This thesis investigates the observability of temperature distribution within the Gross Tumor Volume during superficial hyperthermia treatment, an adjuvant therapy employed alongside radiotherapy or chemotherapy for relapses and treatmentresistant tumors. A critical clinical challenge is addressed: the real-time assessment of temperature at the target site—essential for treatment feedback. Current protocols rely on invasive thermocouples, causing significant patient discomfort while providing only discrete, imprecisely localized measurements. This challenge is further compounded by uncertainties in patient-specific properties, particularly blood perfusion rates that fluctuate during tissue heating. We aim at predicting the temperature distribution inside a mono-dimensional domain represented by an ideal needle that enters perpendicularly from the surface of the patient, relying only on superficial boundary measurements. Given the uncertainties in the properties of the patient, this study is framed as a state estimation problem of the governing physical law of heat transfer in biological systems—Pennes’ Bio-Heat Equation—rather than a direct solution of the equation itself. The methodology is based on a multiple-model adaptive estimation framework using an adaptive observer design. The validation process has highlighted some limitations in traditional numerical simulation approaches, that served as a prompt to implement the estimator using Deep Learning techniques. Specifically, PhysicsInformed Learning is leveraged, an approach particularly suited to scenarios with known governing equations but limited training data. This Physics-Informed approach offers three key advantages: elimination of domain discretization requirements, enabling continuous spatial-temporal probing; real-time prediction of the solution to partial differential equation; and scalability to higher-dimensional input spaces without the constraints of conventional numerical methods. The final section extends our methodology to wave equation solutions for the deformation of soft material, with applications in mixed reality surgical training environments. This interdisciplinary research integrates applied mathematics, deep learning, and control engineering to address complex clinical challenges.
Questa tesi affronta una sfida cruciale nell’ambito dell’ipertermia oncologica superficiale: la valutazione non invasiva ed in tempo reale della temperatura al target. Il trattamento viene utilizzato in combinazione con la Radioterapia o la Chemoterapia, specialmente per recidive e i tumori più resistenti, con l’obiettivo di aumentarne l’efficacia senza somministrare ulteriore dose al paziente. L’efficacia è strettamente connessa con il raggiungimento delle temperature prestabilite in maniera ripetibile tra i trattamenti. Allo stato attuale, l’unico metodo per monitorare il trattamento è rappresentato dalla misurazione invasiva tramite l’uso di termocoppie, un metodo che spesso viene evitato, poiché comporta notevole disagio per i pazienti e fornisce solo misurazioni discrete e difficili da localizzare. La situazione è ulteriormente complicata dalla presenza di incertezze nelle proprietà specifiche del paziente, in particolare nel tasso di perfusione sanguigna, che non è direttamente misurabile e varia durante il trattamento, influenzando notevolmente la distribuzione di temperatura. L’obiettivo consiste nell’ottenere predizioni della distribuzione della temperatura in tutto il dominio soggetto al trattamento utilizzando solo misurazioni al contorno, i dati clinici del setup e l’anatomia del paziente. La metodologia sviluppata consiste nella stima adattiva dell’equazione utilizzando un osservatore multi-modello. Il processo di validazione ha messo in evidenza alcune limitazioni dei tradizionali metodi di simulazione numerica, che hanno motivato il ricorso a tecniche di Deep Learning. In particolare, il Physics-Informed Learning si è dimostrato essere un approccio particolarmente adatto a scenari in cui le leggi fisiche sono note, mentre i dati per il training sono limitati ad alcuni punti o del tutto assenti. L’approccio Physics-Informed si distingue per tre vantaggi fondamentali: elimina la necessità di discretizzare il dominio spazio-temporale, permettendo un campionamento continuo; in secondo luogo, dopo un periodo di training, la soluzione alle equazioni è fornita in tempo reale; infine, lo stesso modello può incorporare variazioni dei parametri dell’equazione come input, garantendo così un’elevata adattabilità a diversi scenari clinici e anatomici. L’ultima parte di questo studio si concentra sulla risoluzione dell’equazione d’onda per la simulazione di materiali morbidi, con applicazioni nella formazione chirurgica in ambienti di realtà mista. Questa ricerca interdisciplinare integra matematica applicata, informatica e ingegneria per affrontare complesse sfide cliniche.
AI-based solution methods for PDEs with application to oncological hyperthermia
CAPPELLINI, GUGLIELMO
2025
Abstract
This thesis investigates the observability of temperature distribution within the Gross Tumor Volume during superficial hyperthermia treatment, an adjuvant therapy employed alongside radiotherapy or chemotherapy for relapses and treatmentresistant tumors. A critical clinical challenge is addressed: the real-time assessment of temperature at the target site—essential for treatment feedback. Current protocols rely on invasive thermocouples, causing significant patient discomfort while providing only discrete, imprecisely localized measurements. This challenge is further compounded by uncertainties in patient-specific properties, particularly blood perfusion rates that fluctuate during tissue heating. We aim at predicting the temperature distribution inside a mono-dimensional domain represented by an ideal needle that enters perpendicularly from the surface of the patient, relying only on superficial boundary measurements. Given the uncertainties in the properties of the patient, this study is framed as a state estimation problem of the governing physical law of heat transfer in biological systems—Pennes’ Bio-Heat Equation—rather than a direct solution of the equation itself. The methodology is based on a multiple-model adaptive estimation framework using an adaptive observer design. The validation process has highlighted some limitations in traditional numerical simulation approaches, that served as a prompt to implement the estimator using Deep Learning techniques. Specifically, PhysicsInformed Learning is leveraged, an approach particularly suited to scenarios with known governing equations but limited training data. This Physics-Informed approach offers three key advantages: elimination of domain discretization requirements, enabling continuous spatial-temporal probing; real-time prediction of the solution to partial differential equation; and scalability to higher-dimensional input spaces without the constraints of conventional numerical methods. The final section extends our methodology to wave equation solutions for the deformation of soft material, with applications in mixed reality surgical training environments. This interdisciplinary research integrates applied mathematics, deep learning, and control engineering to address complex clinical challenges.File | Dimensione | Formato | |
---|---|---|---|
Tesi_dottorato_Cappellini.pdf
accesso aperto
Dimensione
11.8 MB
Formato
Adobe PDF
|
11.8 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/212180
URN:NBN:IT:UNIROMA1-212180