Array geometry dependency in sound speed reconstruction using Deep Neural Network

Deep Neural Networks (DNN) are expected to provide better solutions for the accuracy and processing time of sound speed reconstruction in Ultrasound CT (USCT). In this study, we investigated the effect of array geometry on the image quality of reconstruction by DNN in comparison with the conventional method: Full Waveform Inversion (FWI). Pressure signals for a single frequency were computed by the finite difference method for three array geometries with different element pitch sizes with a ring array transducer. The DNN was trained end-to-end on the task of reconstructing the sound speed distribution from the observed signal. For test data, we utilized numerical breast models which containing a tumor model with different diameters and sound speeds. For training data, we utilized artificial sound speed distribution based on a natural images dataset that is different from the test data and is rich in diverse patterns for appropriate evaluation of the generalization performance. The results for the test data showed that when the element pitch size is around half of a wavelength, DNN can improve the image quality by further reducing the pitch, while FWI cannot. In addition, when the element pitch is larger than one wavelength, DNN can reduce artifacts compared to FWI. When the element pitch is smaller than half a wavelength, DNN has the advantage of better reconstruction of tumor shape and smaller diameter error than FWI for large diameter tumors with large sound speed difference. On the other hand, for small diameter tumors with small sound speed difference, FWI has the advantage of capturing the structure that cannot be captured by DNN. From the above, DNN may be able to improve image quality by designing a probe with a finer pitch than that of conventional probes, while FWI is not able to. In addition, using a probe with a coarser pitch than that of a conventional one may reduce the number of AD converters in the probe and significantly lower the cost, since DNN does not produce the serious artifacts that arise in FWI.