Magnetic resonance imaging (MRI) has recently emerged as a useful diagnostic tool for pulmonary investigations thanks to the advantages provided over the current gold standard, Computed Tomography (CT). However, MRI is still unable to reach CT performance due to unfavourable lung tissue properties and the lack of standard acquisition protocols. To increase MRI image quality, respecting the constraint on the maximum duration of a diagnostic exam, it is necessary to design new protocols and standardize them; this process would benefit from the development of a dedicated MRI lung phantom, to be used as a reference. This thesis aims to conduct a preliminary study for the development of a dedicated MRI lung phantom investigating the relaxometric and radiomic properties of the lung tissue, thanks to a collaboration with the Policlinic San Matteo of Pavia. The lung tissue relaxometric properties were investigated through NMR measurements of ex-vivo samples from patients with lung cancer. The obtained values were compared with those reported in the literature, extrapolated from in-vivo MRI maps; the comparison showed that there is no agreement between these, coming to the conclusion that ex-vivo pulmonary samples are not adequate for the characterization of NMR relaxation times of the lung. The following radiomic analysis was carried out from CT and MRI images of 19 patients with NSCLC (non-small-cell lung cancer). Firstly, the agreement between two software dedicated to radiomic features extraction, LIFEx and Pyradiomics, was verified by checking the values match of selected features extracted from the same image, and the correlation between these. From the analysis emerged that there is no correspondence between features values and that features extracted from CT images are more correlated. In a second phase, the correlation between features extracted through the same software was investigated for both CT and MRI acquisitions, graphically and by the Spearman test. From the analysis, it was seen that the correlation of features extracted through Pyradiomics turns out to be greater. Finally, the presence of features related to the PD-L1 biomarker, whose value is relevant to select the most suitable therapy, was investigated through the Kruskal-Wallis test and through four Machine Learning models: Multinomial logistic regression, Random Forest, Decision Tree, and Support Vector Machine. Among the features extracted with LIFEx, the Kruskal-Wallis test indicates the presence of two weakly correlated features for CT images, and one for MRI images. Considering all the remaining models applied, the accuracy is less than 0.5, so the importance obtained is not reliable. From the work carried out it was concluded: (i) the discrepancy between the ex-vivo NMR values and the in-vivo MRI values is mainly due to the blood circulation diffusion, present in the in-vivo measurements and absent in the ex-vivo ones; (ii) there are some image characteristics on which both software are in agreement, but that in general the comparison between different software applied on the same images has to be done cautiously, even considering features with the same definition and IBSI compliant (approved by a collaboration aimed at harmonizing the radiomics process); (iii) if there is an intention to conduct a multi-messenger clinical investigation based on information extracted both from CT and MRI images, the software that appears to be more stable is Pyradiomics; (iv) it was not possible to identify one or more features significantly correlated to the PD-L1 value due to the narrowness of the starting data set. The first chapter briefly describes the physical principles of the techniques involved, NMR, MRI, and CT, the second is dedicated to Radiomics, the third is a review of the current lung MRI state of the art, and the fourth and fifth include respectively the NMR investigations and the radiomic analysis conducted.
La MRI è recentemente emersa come utile strumento diagnostico per le indagini polmonari grazie ai vantaggi forniti rispetto all’attuale standard di riferimento, la CT. Tuttavia, attualmente la MRI non è in grado di eguagliare le prestazioni della CT a causa di caratteristiche sfavorevoli del tessuto polmonare e della mancanza di protocolli standard d’acquisizione. Per aumentare la qualità delle immagini MRI, rispettando il vincolo sulla durata massima di un esame diagnostico, è necessario progettare nuovi protocolli e standardizzarli; tale processo beneficerebbe dello sviluppo di un fantoccio MRI dedicato da usare come riferimento. Questo lavoro di tesi è volto a condurre uno studio preliminare per lo sviluppo di un fantoccio MRI dedicato indagando le caratteristiche rilassometriche e radiomiche del tessuto polmonare, grazie a una collaborazione con il Policlinico San Matteo di Pavia. L’indagine preliminare delle proprietà rilassometriche del tessuto polmonare è stata condotta tramite misure NMR di campioni ex-vivo provenienti da pazienti affetti da carcinoma polmonare. I valori ottenuti dalle curve di rilassamento sono stati confrontati con quelli riportati in letteratura, estrapolati da mappe MRI in-vivo, e dal confronto è emerso che non vi è accordo fra questi. La successiva analisi radiomica è stata svolta a partire da immagini CT ed MRI di 19 pazienti affetti da NSCLC, non-small-cell lung cancer. In una prima fase si è verificato l’accordo fra due software dedicati all’estrazione di features radiomiche, LIFEx e Pyradiomics, controllando la coincidenza dei valori associati a features selezionate estratte a parità d’immagine, e la correlazione fra queste. È emerso che i valori delle features non coincidono e che le features estratte a partire da immagini CT risultano essere maggiormente correlate. In una seconda fase si è indagata la correlazione fra features estratte a parità di software a partire da acquisizione CT ed MRI, graficamente e tramite il test di Spearman. Dall’analisi si è visto che la correlazione per features estratte tramite