Cancer is considered nowadays one of the leading causes of death in the world. Radiation and particle therapies are well-established techniques to treat the disease, however they suffer the major limitation of delivering therapeutic dose to healthy tissues surrounding the tumor. A different therapy modality is the Neutron Capture Therapy (NCT), whose rapid expansion is due to the recent advancements in neutron-producing accelerators, and in particular Boron Neutron Capture Therapy (BNCT), which target the tumor with 10-B pharmaceuticals and irradiate the patient with epithermal and thermal neutrons. When neutrons undergo a capture reaction with boron in BNCT, high Linear Energy Transfer (LET) particles are released within the tumor cell, leading to its destruction and making the dose delivering highly selective. However, since the local boron concentration cannot be directly measured in vivo, determining the optimal neutron irradiation time and the corresponding absorbed radiation dose remains one of the major challenges in BNCT. Prompt Gamma-ray Single Photon Emission Computed Tomography (PG-SPECT) addresses this issue by employing the detection of the 478 keV prompt gamma rays emitted as a by-product of the neutron capture reaction, enabling real-time imaging of the boron distribution without the need for additional radio-tracers. This approach allows for a direct estimation of the dose delivered to the patient. The RadLab Laboratory of Politecnico di Milano has recently developed a PG-SPECT imaging system. However, the adoption of a new collimator, introduced for enhancing the shielding performance, has also increased the acceptance angle for the gamma rays reaching BeNEdiCTE (Boron NEutron CapTurE) detector. This modification significantly amplifies parallax errors during image reconstruction. Previous studies implemented an Artificial Neural Network (ANN) on the detector's FPGA board to determine the 2D gamma ray interaction locations within the scintillator crystal on-line. The aim of this work is to extend the ANN to simultaneously predict the Depth Of Interaction (DOI), enabling parallax-error correction, while keeping the architectural changes minimal to preserve FPGA compatibility. For the ANN training, real experimental data were preferred over Monte Carlo (MC) simulations to ensure consistency with the detector performance under clinically realistic operating conditions. This choice, however, posed a significant challenge in obtaining a 3D-labeled dataset. To overcome this, multiple 2D datasets were acquired by irradiating the crystal from different orientations. An automated GUI was therefore developed, leading to the acquisition of the front face and two lateral faces in significantly reduced acquisition times. The proposed solution adopts a "combined ANN" approach, which can be exemplified by the following procedure: first, an intermediary ANN is trained with the lateral dataset (labeled with Y,Z coordinates); second, this intermediary ANN is used to predict the Z coordinate of the events contained in the frontal dataset (labeled with X,Y coordinates), creating a 3D-labeled dataset; third, a final ANN is trained on this 3D-labeled dataset. In practice, three variations of this approach were designed and evaluated. Furthermore, the Z predictions obtained from these networks were used to perform parallax-error correction in the tomographic reconstruction of two boron-filled vials, starting from their projections acquired with the BeNediCTE module at LENA nuclear reactor (Pavia). The selected ANN achieves a FWHM of 3.6 mm for the X and Y estimations, and 5.4 mm for the Z estimation across the whole crystal. The X and Y performance degraded by approximately 0.7 mm with respect to the previous two-output ANN, due to the additional Z-prediction task, which reallocates part of the ANN’s learning capacity. Nevertheless, the tomographic reconstructions with DOI correction demonstrate superior geometric accuracy. Finally, an explainability analysis was performed on the selected network to investigate the underlying prediction process of the three coordinates, showing strong consistency with the geometrical properties of the system and the mathematical estimation techniques reported in the literature.
