This dissertation focuses on the development and validation of deep learning (DL)-based methodologies applied to the different stages of the ultrasound imaging acquisition and reconstruction chain, with the goal of improving image quality and automation, from raw signal processing to clinical interpretation. Ultrasound imaging is one of the most widespread diagnostic modalities due to its non-invasiveness, portability, and real-time acquisition capability. However, the process that transforms acoustic echoes into clinically meaningful images is complex and subject to several limitations, including noise contamination, trade-offs between image quality and hardware complexity, and strong dependence on the operator’s experience. Traditional signal processing methods, typically based on deterministic physical models, are robust and interpretable but often struggle to adapt to varying operating conditions and hardware constraints. In this context, deep learning introduces a data-driven approach capable of learning complex relationships directly from data and jointly optimizing the different stages of the imaging chain. Neural networks can enhance signal and image quality, reduce noise, and automate diagnostic tasks, achieving performance levels comparable to those of experienced operators. A first line of research addressed the suppression of electronic switching noise, an artifact introduced by the acquisition hardware that degrades the radio-frequency (RF) signals. A DL-based approach operating in the frequency domain was developed to identify and remove this noise directly from raw data, supported by an experimental protocol designed to collect paired noisy and clean signals. Subsequently, DL was applied to Synthetic Aperture (SA) imaging, a technique that combines multiple transmissions to improve spatial resolution. Two complementary approaches were proposed: the first, a beamforming-based method, maps RF data acquired with simplified configurations into images equivalent to those obtained with more complex systems; the second operates in the image domain, enhancing reconstructions produced by conventional methods through image-to-image mapping. Another research direction focused on three-dimensional imaging using sparse arrays, where a DL-based adaptive fusion framework was introduced to combine the outputs of multiple beamformers, optimizing the quality and consistency of volumetric reconstructions. Finally, the research addressed the automatic segmentation of echocardiographic images, a key step for the quantitative assessment of cardiac function. DL models were explored, with particular attention to Vision Transformer architectures, which demonstrated superior ability to capture long-range spatial dependencies and ensure anatomical consistency compared to traditional convolutional networks. Overall, the dissertation demonstrates how integrating deep learning into the ultrasound imaging chain can make ultrasound a more accurate, adaptive, and automated modality, enhancing diagnostic quality while reducing operator dependency.
Questo elaborato tratta lo sviluppo e la validazione di metodologie basate su deep learning (DL) applicate alle diverse fasi della catena di acquisizione e formazione delle immagini ecografiche, con l’obiettivo di migliorarne la qualità e l’automazione, dal segnale grezzo fino all’interpretazione clinica. L’imaging a ultrasuoni rappresenta una delle tecniche diagnostiche più diffuse per la sua natura non invasiva, portabilità e capacità di acquisizione in tempo reale. Tuttavia, il processo che trasforma gli echi acustici in immagini clinicamente utili è complesso e soggetto a diverse limitazioni, tra cui il rumore, i compromessi tra qualità e complessità hardware e la forte dipendenza dall’esperienza dell’operatore. I metodi tradizionali di elaborazione, basati su modelli fisici deterministici, offrono robustezza e interpretabilità, ma si adattano con difficoltà a condizioni operative variabili e vincoli tecnici. In questo contesto, il DL ha introdotto un approccio data-driven capace di apprendere relazioni complesse direttamente dai dati e di ottimizzare in modo congiunto le diverse fasi della catena di imaging. Le reti neurali possono migliorare la qualità del segnale e dell’immagine, ridurre il rumore e automatizzare compiti diagnostici, avvicinandosi alle prestazioni di operatori esperti. Una prima linea di ricerca ha riguardato la soppressione del rumore elettronico di commutazione, un artefatto introdotto dall’hardware che degrada i segnali a radiofrequenza (RF). È stato sviluppato un approccio DL nel dominio delle frequenze capace di riconoscere e rimuovere il rumore direttamente dai dati grezzi, supportato da un protocollo sperimentale per la raccolta di segnali rumorosi e puliti. Successivamente, il DL è stato applicato al miglioramento dell’imaging a apertura