Pyradiomics risulta essere maggiore. In ultimo, si è cercata la presenza di features correlate al biomarcatore PD-L1, il cui valore è rilevante per stabilire la terapia più adatta, tramite il test di Kruskal-Wallis e 4 modelli di Machine Learning: Multinomial logistic regression, Random Forest, Decision Tree, e Support Vector Machine. Fra le features estratte attraverso LIFEx, tramite il test di Kruskal-Wallis sono state individuate due features debolmente correlate a partire da immagini CT, e una a partire da immagini MRI. Per tutti i restanti modelli applicati l’accuracy risulta essere inferiore a 0,5, dunque le "importance" ottenute non sono affidabili. Dal lavoro svolto si è concluso: (i) la discrepanza fra i valori NMR ex-vivo ed MRI in-vivo è dovuta principalmente alla diffusione del circolo sanguigno, presente nelle misure in-vivo ed assente in quelle ex-vivo; (ii) vi sono alcune features su cui entrambi i software sono in accordo ma in generale bisogna prestare attenzione a confrontare software diversi a parità d’immagine, anche se per il confronto si utilizzano features con stessa definizione e IBSI compliant; (iii) se si intende condurre un’indagine clinica multi-messaggera basata su informazioni estratte a partire da immagini CT ed MRI il software che pare essere più stabile è Pyradiomics; (iv) non è stato possibile individuare una o più feature significativamente correlate al valore del biomarcatore PD-L1 a causa della ristrettezza del data set di partenza. Nel primo capitolo vengono descritti brevemente i principi fisici delle tecniche coinvolte, NMR, MRI, e CT; il secondo capitolo è dedicato alla Radiomica, il terzo è una rassegna sull’attuale stato dell’arte della MRI su polmone, il quarto e il quinto racchiudono rispettivamente le indagini NMR e l’analisi radiomica svolte.
Studio delle proprietà rilassometriche e radiomiche in pazienti affetti da carcinoma polmonare. Indagine preliminare per lo sviluppo di un fantoccio MRI dedicato.
ROBUSTELLI TEST, AGNESE
2021/2022
Abstract
Magnetic resonance imaging (MRI) has recently emerged as a useful diagnostic tool for pulmonary investigations thanks to the advantages provided over the current gold standard, Computed Tomography (CT). However, MRI is still unable to reach CT performance due to unfavourable lung tissue properties and the lack of standard acquisition protocols. To increase MRI image quality, respecting the constraint on the maximum duration of a diagnostic exam, it is necessary to design new protocols and standardize them; this process would benefit from the development of a dedicated MRI lung phantom, to be used as a reference. This thesis aims to conduct a preliminary study for the development of a dedicated MRI lung phantom investigating the relaxometric and radiomic properties of the lung tissue, thanks to a collaboration with the Policlinic San Matteo of Pavia. The lung tissue relaxometric properties were investigated through NMR measurements of ex-vivo samples from patients with lung cancer. The obtained values were compared with those reported in the literature, extrapolated from in-vivo MRI maps; the comparison showed that there is no agreement between these, coming to the conclusion that ex-vivo pulmonary samples are not adequate for the characterization of NMR relaxation times of the lung. The following radiomic analysis was carried out from CT and MRI images of 19 patients with NSCLC (non-small-cell lung cancer). Firstly, the agreement between two software dedicated to radiomic features extraction, LIFEx and Pyradiomics, was verified by checking the values match of selected features extracted from the same image, and the correlation between these. From the analysis emerged that there is no correspondence between features values and that features extracted from CT images are more correlated. In a second phase, the correlation between features extracted through the same software was investigated for both CT and MRI acquisitions, graphically and by the Spearman test. From the analysis, it was seen that the correlation of features extracted through Pyradiomics turns out to be greater. Finally, the presence of features related to the PD-L1 biomarker, whose value is relevant to select the most suitable therapy, was investigated through the Kruskal-Wallis test and through four Machine Learning models: Multinomial logistic regression, Random Forest, Decision Tree, and Support Vector Machine. Among the features extracted with LIFEx, the Kruskal-Wallis test indicates the presence of two weakly correlated features for CT images, and one for MRI images. Considering all the remaining models applied, the accuracy is less than 0.5, so the importance obtained is not reliable. From the work carried out it was concluded: (i) the discrepancy between the ex-vivo NMR values and the in-vivo MRI values is mainly due to the blood circulation diffusion, present in the in-vivo measurements and absent in the ex-vivo ones; (ii) there are some image characteristics on which both software are in agreement, but that in general the comparison between different software applied on the same images has to be done cautiously, even considering features with the same definition and IBSI compliant (approved by a collaboration aimed at harmonizing the radiomics process); (iii) if there is an intention to conduct a multi-messenger clinical investigation based on information extracted both from CT and MRI images, the software that appears to be more stable is Pyradiomics; (iv) it was not possible to identify one or more features significantly correlated to the PD-L1 value due to the narrowness of the starting data set. The first chapter briefly describes the physical principles of the techniques involved, NMR, MRI, and CT, the second is dedicated to Radiomics, the third is a review of the current lung MRI state of the art, and the fourth and fifth include respectively the NMR investigations and the radiomic analysis conducted.È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/14935