Coronary artery disease is the primary cause of mortality worldwide. The etiology of the pathology is atherosclerosis, characterized by the buildup of fatty deposits on the inner walls of coronary arteries. Atherosclerotic plaques may grow and become “vulnerable”, namely prone to instability and rupture, leading to major adverse cardiac events. Radiomics, which involves extracting quantitative metrics from clinical images, has emerged as a powerful tool for identifying markers of plaque vulnerability. When combined with Machine Learning (ML) tools, it enables the development of predictive models. The goal of the present thesis was to define a radiomics-based ML pipeline for the prediction of vulnerable arteries and patients. The study involved 40 patients who underwent Coronary computed tomography angiography (CCTA) and Optical coherence tomography (OCT), with vulnerable plaques labelled from OCT analysis. After image segmentation and volume of interest extraction, the analysis encompassed pre-processing of CCTA images, feature extraction and selection, and ML model training and testing. Four classification tasks were pursued: the assessment of t1) vulnerable left anterior descending artery (LAD); t2) vulnerable arteries, t3) vulnerable patients and t4) treated patients. Due to the small and imbalanced dataset, leave-one-out cross-validation was used. Three features selection methods and 9 ML algorithms were considered to develop binary classifiers. Besides developing radiomic classifiers, biomechanical and combined radiomics-biomechanical classifiers were built to assess whether their integration would enhance predictive precision compared to using each set alone. Radiomics alone provided better performances than biomechanics alone and their combination supported overall performance improvement. The best averaged performances for radiomics alone and the combined model gave F1-score equal to 0.74 vs 0.77 for t1, 0.66 vs 0.70 for t2, 0.72 vs 0.70 for t3, 0.67 vs 0.66 for t4. The results obtained demonstrated the potentialities of the proposed approach for vulnerable plaque prediction.
La malattia coronarica è la principale causa di mortalità nel mondo. La causa della patologia è l’aterosclerosi, caratterizzata da accumulo di depositi di grasso nelle pareti interne delle arterie coronarie. Le placche aterosclerotiche possono diventare ’vulnerabili’, ossia inclini ad instabilità e rottura, portando a eventi cardiaci avversi gravi. La radiomica è emersa come potente strumento per identificare marcatori di vulnerabilità delle placche. Combinata con strumenti di Machine Learning (ML), consente lo sviluppo di modelli predittivi. L’obiettivo della tesi è definire una pipeline di ML basata sulla radiomica per predizione di arterie e pazienti vulnerabili. Lo studio ha coinvolto 40 pazienti sottoposti ad angiografia coronarica con tomografia computerizzata (CCTA) e tomografia ottica computerizzata (OCT), con placche vulnerabili identificate da analisi OCT. L’analisi ha incluso segmentazione dell’immagine, estrazione del volume di interesse, pre-processing delle immagini CCTA, estrazione e selezione delle features, addestramento e test del modello ML. Sono stati definiti quattro task di classificazione: identificazione di t1) arteria discendente sinistra (LAD) , t2) arterie e t3) pazienti vulnerabili e t4) valutazione di pazienti trattati. A causa del dataset ridotto e sbilanciato, è stata utilizzata la leave-one-out cross-validation. Sono stati considerati tre metodi di selezione di features e 9 algoritmi di ML per sviluppare classificatori binari. Oltre allo sviluppo di classificatori radiomici, sono stati costruiti classificatori biomeccanici e radiomici-biomeccanici combinati per valutare se l’integrazione avrebbe migliorato la precisione predittiva rispetto all’uso di set separati. La sola radiomica ha fornito prestazioni migliori rispetto alla sola biomeccanica e la loro combinazione ha portato ad un miglioramento complessivo delle prestazioni. Le migliori performances mediate per la sola radiomica e il modello combinato hanno dato un F1-score di 0.74 vs a 0.77 per t1, 0.66 vs a 0.70 per t2, 0.72 vs a 0.70 per t3, 0.67 vs a 0.66 per t4. I risultati hanno dimostrato le potenzialità dell’approccio proposto per la predizione di placche vulnerabili.
A radiomics-based machine learning approach for the prediction of coronary plaques' vulnerability
DI LUZIO, GIULIA;Disanti, Sara
2022/2023
Abstract
Coronary artery disease is the primary cause of mortality worldwide. The etiology of the pathology is atherosclerosis, characterized by the buildup of fatty deposits on the inner walls of coronary arteries. Atherosclerotic plaques may grow and become “vulnerable”, namely prone to instability and rupture, leading to major adverse cardiac events. Radiomics, which involves extracting quantitative metrics from clinical images, has emerged as a powerful tool for identifying markers of plaque vulnerability. When combined with Machine Learning (ML) tools, it enables the development of predictive models. The goal of the present thesis was to define a radiomics-based ML pipeline for the prediction of vulnerable arteries and patients. The study involved 40 patients who underwent Coronary computed tomography angiography (CCTA) and Optical coherence tomography (OCT), with vulnerable plaques labelled from OCT analysis. After image segmentation and volume of interest extraction, the analysis encompassed pre-processing of CCTA images, feature extraction and selection, and ML model training and testing. Four classification tasks were pursued: the assessment of t1) vulnerable left anterior descending artery (LAD); t2) vulnerable arteries, t3) vulnerable patients and t4) treated patients. Due to the small and imbalanced dataset, leave-one-out cross-validation was used. Three features selection methods and 9 ML algorithms were considered to develop binary classifiers. Besides developing radiomic classifiers, biomechanical and combined radiomics-biomechanical classifiers were built to assess whether their integration would enhance predictive precision compared to using each set alone. Radiomics alone provided better performances than biomechanics alone and their combination supported overall performance improvement. The best averaged performances for radiomics alone and the combined model gave F1-score equal to 0.74 vs 0.77 for t1, 0.66 vs 0.70 for t2, 0.72 vs 0.70 for t3, 0.67 vs 0.66 for t4. The results obtained demonstrated the potentialities of the proposed approach for vulnerable plaque prediction.File | Dimensione | Formato | |
---|---|---|---|
2023_12_DiLuzio_Disanti_ExecutiveSummary_02.pdf
solo utenti autorizzati dal 30/11/2024
Descrizione: Executive Summary
Dimensione
630.46 kB
Formato
Adobe PDF
|
630.46 kB | Adobe PDF | Visualizza/Apri |
2023_12_DiLuzio_Disanti_Tesi_01.pdf
solo utenti autorizzati dal 30/11/2024
Descrizione: Tesi
Dimensione
10.44 MB
Formato
Adobe PDF
|
10.44 MB | 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/10589/215462