sintetica (Synthetic Aperture), una tecnica che combina più trasmissioni per aumentare la risoluzione spaziale. Sono stati proposti due approcci complementari: il primo, di tipo beamforming, mappa i dati RF acquisiti con configurazioni semplificate in immagini equivalenti a quelle di sistemi più complessi; il secondo, invece, opera nel dominio dell’immagine, migliorando le ricostruzioni ottenute con metodi convenzionali attraverso una mappatura da immagine a immagine. Un’ulteriore area di studio ha riguardato l’imaging tridimensionale con array sparsi, dove è stato introdotto un metodo di fusione adattiva basato su DL per combinare i risultati di diversi beamformer, ottimizzando la qualità e la coerenza delle ricostruzioni volumetriche. Infine, la ricerca ha affrontato la segmentazione automatica delle immagini ecocardiografiche, fondamentale per la valutazione della funzione cardiaca. Sono stati esplorati modelli DL, in particolare architetture basate su Vision Transformer, che mostrano una maggiore capacità di catturare relazioni spaziali a lungo raggio e di garantire coerenza anatomica rispetto alle reti convolutive tradizionali. Nel complesso, la tesi dimostra come l’integrazione del DL nella catena di imaging ecografico possa rendere l’ecografia una tecnica più accurata, adattiva e automatizzata, migliorando la qualità diagnostica e riducendo la dipendenza dall’operatore.
Development of deep learning-based methods for signal and image processing in ultrasound medical imaging
Bosco, Edoardo
2026
Abstract
This dissertation focuses on the development and validation of deep learning (DL)-based methodologies applied to the different stages of the ultrasound imaging acquisition and reconstruction chain, with the goal of improving image quality and automation, from raw signal processing to clinical interpretation. Ultrasound imaging is one of the most widespread diagnostic modalities due to its non-invasiveness, portability, and real-time acquisition capability. However, the process that transforms acoustic echoes into clinically meaningful images is complex and subject to several limitations, including noise contamination, trade-offs between image quality and hardware complexity, and strong dependence on the operator’s experience. Traditional signal processing methods, typically based on deterministic physical models, are robust and interpretable but often struggle to adapt to varying operating conditions and hardware constraints. In this context, deep learning introduces a data-driven approach capable of learning complex relationships directly from data and jointly optimizing the different stages of the imaging chain. Neural networks can enhance signal and image quality, reduce noise, and automate diagnostic tasks, achieving performance levels comparable to those of experienced operators. A first line of research addressed the suppression of electronic switching noise, an artifact introduced by the acquisition hardware that degrades the radio-frequency (RF) signals. A DL-based approach operating in the frequency domain was developed to identify and remove this noise directly from raw data, supported by an experimental protocol designed to collect paired noisy and clean signals. Subsequently, DL was applied to Synthetic Aperture (SA) imaging, a technique that combines multiple transmissions to improve spatial resolution. Two complementary approaches were proposed: the first, a beamforming-based method, maps RF data acquired with simplified configurations into images equivalent to those obtained with more complex systems; the second operates in the image domain, enhancing reconstructions produced by conventional methods through image-to-image mapping. Another research direction focused on three-dimensional imaging using sparse arrays, where a DL-based adaptive fusion framework was introduced to combine the outputs of multiple beamformers, optimizing the quality and consistency of volumetric reconstructions. Finally, the research addressed the automatic segmentation of echocardiographic images, a key step for the quantitative assessment of cardiac function. DL models were explored, with particular attention to Vision Transformer architectures, which demonstrated superior ability to capture long-range spatial dependencies and ensure anatomical consistency compared to traditional convolutional networks. Overall, the dissertation demonstrates how integrating deep learning into the ultrasound imaging chain can make ultrasound a more accurate, adaptive, and automated modality, enhancing diagnostic quality while reducing operator dependency.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_bosco_final_xCaricamento.pdf
embargo fino al 11/03/2027
Licenza:
Tutti i diritti riservati
Dimensione
7.06 MB
Formato
Adobe PDF
|
7.06 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/359466
URN:NBN:IT:UNIPV-359466