Chordoma is a rare type of sarcoma that typically grows in the sacrum or in the skull basis, at the ends of the spine. resection is often unfeasible for skull-base chordomas, which also happen to be highly radioresistant. For these reasons, hadrontherapy using carbon ions (Carbon Ion Radiation Therapy - CIRT), turns out to be one the most efficient treatments for controlling tumour progression and reducing the damage to the nearby healthy tissues. In the last few years, various methods for medical imaging analysis, serving various clinical purposes, have been developed and one such method is named Radiomics. This term indicates the quantitative description of specific regions identified on medical images that is achieved through an automatic or semi-automatic process of features extraction and modelling of, for example, patient outcome. The aim of this thesis was two-fold. The main aim was to develop and evaluate radiomic approaches, going from feature selection to machine learning model validation, able to predict local control (LC) in patients affected by skull-base chordoma and treated with carbon-ion hadrontherapy (CIRT). In particular, using CT and MR images and dose maps, different features were selected and extracted and different models were trained following various pipelines. The second aim consisted in investigating if texture features were able to describe tumour response differently according to the directions along which features were computed, i.e. considering the directionality of the therapeutic beams. By exploiting the functionalities of the Computational Environment for Radiological Research (CERR) and Matlab (v. R2019b), radiomic features were extracted from the GTV for each patient obtaining, for each modality, a matrix n x p, where n is the number of patients (57) and p is the number of the extracted features. Different algorithms and their combinations were tested for feature selection, namely Lasso, Lasso - PCA and ICC - Lasso - PCA. For the classification task two machine learning models were implemented, the first one being Logistic Regression (LR) and the second one Support Vector Machines (SVM). Two approaches were implemented to optimize the combination of feature selection and classification methods: - Approach 1 (A1), in which the N most important features were chosen and then the model has been trained and evaluated. This approach needed two independent cross-validation loops to be applied. - Approach 2 (A2), in which, for each cross-validation loop, N features have been chosen and for each subset the model has been trained and evaluated. This approach requires feature selection and classification algorithms to be included in the same cross-validation loop. Models' performance was evaluated in both approaches through a Receiver Operating Curve (ROC) analysis for which Area Under the Curve (AUC), Sensitivity (SE) and Specificity (SP) were computed. In both cases a 5-fold cross-validation (CV) has been used, in order to split the data in 80% for training and validation and the other 20% for test, at each loop. In order to improve results generalizability, ten random seeds have been used to split the data before CV was applied, following the paradigm of a repeated CV pipeline. The data was standardised before being sent to the models and balanced following three different ways: (i) only using the ADASYN algorithm to balance the dataset, (ii) using ADASYN and implementing Bootstrap only on training data and (iii) applying Bootstrap on the entire dataset, before ADASYN. In this thesis a preliminary evaluation of the dependence of dose texture on directions has been conducted. Other than using 13 standard directions to extract texture features, selected directions allowed evaluating texture in the direction parallel to the beam axis and in the plane perpendicular to it. These directions have been exploited to evaluate their impact on the description of dose texture and, then, to verify if those changes could lead to improvements in terms of LC prediction. Comparing Lasso with Lasso - PCA methods, it appears that the former reaches better results in predicting LC since it reaches AUCs equal to 0.88 and 0.85 in the best cases (dose maps) for balancing and balancing & Bootstrap, respectively. The latter method (Lasso - PCA), instead, achieves maximum AUCs (for dose maps) of 0.74 and 0.67. This could be explained by the fact that Lasso is a very strict algorithm, as it completely discards non informative features and may exclude useful information for PCA. When ICC was used, there was a strong initial reduction of features that were to be presented to Lasso and PCA, which translated to worse models performances. This may be related to the fact that ICC does not imply that the selected features are the more informative ones, but only the most robust ones. In the exploratory study on directional textures in dose maps, features are not fed to a feature selection algorithm due to their reduced number. Features extracted along directions parallel to the major beam axis yielded AUC values (0.72 for balancing and 0.71 for balancing & Bootstrap) 5 % higher than those extracted within the perpendicular plane (0.67 for balancing and 0.66 for balancing & Bootstrap). Parallel textural features were also superior to those extracted from the standard set of directions (0.69 both for balancing and balancing & Bootstrap). These preliminary results suggest that relative angles used to irradiate the tumour may have an impact on treatment efficacy but a thorough dosimetric evaluation is needed. Future developments of this thesis include: to increase the sample size by enrolling more patients, and to explore different classifiers and to improve those already implemented . Also, different features subsets sizes could be tested to determine the optimal number of features to be considered alltogether. Finally, for what concerns the exploratory study, different directions relative to the major beam axis and different beam configurations could be considered to optimize models' performance.
