Computed tomography (CT) ranks amongst the most popular, non-invasive medical imaging techniques. It makes use of computer-processed combinations of X-ray measurements, taken from different angles, in order to provide a detailed reconstruction of the inner structure of the body. The focus of this thesis is a particular setup of CT, called limited-angle CT (LA-CT), where the X-ray measurements are restricted to a small angular range. This task, that arises naturally in a number of important, practical applications such as breast tomosynthesis or dental X-ray scanning, presents the advantage of lowering the X-ray radiation dose and reducing the scanning time, and has thus become one of the predominant research topics in CT. The main challenge inherent to LA-CT comes from the fact that the incompleteness of the collected data makes the reconstruction problem extremely ill-posed. As a consequence, classical methods, such as the filtered-backprojection (FBP), show poor performance and result most frequently in undesirable artefacts. Iterative algorithms as well as machine learning approaches have also been proposed in the literature to address this problem, but they still present drawbacks and limitations. The purpose of this thesis is to propose a novel algorithm, that combines both the trustworthiness of traditional, variational methods, and the powerful technology of deep learning strategies, in order to provide stable and reliable LA-CT reconstructions. More generally, the convolutional neural network (CNN) presented in this work, called $\Psi$DONet, is designed to learn pseudodifferential operators ($\Psi$DOs) in the broad context of linear inverse problems. Thus, although $\Psi$DONet is investigated in this thesis in the special case of LA-CT, the theoretical results presented here can be extended to the broader case of convolutional Fourier integral operators (FIOs) and $\Psi$DOs. We formulate the LA-CT reconstruction problem as a regularised optimisation problem, in which the objective function to be minimised is the sum of a data-fidelity measure and a regularisation term that promotes sparsity in the wavelet basis. A well-known iterative technique for the solution of such a problem is the Iterative Soft-Thresholding Algorithm (ISTA) which, in the case of LA-CT, involves a $\Psi$DO at each of its iterations. The convolutional nature of this very operator makes it possible to implement the unfolded iterations of ISTA as the successive layers of a CNN. We show, furthermore, that it is possible to compute the exact values of the parameters of the CNN in such a way that it reproduces the behaviour of standard ISTA, or a perturbation thereof. The strength of $\Psi$DONet thus rests upon the fact that its parameters can be initialised with such values, and then trained through a learning process made particularly efficient thanks to the CNN technology. Two implementations of $\Psi$DONet are investigated: Filter- Based $\Psi$DONet ($\Psi$DONet-F), where the pseudodifferential operator is approximated by means of a set of filters, whose central part is trainable; and Operator-Based $\Psi$DONet ($\Psi$DONet-O), where the pseudodifferential operator is not approximated but explicitly computed, and the learnable parameters are implemented as an additional operator. Numerical tests are conducted on different datasets of simulated data from limited-angle geometry. Both implementations provide similarly good and noteworthy results that clearly outperform the quality of standard ISTA reconstructions, the main difference being a greater computational efficiency for $\Psi$DONet-O. The presented approach offers promising perspectives and paves the way to applying the same idea to other convolutional FIOs or $\Psi$DOs.
La tomografia computerizzata (CT) si colloca tra le tecniche di imaging medico non invasivo più diffuse. Si basa su combinazioni elaborate al computer di misurazioni a raggi X, prese da diverse angolazioni, al fine di fornire una ricostruzione dettagliata della struttura interna del corpo. Il fulcro di questa tesi è una particolare configurazione della CT, chiamata CT ad angolo limitato (LA-CT), dove le misurazioni dei raggi X sono limitate ad un piccolo intervallo angolare. Questa, che si presenta naturalmente in una serie di applicazioni pratiche come la tomosintesi mammaria o la radiografia dentale a raggi X, offre il vantaggio di abbassare la dose di radiazioni di raggi X e di ridurre il tempo di scansione, ed è così diventato uno dei temi di ricerca predominanti in CT. La sfida principale intrinseca alla LA-CT deriva dal fatto che l'incompletezza dei dati raccolti rende il problema della ricostruzione estremamente mal posto. Di conseguenza, i metodi classici, come la filtered-backprojection (FBP), presentano scarse prestazioni e si traducono solitamente in artefatti indesiderati. Algoritmi iterativi e approcci di apprendimento automatico sono anche stati proposti in letteratura per affrontare questo problema, ma presentano ancora svantaggi e limitazioni. Lo scopo di questa tesi è di proporre un nuovo algoritmo, che combina sia l'affidabilità dei metodi variazionali tradizionali, sia la potente tecnologia delle strategie di apprendimento profondo, al fine di fornire ricostruzioni LA-CT stabili e attendibili. Più generalemente, la rete neurale convoluzionale (CNN) presentata in questo lavoro, chiamata $\Psi$DONet, è progettata per apprendere gli operatori pseudodifferenziali ($\Psi$DOs) nel contesto generico dei problemi inversi lineari. Pertanto, sebbene $\Psi$DONet sia indagato in questa tesi nel caso speciale di LA-CT, i risultati teorici qui presentati possono essere estesi al caso più ampio degli operatori integrali di Fourier (FIOs) e $\Psi$DOs convoluzionali. Formuliamo il problema di ricostruzione LA-CT come un problema di ottimizzazione regolarizzato, in cui la funzione obiettivo da minimizzare è la somma di una misura di fedeltà ai dati e un termine di regolarizzazione che promuove la sparsità nella base della wavelet. Una tecnica iterativa nota per la soluzione di tale problema è l'Iterative Soft-Thresholding Algorithm (ISTA) che, nel caso di LA-CT, coinvolge un $\Psi$DO a ciascuna delle sue iterazioni. La natura convoluzionale di questo operatore consente di implementare le iterazioni 'spiegate' di ISTA come i successivi strati di una CNN. Mostriamo, inoltre, che è possibile calcolare i valori esatti dei parametri della CNN in modo tale da riprodurre il comportamento di standard ISTA, o una sua perturbazione. Il punto di forza di $\Psi$DONet risiede quindi nel fatto che i suoi parametri possono essere inizializzati con tali valori, e poi addestrati attraverso un processo di apprendimento reso particolarmente efficiente grazie alla tecnologia dei CNNs. Vengono indagate due implementazioni di $\Psi$DONet: Filter- Based $\Psi$DONet ($\Psi$DONet-F), dove l'operatore pseudodifferenziale è approssimato mediante un insieme di filtri, la cui parte centrale è addestrabile; e Operator-Based $\Psi$DONet ($\Psi$DONet-O), dove l'operatore pseudodifferenziale non è approssimato ma esplicitamente calcolato, e i parametri apprendibili sono implementati come un operatore aggiuntivo. I test numerici vengono condotti su diversi set di dati simulati con geometria ad angolo limitato. Entrambe le implementazioni forniscono risultati simili e degni di nota, che superano nettamente la qualità delle ricostruzioni ISTA standard, essendo la differenza principale una maggiore efficienza computazionale per $\Psi$DONet-O. L'approccio presentato offre prospettive promettenti ed apre la strada per applicare la stessa idea ad altri FIOs o $\Psi$DOs convoluzionali.
