Magnetic Resonance Diffusion Imaging (MRI-DWI) is an advanced noninvasive imaging technique that allows the exploration of the microstructural properties of biological tissues by assessing the Brownian motion of water molecules. Initially introduced in the field of neuroradiology, DWI has gradually established itself as an essential diagnostic tool in other clinical settings as well, such as oncology, nephrology, and hepatology. The underlying physical principle is based on sensitizing the MR signal to changes in molecular diffusion using magnetic field gradients, resulting in quantitative maps capable of highlighting tissue changes that are often undetectable with conventional sequences. In the context of DWI, the Intra Voxel Incoherent Motion (IVIM) model represents an advanced extension of simple Gaussian diffusion, highlighting in the MR signal the contribution of pseudo-diffusion due to microvascular capillary perfusion. Introduced by Le Bihan et al. in the 1980s, the IVIM model in its most general and complete form is based on a bi-exponential equation that allows separate estimation of physiologically relevant parameters: “pure” diffusion coefficient (D), pseudo-diffusion coefficient (D*) and perfusion fraction (f). This technique allows more detailed characterization of tissue microstructure and microcirculation than conventional MR imaging techniques, which are unable to separate the contributions of perfusion and diffusion in the same acquisition. However, the high sensitivity to noise disturbances, dependence on technical parameters, and complexity of parameter estimation still pose significant challenges to robust and reproducible application of the IVIM model in clinical settings. The purpose of the following research is to identify an effective and robust automated pipeline that applies the IVIM model to a cohort of patients from different Italian hospital centers. The prediction of IVIM parameters is done by using a Physics-Informed neural network, through which optimal parametric maps with strong diagnostic power can be derived. In the paper, different network training techniques (single-patient, single-center and multicenter) and two IVIM models (bi-exponential and tri-exponential) will be evaluated in order to identify the best combination for predictive power, performance and computational cost. The study will focus on a specific organ: the liver; through appropriate segmentations, regions of interest (ROIs) will be obtained from which the different analysis techniques can be performed.Finally, we will proceed with a federated approach (Federated Learning) that allows different hospitals to train a model with their own patient data, never directly exchanging images or clinical information, but still contributing to the creation of a shared and more generalizable model. The consistent integration of IVIM imaging into healthcare procedures offers numerous clinical, diagnostic, and research advantages. It allows for a more detailed and non-invasive assessment of tissue microstructure, which can translate into improved diagnosis and characterization of lesions and pathological tissues. This thesis begins a journey to provide methodologies for estimating parametric maps of IVIM models with physiological constraints in hospital settings that also leverage federated learning techniques to overcome the limitations imposed by privacy laws and protect sensitive patient data.
La Risonanza Magnetica per Diffusione (MRI-DWI) è una tecnica avanzata di imaging non invasivo che consente di esplorare le proprietà microstrutturali dei tessuti biologici valutando il movimento browniano delle molecole d’acqua. Introdotta inizialmente in ambito neuroradiologico, la DWI si è progressivamente affermata come strumento diagnostico essenziale anche in altri contesti clinici, quali oncologia, nefrologia ed epatologia. Il principio fisico alla base si fonda sulla sensibilizzazione del segnale MR alle variazioni della diffusione molecolare tramite l’impiego di gradienti di campo magnetico, ottenendo mappe quantitative in grado di evidenziare alterazioni tissutali spesso non rilevabili con sequenze convenzionali. Nel contesto della DWI, il modello dell’Intra Voxel Incoherent Motion (IVIM) rappresenta un’estensione avanzata della semplice diffusione gaussiana, evidenziando e separando nel segnale MR il contributo della pseudo-diffusione dovuta alla perfusione capillare microvascolare. Introdotto da Le Bihan et al. negli anni ’80, il modello IVIM nella sua forma più generale e completa si basa su un'equazione bi-esponenziale che consente la stima separata di parametri fisiologicamente rilevanti: coefficiente di diffusione “puro” (D), coefficiente di pseudo-diffusione (D*) e frazione di perfusione (f). Tuttavia, l’elevata sensibilità ai disturbi di rumore, la dipendenza da parametri tecnici e la complessità della stima dei parametri rappresentano ancora oggi sfide significative per un’applicazione robusta e riproducibile del modello IVIM in ambito clinico. Lo scopo della seguente ricerca consiste nell’individuare una pipeline automatica efficace e robusta che applichi il modello IVIM ad una coorte di pazienti proveniente da differenti centri ospedalieri italiani. La predizione dei parametri IVIM avviene mediante l’impiego di una rete neurale Physics-Informed, attraverso la quale si possono ricavare mappe parametriche ottimali con forte potere diagnostico. Nell’elaborato saranno valutate differenti tecniche di training della rete (singolo paziente, monocentrico e multicentrico) e due modelli IVIM (bi-esponenziale e tri-esponenziale) al fine di individuare la migliore combinazione per potere predittivo, prestazioni e costo computazionale. Lo studio verterà in un organo specifico: il fegato; tramite opportune segmentazioni si ottengono delle regioni di interesse (ROI) dalle quali poter eseguire le diverse tecniche di analisi. Infine, si procederà approcci federati (Federated Learning) che permette ad ospedali diversi di allenare un modello con i propri dati di pazienti, senza mai scambiarsi direttamente le immagini o le informazioni cliniche, ma contribuendo comunque alla creazione di un modello condiviso e più generalizzabile. Una integrazione costante dell'imaging IVIM nelle procedure sanitarie offre numerosi vantaggi clinici, diagnostici e di ricerca. Consentendo una valutazione più dettagliata e non invasiva della microstruttura tissutale, che può tradursi in una migliore diagnosi e caratterizzazione di lesioni e tessuti patologici. Questa tesi inizia un percorso per fornire metodologie di stima delle mappe parametriche dei modelli IVIM con vincoli fisiologici in ambienti ospedalieri che sfruttano anche le tecniche di learning federato, per superare i limiti imposti dalle leggi sulla privacy e tutelare i dati sensibili dei pazienti.