I cordomi sono particolari forme di sarcoma che tipicamente insorgono nel sacro o nella base cranica, ai due estremi della colonna vertebrale. La resezione totale è spesso infattibile per cordomi della basi-cranio, il quale spesso è anche altamente radioresistente. Per queste ragioni, l'adroterapia tramite ioni carbonio (Carbon Ion Radiation Therapy - CIRT), risulta essere uno dei più efficaci trattamenti per controllare la progressione del tumore e ridurre i danni ai tessuti sani adiacenti ad esso. Negli ultimi anni diversi metodi di analisi delle immagini biomediche sono state sviluppate ed uno di questi è nominato Radiomica. Con questo termine si indica la descrizione quantitativa di specifiche regioni identificate nelle immagine mediche attraverso un processo automatico o semi-automatico di estrazione di features e modeling di outcome del paziente. Per questa tesi si possono individuare due principali obbiettivi. L'obiettivo primario è quello di sviluppare e valutare processi radiomici, partendo dagli step di features selection fino alla validazione del modello costruito tramite algoritmi di machine learning, capaci di predire il controllo locale (Local Control, LC) in pazienti affetti da cordoma e trattati con adroterapia a ioni carbonio (CIRT). L'obiettivo secondario consiste invece nell'investigare se le features di texture fossero in grado di descrivere la risposta del tumore in funzione delle direzioni con cui tali features vengono estratte, i.e. considerando la direzionalità del fascio di radiazione. Tramite l'utilizzo combinato di Computational Environment for Radiological Research (CERR) e Matlab (v.R2019b) sono state estratte le features per ogni paziente, ottenendo per ogni modalità una matrice di dimensione n x p, dove n è il numero dei pazienti (57) e p il numero delle features estratte. Diversi algoritmi e delle loro combinazioni sono stati investigati per la selezione delle features, quali Lasso, Lasso - PCA ed Interclass Correlation Coefficient (ICC) - Lasso - PCA. Per la classificazione sono stati invece implementati due modelli di machine learning, il primo basato su Logistic Regression (LR) ed il secondo sfruttando Support Vector Machine (SVM). Due approcci sono stati utilizzati per ottimizzare la combinazione di features selection e metodi di classificazione: - L'approccio 1 (A1), in cui prima sono state scelte le N features più significative e con queste si è addestrato il modello. Questo approccio necessita di due cicli indipendenti di cross-validation. - Nell'approccio 2 (A2), in cui, ad ogni loop di cross-validation, sono state scelte N features e per ognuno di questi subset selezionati si è addestrato e valutato un modello. Questo approccio contiene in un solo ciclo di cross-validation sia la features selection che l'algoritmo di classificazione. Le performance dei modelli sono state valutate in entrambi gli approcci attraverso l'analisi della Curva di funzionamento del ricevitore (Reciver Operating Group - ROC) e di conseguenza l'Area sotto la curva (Area Under the Curve - AUC), Sensitivity (SE) e Specificity (SP). In entrambi i casi è stata utilizzata la cross validazione con un Kfold di 5, in modo da partizionare i dati in 80% per il training e per la validazione mentre il rimanente 20 % per il test, ad ogni loop. Per migliorare la generalizzabilità dei risultati, sono stati usati 10 random seed per partizionare i dati prima di applicare la CV. Il dataset è stato standardizzato e