Applying Machine Learning to Ultrasound Computational Simulations to Determine Scatterer Location

In the field of ultrasound imaging, it has been theorized that imaging noise, known as speckle, may be the product of unresolvable scatterers. This would result in certain speckling patterns revealing themselves over large datasets, which could be utilized to predict the early formation of tumors. Such a dataset would be difficult to analyze by hand, but machine learning algorithms can be used to recognize these patterns in an effective manner. As of now, few attempts have been made apply machine learning to predict scatterer placement from ultrasound scans.
We have constructed a computational simulation which consistently and accurately reproduces experimental data, and can be used to train a machine learning algorithm. Using Field II, a Matlab-based program for ultrasound modelling, simulations were created to replicate data produced from experimental phantoms composed of glass beads and agarose gel. Simulations were designed to account for bead placement and size, as well as experimental conditions. Comparisons between simulated and experimental data using statistical analysis show that computational methods can accurately predict ultrasound images. This verification allows us to begin training a machine learning algorithm using both simulated and experimental data.