Il cancro è considerato al giorno d'oggi una delle principali cause di morte in tutto il mondo. La radioterapia e l'adroterapia sono terapie affermate nel mondo clinico, che però causano l'erogazione di una parte della dose terapeutica ai tessuti sani circondanti il tumore. Una terapia alternativa è la Neutron Capture Therapy (NCT), la cui recente rapida espansione è dovuta agli ultimi avanzamenti sugli acceleratori di neutroni. La NCT comprende in particolare la Boron Neutron Capture Therapy (BNCT), la quale consiste nel marcare le cellule tumorali con composti farmaceutici a base di 10-B e successivamente irradiare il paziente con un flusso di neutroni termali o epitermali. Quando i neutroni reagiscono con il boro attraverso una reazione di cattura, la reazione rilascia particelle a elevato coefficiente Linear Transfer Energy (LET), che comporta la distruzione delle cellule contenenti i composti farmaceutici, cioè le cellule tumorali, rendendo l'erogazione di dose altamente selettiva. Tuttavia, essendo impossibile misurare la concentrazione locale di boro in-vivo in maniera diretta, stimare il tempo di irraggiamento ottimale e la dose terapeutica assorbita dal paziente è una delle maggiori sfide della BNCT. La Prompt Gamma-ray Single Photon Emission Computed Tomography (PG-SPECT) affronta questo problema sfruttando il rilevamento dei raggi gamma a 478 keV emessi come prodotti secondari dalla reazione di cattura, permettendo l'imaging real-time della distribuzione del boro senza la necessità di ulteriori radiofarmaci. Questa tecnica permette quindi una stima diretta della dose terapeutica erogata al paziente. Il laboratorio RadLab del Politecnico di Milano ha recentemente sviluppato un sistema di imaging per la PG-SPECT. Tuttavia, l'adozione di un nuovo collimatore, introdotto per migliorare la schermatura, ha comportato l'aumento dell'angolo di accettazione dei raggi gamma da parte del rilevatore BeNEdiCTE (Boron NEutron CapTurE). Questa modifica ha aumentato significativamente gli errori di parallasse durante la ricostruzione delle immagini. Studi precedenti hanno implementato una Artificial Neural Network (ANN) sulla scheda FPGA del detector per determinare in tempo reale la posizione 2D degli eventi di scintillazione prodotti dall'interazione dei raggi gamma con il cristallo. L'obiettivo di questo lavoro è estendere la ANN affinché predica allo stesso modo la Depth Of Interaction (DOI) degli eventi, così da poter correggere gli errori di parallasse, mantenendo però minime le modifiche apportate alla sua architettura per preservare la compatibilità sull'FPGA. Per il training della nuova ANN, si è preferito usare dati sperimentali reali rispetto a dati provenienti da simulazioni Monte Carlo (MC) per garantire medesime performance del network in condizioni clinicamente realistiche. Questa scelta, tuttavia, comporta significative difficoltà nell'ottenere un dataset con label 3D degli eventi da usare per il training. Di conseguenza, più dataset con label 2D sono stati acquisiti irraggiando il cristallo da diverse orientazioni. A tal fine, è stata sviluppata una Graphical User Interface (GUI) per acquisizioni automatiche, che ha consentito di raccogliere i dati della faccia frontale e delle due facce laterali del cristallo in tempi significativamente ridotti. La soluzione proposta adotta un approccio "combined ANN", che può essere spiegato a titolo di esempio come di seguito: prima, una ANN intermediaria è allenata sul dataset laterale (con label Y,Z); secondo, questa ANN intermediaria viene usata per predire le coordinate Z degli eventi contenuti nel dataset frontale (con label X,Y); terzo, una ANN finale è allenata su quest'ultimo dataset con le tre label X,Y,Z. Tre variazioni di questo approccio sono state progettate e provate. Inoltre, le predizioni della coordinata Z ottenute da questi network sono state usate per correggere gli errori di parallasse durante delle ricostruzioni tomografiche di due provette contenenti boro, le cui proiezioni sono state acquisite con BeNediCTE presso il reattore nucleare LENA di Pavia. La ANN infine selezionata raggiunge una Full Width Half Maximum (FWHM) di 3.6 mm sulle predizioni X e Y, a una FWHM di 5.4 mm sulle predizioni Z considerando l'intero cristallo. Le performance lungo gli assi X e Y risultano leggermente inferiori, ridotte di circa 0.7 mm, rispetto alla ANN precedente a due output, a causa del compito aggiuntivo di predire la coordinata Z, che ridistribuisce le capacità di apprendimento della rete. Ciononostante, le ricostruzioni tomografiche basate sulle correzioni DOI hanno mostrato una superiore accuratezza geometrica. Infine, la rete selezionata è stata analizzata con tecniche di explainability per investigare i processi sottostanti alla predizione delle tre coordinate, dimostrando una forte coerenza con le proprietà geometriche attese del sistema e con le tecniche matematiche di predizione riportate in letteratura.