bilanciato, prima di essere dato in ingresso al modello, in tre modi differenti: (i) utilizzando la funzione ADASYN per bilanciare il dataset, (ii) Usando ADASYN e implementando Boostrap ai soli dati di training e (iii) Applicando Boostrap su tutto il dataset, prima del bilanciamento. In entrambi i casi è stata utilizzata la cross validazione con un Kfold di 5, in modo da partizionare i dati in 80% per il training e per la validazione mentre il rimanente 20% per il test, ad ogni loop. Per migliorare la generalizzabilità dei risultati, sono stati usati 10 random seed per partizionare i dati prima di applicare la CV. In questa tesi è stata inoltre condotta una valutazione preliminare della dipendenza della texture della dose rispetto alle direzioni. Oltre alle 13 direzioni standard utilizzate per estrarre le features di texture, le direzioni selezionate hanno permesso di valutare la texture nella direzione parallela all'asse del fascio e nel piano perpendicolare ad esso. Queste direzioni sono state sfruttate per valutare il loro impatto sulla descrizione della texture della dose, e poi, per verificare se questi cambiamenti potessero portare a miglioramenti in termini di previsione LC. Confrontando il metodo di Lasso con Lasso - PCA si evince che il primo ottiene risultati migliori in termini di predittività del Local Control, in quanto raggiunge nel caso migliore (per mappe di dose) livelli di AUC pari a 0.88 e 0.85, per il caso bilanciato e bilanciato & Boostrap, rispettivamente. Per il secondo step invece (Lasso - PCA) i valori di AUC maggiori (per le mappe di dose) sono 0.74 e 0.67. Quando l'ICC è usato invece si è notata una riduzione di features a monte degli step di Lasso e PCA, il che si traduce in perfomance dei modelli peggiori. Questo è legato al fatto che l'ICC non dà certezza che le features selezionate siano quelle più informative ma solo quelle più robuste. Nello studio esplorativo relativo alle mappe di dose non sono stati utilizzati metodi di feature selection a causa del basso numero di features utilizzate. Le features estratte utilizzando le direzioni parallele all'asse del fascio principale porta a valori di AUC (0.72 per il caso bilanciato e 0.71 per il caso bilanciato & Bootstrap) migliori del 5% rispetto al caso perpendicolare (0.67 per il bilanciato mentre è 0.66 per bilanciato & Bootstrap). Le features di texture parallele sono anche superiori di quelle estratte utilizzando le direzioni standard (0.69 sia per il caso bilanciato sia per il bilanciato & Bootstrap). Questi risultati preliminari mostrano che gli angoli relativi utilizzati per irradiare il tumore potrebbero avere un impatto sull'efficacia del trattamento, ma è necessaria una valutazione dosimetrica approfondita. Da questi risultati preliminari, la texture lungo il fascio sembra essere in qualche modo legata all'outcome del trattamento, ma ulteriori studi sono necessari per capire queste differenze. Dei possibili sviluppi futuri per questa tesi potrebbero essere: aumentare il numero di dati, quindi pazienti, ed esplorare diversi classificatori o migliorare quelli già utilizzati. Anche subset di features di differenti dimensioni potrebbero essere testati per determinare il numero ottimo di features da considerare. Infine, per quanto riguarda lo studio esplorativo, diverse direzioni relative all'asse del fascio principale e configurazioni differenti di fasci potrebbero essere considerate per ottimizzare le performance del modello.
