Endovascular procedures represent a widely adopted minimally inva sive approach for the treatment of cardiovascular diseases, such as carotid artery stenosis (CAS). These interventions rely on the navigation of flexible instruments through complex vascular anatomies to reach specific target sites. Despite their clinical advantages, they still present several challenges, including risks of compli cations such as vessel perforation, as well as prolonged exposure of the surgical team to ionizing radiation. Moreover, procedural outcomes strongly depend on the operator’s experience. In particular, anatomical variability, such as different aortic arch configurations, significantly affects the difficulty of catheterization and the success rate of the procedure. These limitations have motivated the develop ment of robotic and autonomous solutions aimed at improving precision, reducing operator dependency, and enhancing procedural safety. In this context, this the sis presents the development of a simulation framework for AI-based autonomous endovascular navigation. The system is implemented in Unity and reproduces a realistic geometry of the aortic arch, enabling the simulation of guidewire naviga tion under different anatomical conditions. The guidewire is modeled as a flexible structure composed of interconnected elements, enabling its interaction with vessel walls. A reinforcement learning approach based on Proximal Policy Optimization (PPO) and Soft-Actor Critic (SAC) is adopted to train an agent to autonomously navigate the guidewire from the descending aorta to target branches of the aortic arch, mimicking clinical scenarios such as carotid artery stenting, which represents a technically demanding step for inexperienced operators. Additionally, the frame work includes a dedicated actuator simulation module for analyzing the control system independently of the anatomical environment. The results demonstrate the feasibility of learning effective navigation policies within a simulated environment, highlighting the potential of AI-based approaches for robotic catheterization. This work represents an initial step within a broader research effort carried out at the Robot-Assisted Surgery (RAS) research group at KU Leuven.
Le procedure endovascolari rappresentano un approccio minimamente invasivo ampiamente adottato per il trattamento delle malattie cardiovascolari, come la stenosi dell’arteria carotide (CAS). Questi interventi si basano sulla navigazione di strumenti flessibili attraverso anatomie vascolari complesse per raggiungere specifici siti target. Nonostante i loro vantaggi clinici, presentano ancora diverse criticità, tra cui il rischio di complicazioni come la perforazione dei vasi, nonché una prolungata esposizione del team chirurgico a radiazioni ionizzanti. Inoltre, gli esiti procedurali dipendono fortemente dall’esperienza dell’operatore. In particolare, la variabilità anatomica, come le diverse configurazioni dell’arco aortico, influisce significativa mente sulla difficoltà della cateterizzazione e sul tasso di successo della procedura. Queste limitazioni hanno motivato lo sviluppo di soluzioni robotiche e autonome volte a migliorare la precisione, ridurre la dipendenza dall’operatore e aumentare la sicurezza della procedura. In questo contesto, la presente tesi presenta lo sviluppo di un framework di simulazione per la navigazione endovascolare autonoma basata su intelligenza artificiale. Il sistema è implementato in Unity e riproduce modelli realistici dell’arco aortico, consentendo la simulazione della navigazione della guida (guidewire) in diverse condizioni anatomiche. La guida è modellata come una struttura flessibile composta da elementi interconnessi, permettendo l’interazione con le pareti dei vasi. Un approccio di reinforcement learning basato su Proximal Policy Optimization (PPO) e Soft Actor-Critic (SAC) è adottato per addestrare un agente a navigare autonomamente la guida dall’aorta discendente verso i rami target dell’arco aortico, simulando scenari clinici come lo stenting carotideo. Inoltre, il framework include un modulo dedicato alla simulazione dell’attuazione, utilizzato per analizzare il sistema di controllo indipendentemente dall’ambiente anatomico. I risultati dimostrano la fattibilità dell’apprendimento di politiche di navigazione efficaci all’interno di un ambiente simulato, evidenziando il potenziale degli approcci basati su intelligenza artificiale per la cateterizzazione robotica. Questo lavoro rappresenta un primo passo all’interno di un più ampio progetto di ricerca condotto presso il gruppo di Robot-Assisted Surgery (RAS) della KU Leuven.
Autonomous endovascular guidewire navigation in a simulated unity environment
MONTALBANO, ANDREA
2025/2026
Abstract
Endovascular procedures represent a widely adopted minimally inva sive approach for the treatment of cardiovascular diseases, such as carotid artery stenosis (CAS). These interventions rely on the navigation of flexible instruments through complex vascular anatomies to reach specific target sites. Despite their clinical advantages, they still present several challenges, including risks of compli cations such as vessel perforation, as well as prolonged exposure of the surgical team to ionizing radiation. Moreover, procedural outcomes strongly depend on the operator’s experience. In particular, anatomical variability, such as different aortic arch configurations, significantly affects the difficulty of catheterization and the success rate of the procedure. These limitations have motivated the develop ment of robotic and autonomous solutions aimed at improving precision, reducing operator dependency, and enhancing procedural safety. In this context, this the sis presents the development of a simulation framework for AI-based autonomous endovascular navigation. The system is implemented in Unity and reproduces a realistic geometry of the aortic arch, enabling the simulation of guidewire naviga tion under different anatomical conditions. The guidewire is modeled as a flexible structure composed of interconnected elements, enabling its interaction with vessel walls. A reinforcement learning approach based on Proximal Policy Optimization (PPO) and Soft-Actor Critic (SAC) is adopted to train an agent to autonomously navigate the guidewire from the descending aorta to target branches of the aortic arch, mimicking clinical scenarios such as carotid artery stenting, which represents a technically demanding step for inexperienced operators. Additionally, the frame work includes a dedicated actuator simulation module for analyzing the control system independently of the anatomical environment. The results demonstrate the feasibility of learning effective navigation policies within a simulated environment, highlighting the potential of AI-based approaches for robotic catheterization. This work represents an initial step within a broader research effort carried out at the Robot-Assisted Surgery (RAS) research group at KU Leuven.| File | Dimensione | Formato | |
|---|---|---|---|
|
Andrea_Montalbano_AF.pdf
accessibile in internet per tutti
Descrizione: Article Format Autonomous Endovascular Guidewire Navigation in a Simulated Unity Environment
Dimensione
11.88 MB
Formato
Adobe PDF
|
11.88 MB | Adobe PDF | Visualizza/Apri |
|
Andrea_Montalbano_ES.pdf
accessibile in internet per tutti
Descrizione: Executive Summary Autonomous Endovascular Guidewire Navigation in a Simulated Unity Environment
Dimensione
4.01 MB
Formato
Adobe PDF
|
4.01 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/10589/261412