Boron Neutron Capture Therapy (BNCT) requires an accurate patient positioning, as geometric variations can alter the dose distribution between the Gross Tumor Volume (GTV) and the organs at risk (OAR). This work develops a numerical framework for optimizing patient setup through rigid rotations and translations of anatomical volumes, with the goal of maximizing the dosimetric advantage in favor of the GTV. The entire pipeline was implemented in Python, leveraging CuPy for GPU-accelerated voxel-wise operations and PyTorch for differentiable 3D transformations via sampling grids. The adopted score function combines the mean GTV dose, OAR dose penalties, and geometric constraints related to forbidden zones and tumor coverage. Several optimization algorithms were implemented and compared, including Grid Search, Guided Random Walk, Differential Evolution, Cross-Entropy Method, and Adam. They were evaluated both on a cylindrical phantom with the tumor placed at different depths and on a real clinical CT dataset. Results show that stochastic optimizers explore the roto-translation space more efficiently, while Adam converges more rapidly near local minima. In all examined scenarios, geometric optimization improves GTV dose while reducing irradiation to OARs. This work demonstrates the feasibility and effectiveness of a GPU-accelerated optimization approach based on rigid transformations, paving the way for future integration with machine-learning-based methods and predictive models for automatic BNCT treatment control.
La Boron Neutron Capture Therapy (BNCT) richiede un’elevata precisione nel posizionamento del paziente, poiché variazioni geometriche possono modificare la distribuzione di dose tra il Gross Tumor Volume (GTV) e gli Organi A Rischio (OAR). In questo lavoro viene sviluppato un framework numerico per l’ottimizzazione del posizionamento mediante rotazioni e traslazioni rigide dei volumi anatomici, con l’obiettivo di massimizzare il rapporto dosimetrico a favore del GTV. L’intera pipeline è stata implementata in Python, sfruttando sia CuPy per l’accelerazione GPU delle operazioni voxel-wise, sia PyTorch per la gestione differenziabile delle trasformazioni 3D tramite griglie di campionamento. La funzione di score adottata combina la dose media al GTV, la penalizzazione della dose negli OAR e vincoli geometrici relativi alle zone proibite e alla copertura tumorale. Sono stati implementati e confrontati diversi algoritmi di ottimizzazione: Grid Search, Guided Random Walk, Differential Evolution, Cross-Entropy Method e Adam. Essi sono stati valutati sia su un fantoccio cilindrico con tumore a differenti profondità, sia su un caso clinico reale. I risultati mostrano che gli ottimizzatori stocastici esplorano in modo più efficiente lo spazio delle rototraslazioni, mentre Adam converge più rapidamente in prossimità dei minimi locali. In tutti gli scenari analizzati si osserva come l’ottimizzazione geometrica possa migliorare in modo significativo la dose al GTV riducendo l’irradiazione degli OAR. Questo lavoro dimostra la fattibilità e l’efficacia di un approccio di ottimizzazione del posizionamento basato su trasformazioni rigide e accelerato da GPU, aprendo la strada a future integrazioni con tecniche di machine learning e modelli predittivi per il controllo automatico del trattamento BNCT.
Development of an automatic beam positioning system for BNCT treatment planning
Catania, Alessandra
2024/2025
Abstract
Boron Neutron Capture Therapy (BNCT) requires an accurate patient positioning, as geometric variations can alter the dose distribution between the Gross Tumor Volume (GTV) and the organs at risk (OAR). This work develops a numerical framework for optimizing patient setup through rigid rotations and translations of anatomical volumes, with the goal of maximizing the dosimetric advantage in favor of the GTV. The entire pipeline was implemented in Python, leveraging CuPy for GPU-accelerated voxel-wise operations and PyTorch for differentiable 3D transformations via sampling grids. The adopted score function combines the mean GTV dose, OAR dose penalties, and geometric constraints related to forbidden zones and tumor coverage. Several optimization algorithms were implemented and compared, including Grid Search, Guided Random Walk, Differential Evolution, Cross-Entropy Method, and Adam. They were evaluated both on a cylindrical phantom with the tumor placed at different depths and on a real clinical CT dataset. Results show that stochastic optimizers explore the roto-translation space more efficiently, while Adam converges more rapidly near local minima. In all examined scenarios, geometric optimization improves GTV dose while reducing irradiation to OARs. This work demonstrates the feasibility and effectiveness of a GPU-accelerated optimization approach based on rigid transformations, paving the way for future integration with machine-learning-based methods and predictive models for automatic BNCT treatment control.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_alessandra_catania_executive_summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
69.36 kB
Formato
Adobe PDF
|
69.36 kB | Adobe PDF | Visualizza/Apri |
|
2025_12_alessandra_catania.pdf
non accessibile
Descrizione: Thesis
Dimensione
8.82 MB
Formato
Adobe PDF
|
8.82 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/10589/247136