laughingrice/Ultrasound_planewave_sos_inversion

Plane wave raw ultrasound simualted data for deep learning speed of sound inversion

Cite this dataset as:

Feigin M, Freedman D, Anthony BW. Computing Speed-of-Sound from ultrasound: user-agnostic recovery and a new benchmark. IEEE Trans Biomed Eng. 2023; doi:10.1109/TBME.2023.3327147

Simulation

The full dataset consists of 112640 simulations split into 9216 simulations in the training set, 1024 in the validation set, and 1024 in the test set. The measured signal is simulated using the k-wave MATLAB toolbox. Simulations were performed for nine plane waves at $00$, $\pm 8\backslash pm\; 8$, $\pm 16\backslash pm\; 16$, $\pm 24\backslash pm\; 24$, and $\pm 32\backslash pm\; 32$ element offsets, with corresponding wavefront angles of $00$, $\pm 6.7\backslash pm\; 6.7$, $\pm 13.7\backslash pm\; 13.7$, $\pm 20.2\backslash pm\; 20.2$, and $\pm 26.3\backslash pm\; 26.3$ (the time delay is calculated based on 1540 m/s speed of sound so the actual angle will differ per sample), set to pass through the center of the domain. See the figures for details (three of the 9 plane waves are shown to reduce clutter). Each simulation was performed with two center frequencies, 2.5 MHz and 5 MHz, with a Gaussian window (pulse width) of 5 oscillations. An additional simulation at 4.4 MHz is available under the validation directory to allow testing for transfer learning.

Each simulation comprised of $1152\times 11521152\; \backslash times\; 1152$ random speed-of-sound and $\alpha \backslash alpha$ (attenuation) coefficient maps following power law attenuation [\(\mbox{dB} / \mbox{cm} / \mbox{MHz}^2\)] in a domain $42.35\times 42.3542.35\; \backslash times\; 42.35$ mm in size

The domain is constructed by layering a randomly selected set of ellipses and half-planes. For each of the resulting domains (organs), we randomly selected the speed of sound, attenuation coefficient, speckle density, and speckle amplitude. Domains were verified to not slice the probe face; i.e. the resulting maps are verified not to have a discontinuity at the probe face.

The speed of sound range is 1300 m/s to 1800 m/s. The $\alpha \backslash alpha$ coefficient range is $0.050.05$ to $0.150.15$ dB/cm/MHz\({}^2\). Background density is set to 0.9 g/cm\({}^3\) (density of fat).

Speckle noise is randomly generated in the density domain so as not to affect the wavefront propagation speed (uniformly distributed point sources with 2-10 points per wavelength and uniformly distributed amplitude at $\pm 10\mathrm{\%}\backslash pm\; 10\backslash \%$).

Probe

To match our physical hardware, we simulated a 128-element array with 64 active transmit elements. The simulation was carried out with two pulse center frequencies, 2.5 MHz and 5 MHz with a Gaussian window of 5 oscillations.

The central plane wave (zero degrees) is centered at elements 33 to 96. The probe face is placed at $y=60y\; =\; 60$ (outside the perfectly matched layer) and centered on the $xx$ axis. The numerical receive array is 4 elements per sensor element, with a matching kerf (spacing) value, i.e., 4 on 4 off. The signal for each receiver is summed across the 4 receiver elements to generate the 128 receive channels, and the signal is down-sampled to a 40 MHz sampling rate (ADC rate). For the transmit signal, we use a continuous array, as we found that it better matches real-world signals, so for the centered plane wave, a source is placed on all pixels with $y=60y\; =\; 60$ and $322\le x\le 830322\; \backslash le\; x\; \backslash le\; 830$ with a zero time delay on all elements.

File format

The data is in Matlab v7.3 (HDF5) file format create by the python hdf5storage package. Fields in each file:

• alpha_coeff: alpha coeffienct (attenuation) map {1024 - samples, 1channel, 1152 - X dimension, 1152 - Z dimension}
• c0: speed of sound map {1024 - samples, 1, 1152 - X dimension, 1152 - Z dimension}
• cycles: number of cycles in the acoustic wavelet
• f: frequencies simulated
• offsets: plane wave offsets
• p_f_o: simulated ultrasound signal {1024 - samples, 1, 128 - reciever, 2667 - time sample}

the k-wave simulation setup. The US array is placed at line 60 of the numerical grid. Due to kerf, slightly less than half of the array (64 elements) is excited to generate the outgoing plane wave. To better match the actual signal and avoid artifacts, a continuous section is excited. The angle is set based on an assumed 1540 m/s speed of sound so that the plane wave overlaps the center of the domain

Array structure, with 4 active elements and 4 kerf elements interleaved. The recorded signal is the average of the 4 receiving cells for each element

The data simulates plane wave ultrasound data in random media