Image-guided navigation is central to minimally invasive cardiovascular interventions. Fluoroscopy provides real-time device visualization but lacks soft-tissue information and involves ionizing radiation, whereas ultrasound (US) provides anatomical detail without radiation but suffers from noise and limited field of view. Reliable alignment between preoperative CT models and intraoperative US remains a key challenge for simulation-assisted navigation. This thesis presents a CT–US registration framework for updating patient-specific vascular geometries using robot-tracked US acquisitions. The pipeline integrates automatic vessel segmentation from 2D US via convolutional neural networks (CNN), 3D surface reconstruction from tracked sweeps, and a two-stage registration strategy introducing a symmetry-aware rigid alignment based on Iterative Closest Point (ICP) with deformable refinement using Coherent Point Drift (CPD). The framework is evaluated \textit{in vitro} on custom vascular phantoms reproducing artificial and patient-specific geometries, assessing rigid alignment accuracy, initialization sensitivity and the contribution of deformable refinement. The proposed rigid strategy improves robustness to initialization over single-start ICP, while deformable refinement consistently reduces residual alignment errors, achieving millimetric accuracy within intraoperative time and non-invasiveness constraints. Since registration validation is limited to phantom settings, segmentation — a key pipeline component — is analysed to characterize variability affecting the overall framework. Multiple architectures are compared across phantom and clinical datasets to assess how dataset heterogeneity, acquisition variability and vessel type affect segmentation reliability within the pipeline. Simpler CNN models provide more consistent boundary delineation across heterogeneous and data-limited conditions, reducing variability that would otherwise propagate to registration. These results demonstrate the feasibility of combining US-based segmentation with multimodal point cloud registration into a coherent workflow for simulation-driven endovascular navigation.
La navigazione guidata da immagini è centrale nelle procedure cardiovascolari mini-invasive. La fluoroscopia consente la visualizzazione in tempo reale dei dispositivi ma non fornisce informazioni sui tessuti molli ed espone a radiazioni ionizzanti, mentre l’ecografia (US) è affetta da rumore e campo visivo limitato. L’allineamento affidabile tra modelli TC preoperatori e dati US intraoperatori rappresenta quindi una sfida cruciale per la navigazione endovascolare. Questa tesi propone un framework di registrazione TC–US per l’aggiornamento di geometrie vascolari paziente-specifiche mediante acquisizioni US tracciate da robot. La pipeline integra la segmentazione automatica del vaso in immagini US 2D tramite reti neurali convoluzionali (CNN), la ricostruzione 3D della superficie da sweep tracciati e una registrazione in due fasi che combina un allineamento rigido basato su Iterative Closest Point (ICP) con una successiva fase deformabile mediante Coherent Point Drift (CPD). La valutazione sperimentale \textit{in vitro} su fantocci vascolari, riproducenti geometrie artificiali e paziente-specifiche, analizza accuratezza dell’allineamento rigido, sensibilità all’inizializzazione e contributo della fase deformabile. La strategia rigida proposta migliora la robustezza rispetto a ICP \textit{single-start}, mentre CPD ne riduce sistematicamente l’errore residuo, raggiungendo accuratezza millimetrica compatibile con vincoli intraoperatori di tempo e non invasività. Poiché la registrazione è testabile solo in contesti controllati su fantoccio, la segmentazione — componente chiave della pipeline — è analizzata anche su dataset clinici per caratterizzare le principali fonti di variabilità che possono influenzare l’intero sistema. Diverse architetture sono confrontate per valutare l’effetto di eterogeneità del dataset e variabilità di acquisizione sull’affidabilità della segmentazione. Modelli CNN più semplici mostrano maggiore stabilità in condizioni eterogenee e a limitata disponibilità di dati, limitando la propagazione della variabilità alla registrazione. Nel complesso, i risultati dimostrano la fattibilità dell’integrazione tra segmentazione US e registrazione multimodale di point cloud in un workflow coerente per la navigazione endovascolare assistita da simulazione.
Development and evaluation of an automated intraoperative CT-US registration pipeline for endovascular navigation
Vantini, Tommaso
2024/2025
Abstract
Image-guided navigation is central to minimally invasive cardiovascular interventions. Fluoroscopy provides real-time device visualization but lacks soft-tissue information and involves ionizing radiation, whereas ultrasound (US) provides anatomical detail without radiation but suffers from noise and limited field of view. Reliable alignment between preoperative CT models and intraoperative US remains a key challenge for simulation-assisted navigation. This thesis presents a CT–US registration framework for updating patient-specific vascular geometries using robot-tracked US acquisitions. The pipeline integrates automatic vessel segmentation from 2D US via convolutional neural networks (CNN), 3D surface reconstruction from tracked sweeps, and a two-stage registration strategy introducing a symmetry-aware rigid alignment based on Iterative Closest Point (ICP) with deformable refinement using Coherent Point Drift (CPD). The framework is evaluated \textit{in vitro} on custom vascular phantoms reproducing artificial and patient-specific geometries, assessing rigid alignment accuracy, initialization sensitivity and the contribution of deformable refinement. The proposed rigid strategy improves robustness to initialization over single-start ICP, while deformable refinement consistently reduces residual alignment errors, achieving millimetric accuracy within intraoperative time and non-invasiveness constraints. Since registration validation is limited to phantom settings, segmentation — a key pipeline component — is analysed to characterize variability affecting the overall framework. Multiple architectures are compared across phantom and clinical datasets to assess how dataset heterogeneity, acquisition variability and vessel type affect segmentation reliability within the pipeline. Simpler CNN models provide more consistent boundary delineation across heterogeneous and data-limited conditions, reducing variability that would otherwise propagate to registration. These results demonstrate the feasibility of combining US-based segmentation with multimodal point cloud registration into a coherent workflow for simulation-driven endovascular navigation.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Vantini_ExecutiveSummary.pdf
accessibile in internet per tutti
Descrizione: Executive summary
Dimensione
934.84 kB
Formato
Adobe PDF
|
934.84 kB | Adobe PDF | Visualizza/Apri |
|
2026_03_Vantini_Thesis.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
23.41 MB
Formato
Adobe PDF
|
23.41 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/10589/253125