Radiomica multi-parametrica per la predizione del controllo locale in pazienti affetti da cordoma della basi-cranio
MARZORATI, LORENZO
2019/2020
Abstract
Chordoma is a rare type of sarcoma that typically grows in the sacrum or in the skull basis, at the ends of the spine. resection is often unfeasible for skull-base chordomas, which also happen to be highly radioresistant. For these reasons, hadrontherapy using carbon ions (Carbon Ion Radiation Therapy - CIRT), turns out to be one the most efficient treatments for controlling tumour progression and reducing the damage to the nearby healthy tissues. In the last few years, various methods for medical imaging analysis, serving various clinical purposes, have been developed and one such method is named Radiomics. This term indicates the quantitative description of specific regions identified on medical images that is achieved through an automatic or semi-automatic process of features extraction and modelling of, for example, patient outcome. The aim of this thesis was two-fold. The main aim was to develop and evaluate radiomic approaches, going from feature selection to machine learning model validation, able to predict local control (LC) in patients affected by skull-base chordoma and treated with carbon-ion hadrontherapy (CIRT). In particular, using CT and MR images and dose maps, different features were selected and extracted and different models were trained following various pipelines. The second aim consisted in investigating if texture features were able to describe tumour response differently according to the directions along which features were computed, i.e. considering the directionality of the therapeutic beams. By exploiting the functionalities of the Computational Environment for Radiological Research (CERR) and Matlab (v. R2019b), radiomic features were extracted from the GTV for each patient obtaining, for each modality, a matrix n x p, where n is the number of patients (57) and p is the number of the extracted features. Different algorithms and their combinations were tested for feature selection, namely Lasso, Lasso - PCA and ICC - Lasso - PCA. For the classification task two machine learning models were implemented, the first one being Logistic Regression (LR) and the second one Support Vector Machines (SVM). Two approaches were implemented to optimize the combination of feature selection and classification methods: - Approach 1 (A1), in which the N most important features were chosen and then the model has been trained and evaluated. This approach needed two independent cross-validation loops to be applied. - Approach 2 (A2), in which, for each cross-validation loop, N features have been chosen and for each subset the model has been trained and evaluated. This approach requires feature selection and classification algorithms to be included in the same cross-validation loop. Models' performance was evaluated in both approaches through a Receiver Operating Curve (ROC) analysis for which Area Under the Curve (AUC), Sensitivity (SE) and Specificity (SP) were computed. In both cases a 5-fold cross-validation (CV) has been used, in order to split the data in 80% for training and validation and the other 20% for test, at each loop. In order to improve results generalizability, ten random seeds have been used to split the data before CV was applied, following the paradigm of a repeated CV pipeline. The data was standardised before being sent to the models and balanced following three different ways: (i) only using the ADASYN algorithm to balance the dataset, (ii) using ADASYN and implementing Bootstrap only on training data and (iii) applying Bootstrap on the entire dataset, before ADASYN. In this thesis a preliminary evaluation of the dependence of dose texture on directions has been conducted. Other than using 13 standard directions to extract texture features, selected directions allowed evaluating texture in the direction parallel to the beam axis and in the plane perpendicular to it. These directions have been exploited to evaluate their impact on the description of dose texture and, then, to verify if those changes could lead to improvements in terms of LC prediction. Comparing Lasso with Lasso - PCA methods, it appears that the former reaches better results in predicting LC since it reaches AUCs equal to 0.88 and 0.85 in the best cases (dose maps) for balancing and balancing & Bootstrap, respectively. The latter method (Lasso - PCA), instead, achieves maximum AUCs (for dose maps) of 0.74 and 0.67. This could be explained by the fact that Lasso is a very strict algorithm, as it completely discards non informative features and may exclude useful information for PCA. When ICC was used, there was a strong initial reduction of features that were to be presented to Lasso and PCA, which translated to worse models performances. This may be related to the fact that ICC does not imply that the selected features are the more informative ones, but only the most robust ones. In the exploratory study on directional textures in dose maps, features are not fed to a feature selection algorithm due to their reduced number. Features extracted along directions parallel to the major beam axis yielded AUC values (0.72 for balancing and 0.71 for balancing & Bootstrap) 5 % higher than those extracted within the perpendicular plane (0.67 for balancing and 0.66 for balancing & Bootstrap). Parallel textural features were also superior to those extracted from the standard set of directions (0.69 both for balancing and balancing & Bootstrap). These preliminary results suggest that relative angles used to irradiate the tumour may have an impact on treatment efficacy but a thorough dosimetric evaluation is needed. Future developments of this thesis include: to increase the sample size by enrolling more patients, and to explore different classifiers and to improve those already implemented . Also, different features subsets sizes could be tested to determine the optimal number of features to be considered alltogether. Finally, for what concerns the exploratory study, different directions relative to the major beam axis and different beam configurations could be considered to optimize models' performance.| File | Dimensione | Formato | |
|---|---|---|---|
|
MT_Marzorati.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi di laurea magistrale
Dimensione
3.31 MB
Formato
Adobe PDF
|
3.31 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/10589/167056