Atherosclerosis is a chronic and progressive condition and a major contributor to cerebrovascular disease, with an estimated 10-20% of ischemic strokes attributable to carotid artery disease. Early detection of rupture-prone, vulnerable plaques is therefore crucial to prevent major cerebrovascular events. This thesis proposes a novel sector-based radiomic approach to investigate the local relationship between baseline radiomic features and two-year carotid plaque remodelling in symptomatic patients. A multicentric cohort of 75 patients (from 4 hospitals) underwent multi-detector computational tomography angiography (MDCTA). Each carotid artery was rigorously sectorized along the reconstructed centerline into 1mm x 45° units, and each sector was classified as plaque-containing when wall thickness was ≥1.5 mm, yielding 7315 sectors for the subsequent analysis. Radiomic features were extracted sector by sector and, for each sector, three longitudinal morphometric outcomes were computed: calcification thickness change, lumen-to-calcification distance change, and plaque wall thickness change. Associations between baseline radiomic features and morphometric changes were first explored through univariate analysis and then extended to multivariate modelling, developing machine-learning classifiers to predict progression versus regression patterns. All combinations between 4 selectors and 5 classifiers were trained with 5-fold cross-validation. The best-performing combination was then refit on the full training set and evaluated on the internal test set and on external test centers. The multicenter design ensured the availability of a true external test and enabled training on different center combinations. Single-center training achieved high internal accuracy but exhibited performance drops on external centers, reflecting demographic and acquisition-related domain shifts. Training on heterogeneous, multicenter data systematically improved external generalizability, indicating that combining datasets effectively mitigates these shifts. In particular, texture-based radiomic features, related to structural heterogeneity, emerged as the most strongly associated with the morphometric outcomes. On the external cohort, the best models achieved for calcium thickness evolution a balanced accuracy of 0.768 ±0.014 and a ROC-AUC of 0.827 ±0.013; and for lumen-to-calcification distance change a balanced accuracy of 0.793 ±0.012 and a ROC-AUC of 0.852 ±0.013. In contrast, plaque wall thickness patterns were less predictable from texture alone (best F1-score= 0.632 ±0.028). In addition an integrated model combining radiomic features with baseline morphometric markers of vulnerability was able to distinguish culprit from non-culprit carotid arteries, achieving on the external center a balanced accuracy of 0.69 ±0.08 and a ROC-AUC of 0.79 ±0.07. These results highlight the potential of jointly modelling for artery-level risk assessment. Overall, the results showed that sector-level radiomic analysis can capture trajectory of calcium remodelling, whereas wall remodelling may require richer, multi-scale features and multimodal integration with clinical and biomechanical factors. These findings support sector-based radiomics as a non-invasive imaging biomarker for carotid plaque evolution and contribute to moving beyond stenosis toward generalizable, patient-specific risk stratification.
L’aterosclerosi è una condizione cronica e progressiva che costituisce una delle principali cause di patologia cerebrovascolare; si stima che il 10–20% degli ictus ischemici sia attribuibile alla stenosi carotidea. Pertanto, l’identificazione precoce delle placche vulnerabili, a rischio di rottura, è cruciale per prevenire eventi cerebrovascolari e ictus. Questa tesi propone un approccio radiomico basato sulla settorizzazione della placca carotidea per esplorare la relazione locale tra le caratteristiche radiomiche di base e il rimodellamento della placca a due anni, in pazienti sintomatici. È stata analizzata una coorte multicentrica di 75 pazienti (da quattro ospedali) sottoposti ad angio-TC a multidetettore (MDCTA). Ciascuna carotide è stata suddivisa lungo la centerline in spicchi (1 mm di spessore × 45°); ognisettore è statoclassificato come contenente placca se lospessore della parete era superiore a 1.5 mm, ottenendo 7315 settori per le analisi successive. Le features radiomiche sono state estratte settore per settore e, per ciascun settore, sono state calcolate tre variabili morfometriche longitudinali: variazione dello spessore della calcificazione, variazione della distanza lume-calcificazione e variazione dello spessore di parete. Le associazioni tra features radiomiche di base e modifiche morfometriche sono state inizialmente indagate con un’analisi univariata e successivamente estese a modelli multivariati, sviluppando classificatori di machine learning per predire pattern di progressione/regressione. Tutte le combinazioni tra 4 selettori e 5 classificatori sono state addestrate con 5-fold cross-validation. La combinazione migliore è stata quindi riaddestrata sull’intero training set e valutata sul test interno e sui centri esterni. Il design multicentrico ha garantito la disponibilità di un vero test esterno e ha consentito l’addestramento su diverse combinazioni di centri. L’addestramento su singolo centro ha mostrato alte prestazioni interne, ma un calo sui centri esterni, riflettendo le differenze demografiche e di acquisizione (domain shift). L’addestramento su dati eterogenei multicentrici ha migliorato in modo sistematico