Synthetic augmentation based autoencoding neural network for speckle removal in intracardiac ultrasound imaging

1 . A system comprising:

one or more non-transitory data storage media; and

one or more first computer hardware processors in communication with the one or more non-transitory data storage media, wherein the one or more first computer hardware processors are configured to execute first computer-executable instructions to at least:

receive a set of tissue images;

determine, from the set of tissue images, a set of synthetic speckled tissue images, wherein to determine the set of synthetic speckled tissue images comprises:

create noise data;

modify the noise data with at least a filter that results in speckle filter data;

apply, to a first tissue image from the set of tissue images, the speckle filter data as an overlay that results in a first intermediate synthetic speckled tissue image; and

transform the first intermediate synthetic speckled tissue image into a first synthetic speckled tissue image in a sector format, wherein the set of synthetic speckled tissue images comprises the first synthetic speckled tissue image;

determine, from the set of tissue images, a set of images as targets;

create a training set comprising (i) the set of synthetic speckled tissue images and (ii) the set of images as targets; and

train a convolutional autoencoder with the training set.

2 . The system of claim 1 , further comprising:

one or more second computer hardware processors configured to execute second computer-executable instructions to at least:

receive a plurality of ultrasound frames during a cardiology procedure; and

for each ultrasound frame from the plurality of ultrasound frames,

generate, via the convolutional autoencoder, a denoised ultrasound frame, and

cause presentation, via a graphical user interface, of the denoised ultrasound frame.

3 . The system of claim 1 , wherein the set of tissue images comprises non-cardiac images.

4 . The system of claim 1 , wherein to transform the first intermediate synthetic speckled tissue image into the first synthetic speckled tissue image in the sector format comprises:

edit the first intermediate synthetic speckled tissue image to have a shape comprising a first straight edge, a second straight edge, and a curved edge.

5 . The system of claim 4 , wherein to edit the first intermediate synthetic speckled tissue image to have the shape comprises:

modify pixels in the first intermediate synthetic speckled tissue image without removing any features from the first intermediate synthetic speckled tissue image.

6 . The system of claim 4 , wherein to edit the first intermediate synthetic speckled tissue image to have the shape comprises:

crop at least a portion of the first intermediate synthetic speckled tissue image.

7 . A computer-implemented method comprising:

receiving a set of initial training images;

determining, from the set of initial training images, a set of synthetic speckled images, wherein determining the set of synthetic speckled images comprises:

creating noise data;

modifying the noise data with at least a filter that results in speckle filter data; and

applying, to a first training image from the set of initial training images, the speckle filter data as an overlay that results in a first synthetic speckled image, wherein the set of synthetic speckled images is determined from the first synthetic speckled image;

determining, from the set of initial training images, a set of images as targets;

creating a training set comprising (i) the set of synthetic speckled images and (ii) the set of images as targets; and

training a machine learning model with the training set.

8 . The computer-implemented method of claim 7 , training the machine learning model comprises:

applying a region-of-interest mask to the training set and implementing a transformer-type attention mechanism that computes attention scores between feature locations, wherein attention in the transformer-type attention mechanism is gated by the region-of-interest mask such a first attention weight outside a masked region is attenuated during training while a second attention weight within the masked region is emphasized, and wherein the transformer-type attention mechanism comprises multi-head self-attention or cross-attention between encoder and decoder features.

9 . The computer-implemented method of claim 7 , further comprising:

receiving a speckle size parameter, wherein modifying the noise data further comprises:

applying a sector geometric transformation and selecting a kernel size using the speckle size parameter that results in the speckle filter data.

10 . The computer-implemented method of claim 9 , wherein the speckle size parameter comprises at least one of a beam spread or a speckle size growth parameter as a function of depth.

11 . The computer-implemented method of claim 7 , wherein modifying the noise data further comprises:

applying a weighted summation to the noise data and filtered data that results in the speckle filter data.

12 . The computer-implemented method of claim 7 , wherein creating the noise data further comprises:

determining the noise data from a Rayleigh distribution.

13 . The computer-implemented method of claim 7 , wherein the filter comprises a median filter or a mean filter.

14 . A system comprising:

one or more non-transitory data storage media; and

one or more first computer hardware processors in communication with the one or more non-transitory data storage media, wherein the one or more first computer hardware processors are configured to execute first computer-executable instructions to at least:

receive a set of initial training images;

determine, from the set of initial training images, a set of synthetic speckled images, wherein to determine the set of synthetic speckled images comprises:

create noise data;

modify the noise data with at least a filter that results in speckle filter data; and

apply, to a first training image from the set of initial training images, the speckle filter data as an overlay that results in a first synthetic speckled image, wherein the set of synthetic speckled images is determined from the first synthetic speckled image;

determine, from the set of initial training images, a set of images as targets;

create a training set comprising (i) the set of synthetic speckled images and (ii) the set of images as targets; and

train a machine learning model with the training set.

15 . The system of claim 14 , further comprising:

one or more second computer hardware processors configured to execute second computer-executable instructions to at least:

receive an ultrasound image;

generate, via the machine learning model, a denoised ultrasound image; and

cause presentation, via a graphical user interface, of the denoised ultrasound image.

16 . The system of claim 14 , wherein to determine the set of synthetic speckled images comprises:

transform the first synthetic speckled image into a sectorized synthetic speckled image, wherein the set of synthetic speckled images comprises the sectorized synthetic speckled image.

17 . The system of claim 14 , wherein the machine learning model corresponds to at least one of a convolutional autoencoder or a U-Net model.

18 . The system of claim 14 , further comprising:

one or more second computer hardware processors configured to execute second computer-executable instructions to at least:

receive a plurality of ultrasound frames during a cardiology procedure; and

for each ultrasound frame from the plurality of ultrasound frames,

generate, via the machine learning model, a denoised ultrasound frame, and

cause presentation, via a graphical user interface, of the denoised ultrasound frame.

19 . The system of claim 14 , wherein to train the machine learning model comprises:

apply a region-of-interest mask to the training set and implement a transformer-type attention mechanism that computes attention scores between feature locations, wherein attention in the transformer-type attention mechanism is gated by the region-of-interest mask such a first attention weight outside a masked region is attenuated during training while a second attention weight within the masked region is emphasized, and wherein the transformer-type attention mechanism comprises multi-head self-attention or cross-attention between encoder and decoder features.

20 . The system of claim 14 , wherein to modify the noise data further comprises:

apply a weighted summation to the noise data and filtered data that results in the speckle filter data.