Multiple sclerosis (MS) is one of the most common disabling diseases that affect young adults, and its incidence is increasing worldwide. Magnetic Resonance Imaging (MRI) is a fundamental technique for the diagnosis of this disease. Traditional MRI is primarily qualitative, so different quantitative MRI (qMRI) approaches have been developed to obtain further information about the examined biological tissues.An example is diffusion MRI (dMRI): the NMR signal from which the images are reconstructed is sensitive to water diffusion inside the tissues, which in turn is influenced by their composition and microstructure. A dMRI study generally comprises four steps: 1) acquisition, 2) image preprocessing (denoising, unring and correction for distortions), 3) image processing and 4) results analysis. However, there is no universally standardized approach for conducting dMRI analyses, and different research groups may employ different algorithms. The objective of this thesis is dual: first, the effect of different algorithms and diffusion models was investigated from a technical point of view, then the dMRI models were applied to a clinical context, specifically targeting subjects affected by multiple sclerosis. A total of 45 subjects were considered: 14 healthy controls and 31 patients affected by MS. Initially, different implementations of the denoising and unring algorithms were compared. In particular, the denoising algorithms considered in this work are based on the Marchenko-Pastur Principal Component Analysis (MPPCA) while the unring ones are based on the local sub-voxel shift method. In both cases the comparison between the different algorithms was conducted using heatscatter plots and difference maps, and computing Pearson correlation coefficient of the resulting images. After this, the images were fitted with different diffusion models: two Neurite Orientation Dispersion and Density Imaging models (Watson-NODDI and Bingham-NODDI) and the spherical mean technique (SMT), which are multicompartment models, and the diffusion kurtosis imaging (DKI), which is a tensor model. For Watson-NODDI and DKI different implementations were compared. Regarding the Watson-NODDI model, the aim was to compare for the first time the original Matlab implementation proposed by Zhang and colleagues (2012) with its counterpart in DMIPY. It was discovered that DMIPY's model is based on different equations with respect to the ones present in Matlab, thus DMIPY was modified so that the signal was fitted with the original equations developed by Zhang. In this way it was possible to conduct a residual analysis and determine the effect that changing the equations had on the resulting estimated parametric maps. Furthermore, a comparison was carried out between Watson-NODDI and Bingham-NODDI model. Regarding DKI, the aim was to compare two software, DIPY and DESIGNER, a toolbox developed in 2014 by Ades-Aron et al. The analysis revealed that the quality of the resulting maps depended on the preprocessing pipeline. The analysis of the different algorithms and model implementations showed that the choice of the software used for the quantitative analysis of dMRI images can influence the results. Therefore, when making comparisons in dMRI it is necessary to pay attention to the software used in order to disentangle the causes of the differences in the microstructural metrics, that can be due both to physiological changes and to algorithms implementation. In the second part of the thesis, the above-mentioned models were employed in a clinical context and applied to subjects affected by MS. The analyses were performed with a Voxel-Based Analysis (VBA), and the results were consistent with both the prior observations between Watson and Bingham NODDI and the existing literature.
La sclerosi multipla (SM) è una delle più comuni malattie invalidanti che colpiscono i giovani adulti, e la sua incidenza è in aumento in tutto il mondo. La risonanza magnetica (MRI) è una tecnica fondamentale per la diagnosi di questa malattia. L'MRI tradizionale è principalmente qualitativa, quindi sono stati sviluppati diversi approcci quantitativi (qMRI) che permettono di ottenere ulteriori informazioni sui tessuti in esame. Un esempio è la risonanza magnetica di diffusione (dMRI): il segnale NMR da cui vengono ricostruite le immagini è sensibile alla diffusione dell'acqua all'interno dei tessuti, che a sua volta è influenzato dalla loro composizione e microstruttura. Uno studio dMRI comprende generalmente quattro fasi: 1) acquisizione 2) pre-elaborazione dell'immagine (denoising, unring e correzione delle distorsioni) 3) elaborazione dell'immagine e 4) analisi dei risultati. Tuttavia, non esiste un approccio universalmente standardizzato per svolgere le analisi di dMRI dunque diversi gruppi di ricerca possono implementare algoritmi diversi. L'obiettivo di questa tesi è duplice: inizialmente, l'effetto di diversi algoritmi e modelli di diffusione è stato studiato da un punto di vista tecnico, quindi i modelli di dMRI sono stati applicati a un contesto clinico, specificamente a soggetti affetti da sclerosi multipla. Un totale di 45 soggetti sono stati considerati: 14 controlli sani e 31 pazienti affetti da SM. Inizialmente, sono state confrontate diverse implementazioni degli algoritmi di denoising e unring. In particolare, gli algoritmi di denoising considerati sono basati sulla Marchenko-Pastur Principal Component Analysis mentre quelli di unring sono basati sul metodo di local sub-voxel shift. In entrambi i casi il confronto tra i diversi algoritmi è stato condotto utilizzando heatscatter e mappe differenza e calcolando il coefficiente di correlazione Pearson tra le immagini risultanti. Successivamente, le immagini sono state fittate con i diversi modelli di diffusione: due modelli