Development of a 3D position reconstruction ANN for scintillation events in BNCT-SPECT applications
POMPIGNA, LEONARDO
2025/2026
Abstract
Cancer is considered nowadays one of the leading causes of death in the world. Radiation and particle therapies are well-established techniques to treat the disease, however they suffer the major limitation of delivering therapeutic dose to healthy tissues surrounding the tumor. A different therapy modality is the Neutron Capture Therapy (NCT), whose rapid expansion is due to the recent advancements in neutron-producing accelerators, and in particular Boron Neutron Capture Therapy (BNCT), which target the tumor with 10-B pharmaceuticals and irradiate the patient with epithermal and thermal neutrons. When neutrons undergo a capture reaction with boron in BNCT, high Linear Energy Transfer (LET) particles are released within the tumor cell, leading to its destruction and making the dose delivering highly selective. However, since the local boron concentration cannot be directly measured in vivo, determining the optimal neutron irradiation time and the corresponding absorbed radiation dose remains one of the major challenges in BNCT. Prompt Gamma-ray Single Photon Emission Computed Tomography (PG-SPECT) addresses this issue by employing the detection of the 478 keV prompt gamma rays emitted as a by-product of the neutron capture reaction, enabling real-time imaging of the boron distribution without the need for additional radio-tracers. This approach allows for a direct estimation of the dose delivered to the patient. The RadLab Laboratory of Politecnico di Milano has recently developed a PG-SPECT imaging system. However, the adoption of a new collimator, introduced for enhancing the shielding performance, has also increased the acceptance angle for the gamma rays reaching BeNEdiCTE (Boron NEutron CapTurE) detector. This modification significantly amplifies parallax errors during image reconstruction. Previous studies implemented an Artificial Neural Network (ANN) on the detector's FPGA board to determine the 2D gamma ray interaction locations within the scintillator crystal on-line. The aim of this work is to extend the ANN to simultaneously predict the Depth Of Interaction (DOI), enabling parallax-error correction, while keeping the architectural changes minimal to preserve FPGA compatibility. For the ANN training, real experimental data were preferred over Monte Carlo (MC) simulations to ensure consistency with the detector performance under clinically realistic operating conditions. This choice, however, posed a significant challenge in obtaining a 3D-labeled dataset. To overcome this, multiple 2D datasets were acquired by irradiating the crystal from different orientations. An automated GUI was therefore developed, leading to the acquisition of the front face and two lateral faces in significantly reduced acquisition times. The proposed solution adopts a "combined ANN" approach, which can be exemplified by the following procedure: first, an intermediary ANN is trained with the lateral dataset (labeled with Y,Z coordinates); second, this intermediary ANN is used to predict the Z coordinate of the events contained in the frontal dataset (labeled with X,Y coordinates), creating a 3D-labeled dataset; third, a final ANN is trained on this 3D-labeled dataset. In practice, three variations of this approach were designed and evaluated. Furthermore, the Z predictions obtained from these networks were used to perform parallax-error correction in the tomographic reconstruction of two boron-filled vials, starting from their projections acquired with the BeNediCTE module at LENA nuclear reactor (Pavia). The selected ANN achieves a FWHM of 3.6 mm for the X and Y estimations, and 5.4 mm for the Z estimation across the whole crystal. The X and Y performance degraded by approximately 0.7 mm with respect to the previous two-output ANN, due to the additional Z-prediction task, which reallocates part of the ANN’s learning capacity. Nevertheless, the tomographic reconstructions with DOI correction demonstrate superior geometric accuracy. Finally, an explainability analysis was performed on the selected network to investigate the underlying prediction process of the three coordinates, showing strong consistency with the geometrical properties of the system and the mathematical estimation techniques reported in the literature.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Pompigna_Tesi.pdf
non accessibile
Descrizione: Tesi
Dimensione
15.81 MB
Formato
Adobe PDF
|
15.81 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Pompigna_Executive_summary.pdf
non accessibile
Descrizione: Executive summary
Dimensione
1.59 MB
Formato
Adobe PDF
|
1.59 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/10589/247403