Khun Jush F (2025)
Publication Language: English
Publication Type: Thesis
Publication year: 2025
URI: https://open.fau.de/handle/openfau/35318
In the realm of breast cancer screening, X-ray mammography has been the gold-standard imaging technique for decades, enabling early detection and better treatment outcomes. However, mammography has certain limitations, particularly in dense breast tissues, where the sensitivity of the technique decreases. To address this challenge, ultrasound imaging has emerged as a complementary modality, offering improved sensitivity in dense breasts and cost efficiency. Despite advancements in b-mode imaging techniques, ultrasound imaging is a qualitative approach that is highly dependent on the operator's expertise and interpretations. Thus, additional quantitative approaches can improve tissue characterization, e.g. Speed-of-Sound (SoS) has great potential in the early detection of breast cancers. Nevertheless, there is no gold-standard method to reconstruct SoS in pulse-echo setup. The primary focus of this dissertation is to address the reconstruction of SoS maps in pulse-echo medical ultrasound from a plane-wave transmission setup using deep-learning-based approaches. Obtaining labeled real-world data for training deep learning models in ultrasound is challenging due to the lack of a reliable gold-standard method for SoS measurements. To overcome the challenges associated with obtaining real-world labeled data, this dissertation adopts the use of simulated data for training deep neural networks. Simulated data offers a large and diverse dataset and enables the creation of arbitrary heterogeneous structures and acoustic properties that would be costly and time-consuming to acquire in real setups, if even possible. The proposed data generation setup in this thesis utilizes an acoustic toolbox for raw data simulation. Two setups are proposed for generating training datasets: one with simplified homogeneous mediums and inclusions with varying SoS values which is an extension of an existing setup proposed by \cite{feigin2019deep}, and the other utilizing Tomosynthesis images to extract realistic tissue structures for acoustic simulation. These synthetic data generation approaches facilitated more comprehensive training data and enhanced the stability and reproducibility of SoS reconstruction. Our method which is trained only on the simulated RF dataset from a single plane-wave acquisition achieved RMSE of 20.96±0.8 m/s and MAE of 14.03±1.3 m/s on simulated test data. Furthermore, we showed that the networks trained on simulated data can be employed to reconstruct SoS directly from measured RF data. On the measured phantom data, for consecutive data acquisitions with inclusions in FoV, the predicted SoS values exhibited a standard deviation (SD) of 26.70 m/s inside the inclusions and 30.32 m/s in the background. Additionally, this dissertation explores the use of encoder-decoder network architectures for SoS reconstruction, tailored to handle RF, and IQ demodulated raw data. The En-De-Net, RF-Net, and IQ-Net are introduced to accommodate different input data formats and processing requirements. The feasibility and stability of SoS reconstruction using encoder-decoder networks are investigated. We showed that the networks trained on different IQ demodulated and decimated formats perform on par with the networks trained using RF data, as long as their phase information is intact. To address the limitations of end-to-end approaches, a novel representation learning method called the AutoSpeed network is proposed. This technique employs the latent space of linked autoencoders to efficiently map RF data to the SoS domain. On the measured phantom data, for the consecutive data acquisitions with inclusions in the FoV, this method demonstrated superiority in terms of stability and reproducibility with a lower SD compared to the end-to-end approach (14.75 m/s within the inclusions and 23.9 m/s in the background). The proposed deep-learning-based methods, combined with synthetic data and representation learning techniques, hold great promise for advancing SoS reconstruction in pulse-echo medical ultrasound, paving the way for more accurate tissue characterization.
APA:
Khun Jush, F. (2025). Speed-of-Sound Reconstruction from Medical Ultrasound Raw Data using Deep Neural Networks (Dissertation).
MLA:
Khun Jush, Farnaz. Speed-of-Sound Reconstruction from Medical Ultrasound Raw Data using Deep Neural Networks. Dissertation, 2025.
BibTeX: Download