Speed-of-Sound Reconstruction from Medical Ultrasound Raw Data using Deep Neural Networks
Mammography has long been the standard for breast cancer screening, but it loses sensitivity in dense breast tissue. Ultrasound offers a cheaper, more sensitive complement for these cases, though conventional ultrasound imaging is qualitative, meaning its results depend heavily on the operator’s skill and interpretation. A quantitative measure called Speed-of-Sound (SoS) could help characterize tissue more objectively and aid early cancer detection, but there is no established gold-standard method for reconstructing SoS in the most common ultrasound setup (pulse-echo, plane-wave transmission). This thesis tackles that reconstruction problem using deep learning.
A major obstacle is that there is no reliable way to obtain labeled real-world SoS data to train models on. To get around this, the thesis relies on simulated data instead, generated with an acoustic simulation toolbox using two approaches: simplified media with inclusions of varying SoS values, and more realistic tissue structures derived from Tomosynthesis images. Networks trained purely on this simulated data, using only a single plane-wave acquisition, achieved strong accuracy on simulated test data, and the same networks proved able to reconstruct SoS directly from real measured data on phantom tests, with consistent, reproducible results across repeated acquisitions.
The thesis also compares different encoder-decoder architectures (En-De-Net, RF-Net, IQ-Net) for handling different types of raw ultrasound data, showing that networks trained on IQ-demodulated data perform just as well as those trained on raw RF data, as long as phase information is preserved. To improve on straightforward end-to-end reconstruction, the thesis introduces AutoSpeed, a representation-learning method that uses linked autoencoders to map raw data into the SoS domain. This approach turned out to be more stable and reproducible than the end-to-end alternative on real phantom measurements.
Together, these deep-learning methods, combined with synthetic training data and representation learning, offer a promising path toward more accurate, reproducible tissue characterization from pulse-echo ultrasound.