The scope of this thesis is to develop advanced computational models of the vascular microenvironment, to provide quantitative insights into the pathophysiology of diseases such as cancer. This work employs a mixed-dimensional framework consisting of a 3D-1D computational model based on parametrized partial differential equations as a full-order model (FOM) that describes blood flow and oxygen transport from the microvasculature to the tissue. In the first part of the work, the geometrical modeling of a vascular network is addressed. We propose two methods for the synthesis of vascular networks: the first one relies on Voronoi-based diagrams to control morphometrical and topological properties, while the second approach uses deep learning techniques to create complex and anatomically consistent vascular geometries efficiently. These models are then applied to radiotherapy (RT) to explore how hypoxia affects tumor resistance to radiation. Computational studies show that regions with irregular vascularization exhibit hypoxia and greater radioresistance. Consequently, a sensitivity analysis identifies key factors that influence oxygenation, with a focus on vascular topology. In the second part of the work, we introduce advanced nonlinear reduced-order modeling (ROM) techniques and deep learning approaches to achieve a relevant speed-up and a negligible loss in accuracy with respect to the FOM. Based on the standard proper orthogonal decomposition (POD) approach and leveraging Mesh-Informed Neural Networks (MINNs), the ROM combines together a neural network approximating the POD coefficients and a closure model acting as fine-scale corrector for the local structures not captured by the former. This approach is validated by means of an error analysis and an uncertainty quantification (UQ) analysis, performed in a multifidelity framework. For the latter, we introduce a deep learning-enhanced multifidelity Monte Carlo (DL-MFMC) method, proposing an optimal budget management policy to accelerate the statistics of quantities of interest, particularly in the context of oxygen transport and RT outcome. Preliminary applications of these computa- tional models in lab-on-chip systems are explored, analyzing vascular permeability before and after RT, showing the potential for future precision medicine in cancer treatment.
L’obiettivo della tesi è sviluppare modelli computazionali avanzati del microambiente vascolare, per fornire approfondimenti quantitativi sulla fisiopatologia di malattie come il cancro. Nel lavoro si adotta un framework misto-dimensionale costituito da un modello computazionale 3D-1D basato su equazioni alle derivate parziali parametriche, come modello ad alta fedeltà (FOM) che descrive il flusso sanguigno e il trasporto di ossigeno dal microcircolo al tessuto. Nella prima parte del lavoro si affronta la modellizzazione geometrica di una rete vascolare, proponendo due metodi per la sintesi delle reti: il primo basato su diagrammi di Voronoi per controllare le proprietà morfometriche e topologiche, mentre il secondo che utilizza tecniche di deep learning per generare in modo efficiente geometrie vascolari complesse e anatomicamente coerenti. Tali modelli sono poi applicati alla radioterapia (RT) per valutare come l’ipossia influenzi la resistenza del tumore alle radiazioni. Gli studi computazionali dimostrano che le regioni con vascolarizzazione irregolare presentano ipossia e maggiore radioresistenza. In seguito, un’analisi di sensitività identifica i fattori chiave che influenzano l’ossigenazione, con particolare attenzione alla topologia vascolare. Nella seconda parte del lavoro, sono introdotte tecniche avanzate di riduzione di modello (ROM) non lineare e approcci di deep learning per ridurre i costi computazionali, con perdite di accuratezza trascurabili rispetto al FOM. Sfruttando la proper orthogonal decomposition (POD) e utilizzando reti neurali con architetture mesh-informed (MINNs), il ROM combina una rete neurale che approssima i coefficienti POD e un modello di chiusura come correttore di scala fine per le strutture locali. Il metodo è validato attraverso un’analisi degli errori e una quantificazione delle incertezze (UQ) in un framework multifedeltà. Per quest’ultima, viene introdotto un metodo Monte Carlo multifedeltà potenziato dal deep learning (DL-MFMC), con un algoritmo di gestione ottimale del budget computazionale per accelerare il calcolo di statistiche di quantità di interesse relative a trasporto di ossigeno e RT. Si esplorano infine applicazioni preliminari di tali modelli in sistemi lab-on-chip, analizzando la permeabilità vascolare prima e dopo la RT, dimostrando il loro potenziale per la medicina di precisione nel trattamento del cancro.
Advanced multiphysics computational models for the vascular microenvironment
Vitullo, Piermario
2024/2025
Abstract
The scope of this thesis is to develop advanced computational models of the vascular microenvironment, to provide quantitative insights into the pathophysiology of diseases such as cancer. This work employs a mixed-dimensional framework consisting of a 3D-1D computational model based on parametrized partial differential equations as a full-order model (FOM) that describes blood flow and oxygen transport from the microvasculature to the tissue. In the first part of the work, the geometrical modeling of a vascular network is addressed. We propose two methods for the synthesis of vascular networks: the first one relies on Voronoi-based diagrams to control morphometrical and topological properties, while the second approach uses deep learning techniques to create complex and anatomically consistent vascular geometries efficiently. These models are then applied to radiotherapy (RT) to explore how hypoxia affects tumor resistance to radiation. Computational studies show that regions with irregular vascularization exhibit hypoxia and greater radioresistance. Consequently, a sensitivity analysis identifies key factors that influence oxygenation, with a focus on vascular topology. In the second part of the work, we introduce advanced nonlinear reduced-order modeling (ROM) techniques and deep learning approaches to achieve a relevant speed-up and a negligible loss in accuracy with respect to the FOM. Based on the standard proper orthogonal decomposition (POD) approach and leveraging Mesh-Informed Neural Networks (MINNs), the ROM combines together a neural network approximating the POD coefficients and a closure model acting as fine-scale corrector for the local structures not captured by the former. This approach is validated by means of an error analysis and an uncertainty quantification (UQ) analysis, performed in a multifidelity framework. For the latter, we introduce a deep learning-enhanced multifidelity Monte Carlo (DL-MFMC) method, proposing an optimal budget management policy to accelerate the statistics of quantities of interest, particularly in the context of oxygen transport and RT outcome. Preliminary applications of these computa- tional models in lab-on-chip systems are explored, analyzing vascular permeability before and after RT, showing the potential for future precision medicine in cancer treatment.File | Dimensione | Formato | |
---|---|---|---|
PhD_Thesis_Piermario_Vitullo.pdf
accessibile in internet per tutti
Dimensione
33.37 MB
Formato
Adobe PDF
|
33.37 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/10589/232633