core Evidence of Difference １０１
• 4.3.3 Clinical Implications: Classification Performance １０６
• 4.3.4 Limitations and Future Directions １０８
• 4.4 GENERAL DISCUSSION AND SYNTHESIS １１０
• 4.4.1 Synthesis of Findings １１０
• 4.4.2 Theoretical Implications １１２
• 4.4.2.1. Quantitative Validation of Peripheral-Central Vascular Coupling １１２
• 4.4.2.2. Morphological Reflections of Distinct Pathophysiological Mechanisms １１３
• 4.4.3 Clinical Translation １１６
• 4.4.4 Methodological Contributions １２１
• 4.4.4.1. Technical Innovation １２１
• 4.4.4.2. Validation Approach １２１
• 4.4.5 Broader Implications １２２
• CHAPTER 5. CONCLUSION １２５
• 5.1 SUMMARY OF KEY FINDINGS １２５
• 5.2 LIMITATIONS １２６
• 5.3 FUTURE RESEARCH DIRECTIONS １２８
• 5.4 FINAL CONCLUSIONS １２９
• Bibliography １３１
• 국문 초록 １４５