رسائل ماجيستير
English
2026
Deep Clustering Approaches for Carotid Artery Calcification Detection in Panoramic Radiographs for Enhancing Cardiovascular Risk Prediction
حسين المصري, رضوان القصراوي, نادين خالد إبراهيم عريقات
كلية المهن الصحية - العلوم الطبية المساندة-جامعة القدس · فلسطين
Cardiovascular disease remains a leading cause of death worldwide, making early identification of vascular risk markers essential for prevention. Carotid artery calcifications can occasionally be visualized on panoramic dental radiographs, offering an opportunistic indicator of atherosclerotic burden during routine dental care. Yet, manual identification is challenged by inter-reader variability and the presence of anatomical mimics, and many published AI solutions rely on supervised learning that requires large, densely labeled datasets. This thesis investigates an alternative pathway by developing an unsupervised deep clustering framework for carotid region analysis on panoramic radiographs, and by integrating questionnaire-based risk factor modeling to support broader cardiovascular risk stratification. This single-center retrospective cross-sectional observational study included 1,107 panoramic radiographs acquired between February 2025 and August 2025 during routine dental examinations at Abraj Dental Clinics affiliated with Al-Quds University, in collaboration with the Faculty of Dentistry. In addition, a cross-sectional questionnaire sub-study was prospectively administered to a subset of participants (n = 438) to capture cardiovascular risk profiles and support complementary non-imaging analyses. The imaging cohort comprised 48% males and 52% females, with an age range of 18–85 years and a mean age of approximately 40 years. Preprocessing consisted of contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) followed by extraction of bilateral carotid regions of interest (ROIs), resized to 128×128 pixels. To represent each ROI, a dual feature strategy was adopted. Interpretable handcrafted radiographic features were computed to quantify intensity distributions, texture patterns, and edge- and morphology-related cues potentially associated with calcification. In parallel, deep representations were learned using a convolutional autoencoder pretrained for 300 epochs with mean squared error loss and a low-dimensional latent space, producing compact feature vectors suitable for downstream clustering. Four clustering methods were evaluated: K-Means, hierarchical clustering, Gaussian Mixture Models, and Deep Embedded Clustering (DEC). Clustering quality was assessed using internal validation metrics that do not require ground truth labels, including the Silhouette score, Davies–Bouldin index, and Calinski–Harabasz index. To connect unsupervised clustering outcomes to clinical relevance, the thesis adopted a patient-level validation strategy, where each patient contributed left and right ROI assignments. A high-risk validation subset of 21 patients was defined using confirmed cardiovascular disease history or documented carotid calcifications in clinical records. Patient-level accuracy was calculated based on the proportion of high-risk patients grouped into the dominant high-risk cluster. Model interpretability was further supported using GradCAM++ visualization to highlight salient ROI regions consistent with expected calcification-related patterns. DEC achieved the strongest clustering performance across the internal metrics, with a Silhouette score of 0.214, a Davies–Bouldin index of 1.752, and a Calinski–Harabasz index of 524, indicating improved within-cluster cohesion and between-cluster separation relative to baseline methods. Patient-level validation also favored DEC, which achieved 95.2% accuracy in aggregating high-risk patients into a dominant cluster with fewer anomalous cases compared with K-Means, hierarchical clustering, and GMM. To incorporate non-imaging determinants of cardiovascular risk, prospectively administered questionnaire data were analyzed for 438 participants (220 high risk, 218 low risk). Univariate association testing using chi-square statistics with Cramér’s V suggested notable associations with age, physical activity, sleep duration, sedentary time, dietary behaviors, selected health awareness indicators, and mental health measures such as relaxation difficulty and feelings of worthlessness. A Random Forest classifier trained on questionnaire features achieved high predictive performance (accuracy = 0.9318, ROC AUC = 0.9821, F1 = 0.9291), and feature importance analysis highlighted socioeconomic factors and psychological distress-related variables among influential predictors alongside age and lifestyle behaviors. In conclusion, this thesis demonstrates that unsupervised deep clustering of anatomically defined carotid ROIs on panoramic radiographs can yield coherent groupings aligned with clinically defined high-risk status, while offering an interpretable and label-efficient screening support pathway. Combining imaging-based signals with questionnaire-derived risk factors may further strengthen cardiovascular risk stratification and support targeted referral for confirmatory vascular assessment. Future work should incorporate gold-standard vascular confirmation, multi-site external validation, and longitudinal outcome linkage to establish clinical reliability and generalizability.