la generalizzabilità esterna, indicando che la combinazione dei dataset mitiga efficacemente tali effetti. In particolare, le features testurali, descrittive dell’eterogeneità strutturale, sono risultate le più associate alle variabili morfometriche. Sulla coorte esterna, i modelli migliori hanno ottenuto, per l’evoluzione dello spessore della calcificazione, una balanced accuracy di 0.768 ±0.014 e una ROC-AUC di 0.827 ±0.013; e, per la distanza lume-calcificazione, una balanced accuracy di 0.793 ±0.012 e una ROC-AUC di 0.852 ±0.013. Al contrario, i pattern di spessore di parete sono risultati meno predicibili dalle sole feature testurali (miglior F1-score= 0.632 ±0.028). È stato inoltre sviluppato un modello integrato che combina feature radiomiche e caratteristiche di vulnerabilità, in grado di distinguere tra carotidi culprit e non culprit, con una BA = 0.69 ±0.08 e una ROC-AUC = 0.79 ±0.07 sul centro esterno. Questi risultati evidenziano il potenziale della modellizzazione congiunta per la valutazione del rischio a livello di singola arteria. Nel complesso, i risultati mostrano che un’analisi radiomica a livello settoriale è in grado di cogliere la traiettoria del rimodellamento calcifico, mentre il rimodellamento di parete potrebbe richiedere descrittori multi-scala e un’integrazione multimodale con fattori clinici e biomeccanici. Queste evidenze supportano la radiomica settoriale come biomarcatore di imaging non invasivo per l’evoluzione della placca carotidea e contribuiscono ad andare oltre la stenosi verso una stratificazione del rischio generalizzabile e personalizzata.
A novel sector-based radiomic approach to assess atherosclerotic carotid plaque dynamics
DE SANTIS, MARINA
2025/2026
Abstract
Atherosclerosis is a chronic and progressive condition and a major contributor to cerebrovascular disease, with an estimated 10-20% of ischemic strokes attributable to carotid artery disease. Early detection of rupture-prone, vulnerable plaques is therefore crucial to prevent major cerebrovascular events. This thesis proposes a novel sector-based radiomic approach to investigate the local relationship between baseline radiomic features and two-year carotid plaque remodelling in symptomatic patients. A multicentric cohort of 75 patients (from 4 hospitals) underwent multi-detector computational tomography angiography (MDCTA). Each carotid artery was rigorously sectorized along the reconstructed centerline into 1mm x 45° units, and each sector was classified as plaque-containing when wall thickness was ≥1.5 mm, yielding 7315 sectors for the subsequent analysis. Radiomic features were extracted sector by sector and, for each sector, three longitudinal morphometric outcomes were computed: calcification thickness change, lumen-to-calcification distance change, and plaque wall thickness change. Associations between baseline radiomic features and morphometric changes were first explored through univariate analysis and then extended to multivariate modelling, developing machine-learning classifiers to predict progression versus regression patterns. All combinations between 4 selectors and 5 classifiers were trained with 5-fold cross-validation. The best-performing combination was then refit on the full training set and evaluated on the internal test set and on external test centers. The multicenter design ensured the availability of a true external test and enabled training on different center combinations. Single-center training achieved high internal accuracy but exhibited performance drops on external centers, reflecting demographic and acquisition-related domain shifts. Training on heterogeneous, multicenter data systematically improved external generalizability, indicating that combining datasets effectively mitigates these shifts. In particular, texture-based radiomic features, related to structural heterogeneity, emerged as the most strongly associated with the morphometric outcomes. On the external cohort, the best models achieved for calcium thickness evolution a balanced accuracy of 0.768 ±0.014 and a ROC-AUC of 0.827 ±0.013; and for lumen-to-calcification distance change a balanced accuracy of 0.793 ±0.012 and a ROC-AUC of 0.852 ±0.013. In contrast, plaque wall thickness patterns were less predictable from texture alone (best F1-score= 0.632 ±0.028). In addition an integrated model combining radiomic features with baseline morphometric markers of vulnerability was able to distinguish culprit from non-culprit carotid arteries, achieving on the external center a balanced accuracy of 0.69 ±0.08 and a ROC-AUC of 0.79 ±0.07. These results highlight the potential of jointly modelling for artery-level risk assessment. Overall, the results showed that sector-level radiomic analysis can capture trajectory of calcium remodelling, whereas wall remodelling may require richer, multi-scale features and multimodal integration with clinical and biomechanical factors. These findings support sector-based radiomics as a non-invasive imaging biomarker for carotid plaque evolution and contribute to moving beyond stenosis toward generalizable, patient-specific risk stratification.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_DeSantis_Tesi.pdf
accessibile in internet per tutti
Descrizione: testo della tesi
Dimensione
29.43 MB
Formato
Adobe PDF
|
29.43 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_DeSantis_ExecutiveSummary.pdf
accessibile in internet per tutti
Descrizione: testo executive summary
Dimensione
805.69 kB
Formato
Adobe PDF
|
805.69 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/247104