di Neurite Orientation Dispersion e Density Imaging (Watson-NODDI e Bingham-NODDI) e la spherical mean technique (SMT), che sono modelli multicompartimento, e la diffusione kurtosis imaging (DKI) che è un modello tensoriale. Per Watson-NODDI e DKI sono state confrontate diverse implementazioni. Per quanto riguarda il modello Watson-NODDI, l'obiettivo era quello di confrontare per la prima volta l'implementazione originale di Matlab proposta da Zhang (2012) con la sua controparte in DMIPY. È stato scoperto che il modello di DMIPY si basa su equazioni diverse da quelle di Matlab, quindi DMIPY è stato modificato in modo che il segnale fosse fittato con equazioni originali sviluppate da Zhang. In questo modo è stato possibile condurre un'analisi dei residui e determinare l'effetto che modificare le equazioni ha sulle mappe parametriche stimate. Inoltre, il modello Watson-NODDI è stato confrontato con il modello Bingham-NODDI. Per quanto riguarda DKI, l'obiettivo era quello di confrontare DIPY e DESIGNER, un software sviluppato nel 2014 da Ades-Aron et al. L'analisi ha rivelato che la qualità delle mappe risultanti dipende dalla pipeline di preprocessing. L'analisi dei diversi algoritmi ha dimostrato che la scelta del software può influenzare la stima dei parametri. Pertanto, quando si effettuano confronti in dMRI è necessario prestare attenzione al software utilizzato e comprendere se eventuali differenze nelle metriche stimate sono dovute a cambiamenti fisiologici o agli algoritmi. Nella seconda parte della tesi, i modelli sono stati applicati a soggetti affetti da SM e analizzati tramite voxel based analysis. I risultati trovati sono coerenti sia con le osservazioni precedenti tra Watson e Bingham NODDI che con la letteratura esistente.
Studio dell'impatto di diverse implementazioni di modelli di diffusione MRI sulle mappe microstrutturali: un'applicazione alla sclerosi multipla
TOGNOLINI, GIADA
2022/2023
Abstract
Multiple sclerosis (MS) is one of the most common disabling diseases that affect young adults, and its incidence is increasing worldwide. Magnetic Resonance Imaging (MRI) is a fundamental technique for the diagnosis of this disease. Traditional MRI is primarily qualitative, so different quantitative MRI (qMRI) approaches have been developed to obtain further information about the examined biological tissues.An example is diffusion MRI (dMRI): the NMR signal from which the images are reconstructed is sensitive to water diffusion inside the tissues, which in turn is influenced by their composition and microstructure. A dMRI study generally comprises four steps: 1) acquisition, 2) image preprocessing (denoising, unring and correction for distortions), 3) image processing and 4) results analysis. However, there is no universally standardized approach for conducting dMRI analyses, and different research groups may employ different algorithms. The objective of this thesis is dual: first, the effect of different algorithms and diffusion models was investigated from a technical point of view, then the dMRI models were applied to a clinical context, specifically targeting subjects affected by multiple sclerosis. A total of 45 subjects were considered: 14 healthy controls and 31 patients affected by MS. Initially, different implementations of the denoising and unring algorithms were compared. In particular, the denoising algorithms considered in this work are based on the Marchenko-Pastur Principal Component Analysis (MPPCA) while the unring ones are based on the local sub-voxel shift method. In both cases the comparison between the different algorithms was conducted using heatscatter plots and difference maps, and computing Pearson correlation coefficient of the resulting images. After this, the images were fitted with different diffusion models: two Neurite Orientation Dispersion and Density Imaging models (Watson-NODDI and Bingham-NODDI) and the spherical mean technique (SMT), which are multicompartment models, and the diffusion kurtosis imaging (DKI), which is a tensor model. For Watson-NODDI and DKI different implementations were compared. Regarding the Watson-NODDI model, the aim was to compare for the first time the original Matlab implementation proposed by Zhang and colleagues (2012) with its counterpart in DMIPY. It was discovered that DMIPY's model is based on different equations with respect to the ones present in Matlab, thus DMIPY was modified so that the signal was fitted with the original equations developed by Zhang. In this way it was possible to conduct a residual analysis and determine the effect that changing the equations had on the resulting estimated parametric maps. Furthermore, a comparison was carried out between Watson-NODDI and Bingham-NODDI model. Regarding DKI, the aim was to compare two software, DIPY and DESIGNER, a toolbox developed in 2014 by Ades-Aron et al. The analysis revealed that the quality of the resulting maps depended on the preprocessing pipeline. The analysis of the different algorithms and model implementations showed that the choice of the software used for the quantitative analysis of dMRI images can influence the results. Therefore, when making comparisons in dMRI it is necessary to pay attention to the software used in order to disentangle the causes of the differences in the microstructural metrics, that can be due both to physiological changes and to algorithms implementation. In the second part of the thesis, the above-mentioned models were employed in a clinical context and applied to subjects affected by MS. The analyses were performed with a Voxel-Based Analysis (VBA), and the results were consistent with both the prior observations between Watson and Bingham NODDI and the existing literature.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/20.500.14239/16544