Reti neurali convoluzionali per problemi inversi con operatori pseudodifferenziali: applicazione alla tomografia ad angolo limitato / Mathilde Emmanuelle Galinier , 2021 Feb 26. 33. ciclo, Anno Accademico 2019/2020.
Reti neurali convoluzionali per problemi inversi con operatori pseudodifferenziali: applicazione alla tomografia ad angolo limitato
GALINIER, Mathilde Emmanuelle
2021
Abstract
Computed tomography (CT) ranks amongst the most popular, non-invasive medical imaging techniques. It makes use of computer-processed combinations of X-ray measurements, taken from different angles, in order to provide a detailed reconstruction of the inner structure of the body. The focus of this thesis is a particular setup of CT, called limited-angle CT (LA-CT), where the X-ray measurements are restricted to a small angular range. This task, that arises naturally in a number of important, practical applications such as breast tomosynthesis or dental X-ray scanning, presents the advantage of lowering the X-ray radiation dose and reducing the scanning time, and has thus become one of the predominant research topics in CT. The main challenge inherent to LA-CT comes from the fact that the incompleteness of the collected data makes the reconstruction problem extremely ill-posed. As a consequence, classical methods, such as the filtered-backprojection (FBP), show poor performance and result most frequently in undesirable artefacts. Iterative algorithms as well as machine learning approaches have also been proposed in the literature to address this problem, but they still present drawbacks and limitations. The purpose of this thesis is to propose a novel algorithm, that combines both the trustworthiness of traditional, variational methods, and the powerful technology of deep learning strategies, in order to provide stable and reliable LA-CT reconstructions. More generally, the convolutional neural network (CNN) presented in this work, called $\Psi$DONet, is designed to learn pseudodifferential operators ($\Psi$DOs) in the broad context of linear inverse problems. Thus, although $\Psi$DONet is investigated in this thesis in the special case of LA-CT, the theoretical results presented here can be extended to the broader case of convolutional Fourier integral operators (FIOs) and $\Psi$DOs. We formulate the LA-CT reconstruction problem as a regularised optimisation problem, in which the objective function to be minimised is the sum of a data-fidelity measure and a regularisation term that promotes sparsity in the wavelet basis. A well-known iterative technique for the solution of such a problem is the Iterative Soft-Thresholding Algorithm (ISTA) which, in the case of LA-CT, involves a $\Psi$DO at each of its iterations. The convolutional nature of this very operator makes it possible to implement the unfolded iterations of ISTA as the successive layers of a CNN. We show, furthermore, that it is possible to compute the exact values of the parameters of the CNN in such a way that it reproduces the behaviour of standard ISTA, or a perturbation thereof. The strength of $\Psi$DONet thus rests upon the fact that its parameters can be initialised with such values, and then trained through a learning process made particularly efficient thanks to the CNN technology. Two implementations of $\Psi$DONet are investigated: Filter- Based $\Psi$DONet ($\Psi$DONet-F), where the pseudodifferential operator is approximated by means of a set of filters, whose central part is trainable; and Operator-Based $\Psi$DONet ($\Psi$DONet-O), where the pseudodifferential operator is not approximated but explicitly computed, and the learnable parameters are implemented as an additional operator. Numerical tests are conducted on different datasets of simulated data from limited-angle geometry. Both implementations provide similarly good and noteworthy results that clearly outperform the quality of standard ISTA reconstructions, the main difference being a greater computational efficiency for $\Psi$DONet-O. The presented approach offers promising perspectives and paves the way to applying the same idea to other convolutional FIOs or $\Psi$DOs.File | Dimensione | Formato | |
---|---|---|---|
PhD_thesis.pdf
Open access
Descrizione: Tesi definitiva Galinier Mathilde
Tipologia:
Tesi di dottorato
Dimensione
5.22 MB
Formato
Adobe PDF
|
5.22 MB | Adobe PDF | Visualizza/Apri |
Pubblicazioni consigliate
I metadati presenti in IRIS UNIMORE sono rilasciati con licenza Creative Commons CC0 1.0 Universal, mentre i file delle pubblicazioni sono rilasciati con licenza Attribuzione 4.0 Internazionale (CC BY 4.0), salvo diversa indicazione.
In caso di violazione di copyright, contattare Supporto Iris