STIMA ROBUSTA DEI PARAMETRI DEL MODELLO INTRA VOXEL INCOHERENT MOTIONS DEL FEGATO UTILIZZANDO RETI NEURALI PHYSICS-INFORMED E APPPRENDIMENTO FEDERATO TRA CENTRI DI IMAGING
PRINCIPI, FEDERICO
2024/2025
Abstract
Magnetic Resonance Diffusion Imaging (MRI-DWI) is an advanced noninvasive imaging technique that allows the exploration of the microstructural properties of biological tissues by assessing the Brownian motion of water molecules. Initially introduced in the field of neuroradiology, DWI has gradually established itself as an essential diagnostic tool in other clinical settings as well, such as oncology, nephrology, and hepatology. The underlying physical principle is based on sensitizing the MR signal to changes in molecular diffusion using magnetic field gradients, resulting in quantitative maps capable of highlighting tissue changes that are often undetectable with conventional sequences. In the context of DWI, the Intra Voxel Incoherent Motion (IVIM) model represents an advanced extension of simple Gaussian diffusion, highlighting in the MR signal the contribution of pseudo-diffusion due to microvascular capillary perfusion. Introduced by Le Bihan et al. in the 1980s, the IVIM model in its most general and complete form is based on a bi-exponential equation that allows separate estimation of physiologically relevant parameters: “pure” diffusion coefficient (D), pseudo-diffusion coefficient (D*) and perfusion fraction (f). This technique allows more detailed characterization of tissue microstructure and microcirculation than conventional MR imaging techniques, which are unable to separate the contributions of perfusion and diffusion in the same acquisition. However, the high sensitivity to noise disturbances, dependence on technical parameters, and complexity of parameter estimation still pose significant challenges to robust and reproducible application of the IVIM model in clinical settings. The purpose of the following research is to identify an effective and robust automated pipeline that applies the IVIM model to a cohort of patients from different Italian hospital centers. The prediction of IVIM parameters is done by using a Physics-Informed neural network, through which optimal parametric maps with strong diagnostic power can be derived. In the paper, different network training techniques (single-patient, single-center and multicenter) and two IVIM models (bi-exponential and tri-exponential) will be evaluated in order to identify the best combination for predictive power, performance and computational cost. The study will focus on a specific organ: the liver; through appropriate segmentations, regions of interest (ROIs) will be obtained from which the different analysis techniques can be performed.Finally, we will proceed with a federated approach (Federated Learning) that allows different hospitals to train a model with their own patient data, never directly exchanging images or clinical information, but still contributing to the creation of a shared and more generalizable model. The consistent integration of IVIM imaging into healthcare procedures offers numerous clinical, diagnostic, and research advantages. It allows for a more detailed and non-invasive assessment of tissue microstructure, which can translate into improved diagnosis and characterization of lesions and pathological tissues. This thesis begins a journey to provide methodologies for estimating parametric maps of IVIM models with physiological constraints in hospital settings that also leverage federated learning techniques to overcome the limitations imposed by privacy laws and protect sensitive patient data.| File | Dimensione | Formato | |
|---|---|---|---|
|
Principi_Federico.pdf
embargo fino al 24/10/2026
Dimensione
5.3 MB
Formato
Adobe PDF
|
5.3 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/95806