Granted Patent B2

US 12,579,441 · App. 18/053,348 · Granted Mar 17, 2026

Inventors: Yang Zhang (Shanghai, CN); Yang Lyu (Shanghai, CN); Chen Xi (Shanghai, CN)

Assignee: SHANGHAI UNITED IMAGING INTELLIGENCE CO., LTD.

G06N3/084G06N3/045G06T11/008G06T2207/10104G06T2207/20084

Abstract

Claims (84)

1 . A system for image reconstruction, comprising:

at least one storage device including a set of instructions; and

at least one processor configured to communicate with the at least one storage device, wherein when executing the set of instructions, the at least one processor is configured to direct the system to perform operations including:

obtaining an initial image of a subject, the initial image being generated based on scan data of the subject that is collected by an imaging device;

generating a gradient image associated with the initial image;

extracting at least one first 2-dimensional (2D) image from the initial image;

extracting at least one second 2D image from the gradient image;

generating one or more concatenated images by concatenating the at least one first 2D image and the at least one second 2D image, wherein the at least one first 2D image and the at least one second 2D image include a first dimension and a second dimension, the at least one first 2D image and the at least one second 2D image are concatenated along a third dimension different from the first dimension and the second dimension to generate the one or more concatenated images including the first dimension, the second dimension, and the third dimension; and

generating, based on the one or more concatenated images, a target image of the subject by applying an image reconstruction model, the target image having a higher image quality than the initial image.

2 . The system of claim 1 , wherein the generating, based on the one or more concatenated images, a target image of the subject based on the initial image and the gradient image by applying an image reconstruction model includes:

generating the target image by inputting the one or more concatenated images into the image reconstruction model.

3 . The system of claim 1 , wherein:

the at least one first 2D image includes at least one first axial image, at least one first sagittal image, and at least one first coronary image extracted from the initial image; and

the at least one second 2D image includes at least one second axial image, at least one second sagittal image, and at least one second coronary image extracted from the gradient image.

4 . The system of claim 3 , wherein the generating one or more concatenated images by concatenating the at least one first 2D image and the at least one second 2D image includes:

generating a first concatenated image by concatenating the at least one first axial image and the at least one second axial image;

generating a second concatenated image by concatenating the at least one first sagittal image and the at least one second sagittal image; and

generating a third concatenated image by concatenating the at least one first coronary image and the at least one second coronary image.

5 . The system of claim 4 , wherein the image reconstruction model includes:

an axial view component configured to generate a first feature map by processing the first concatenated image;

a sagittal view component configured to generate a second feature map by processing the second concatenated image;

a coronary view component configured to generate a third feature map by processing the third concatenated image; and

an integration component configured to generate an output image by processing the first feature map, the second feature map, and the third feature map, wherein the target image is generated based on the output image of the integration component.

6 . The system of claim 1 , wherein the image reconstruction model is a trained cascaded neural network including a plurality of trained models that are sequentially connected, the plurality of trained models include a trained first model and one or more trained second models downstream to the trained first model, and the generating, based on the one or more concatenated images, a target image by applying an image reconstruction model includes:

obtaining an output image of the trained first model by inputting the one or more concatenated images into the trained first model;

for each of the one or more trained second model,

extracting at least one third 2D image from an output image of a previous trained model connected to the trained second model; and

obtaining an output image of the trained second model based on the at least one first 2D image, the at least one second 2D image, the at least one third 2D image, and the trained second model, wherein the target image is generated based on an output image of the last trained second model of the trained cascaded neural network.

7 . The system of claim 1 , wherein the scan data of the initial image corresponds to a first radiation dose associated with the subject, and the target image corresponds to a second radiation dose higher than the first radiation dose.

8 . The system of claim 1 , wherein the image reconstruction model corresponds to a target image resolution, the initial image has an image resolution different from the target image resolution, and the at least one processor is further configured to direct the system to perform the operations including:

generating a resampled initial image having the target image resolution by resampling the initial image;

generating a preprocessed initial image by normalizing the resampled initial image; and

generating a preprocessed gradient image by normalizing the gradient image, and wherein the at least one first 2D image is extracted from the preprocessed initial image, and the at least one second 2D image is extracted from the preprocessed gradient image.

9 . The system of claim 1 , wherein the image reconstruction model is further configured to reduce noise in the initial image.

10 . The system of claim 1 , wherein the image reconstruction model includes a neural network model.

11 . The system of claim 1 , wherein the image reconstruction model is trained according to a training process including:

obtaining a plurality of training samples, each of the plurality of training samples includes a sample initial image of a sample subject, a sample gradient image associated with the sample initial image, and a sample target image of the sample subject, wherein the sample target image has a higher image quality than the sample initial image;

obtaining a preliminary model; and

generating the image reconstruction model by training the preliminary model using the plurality of training samples.

12 . The system of claim 11 , wherein the generating the image reconstruction model by the training the preliminary model using the plurality of training samples includes:

for each of the plurality of training samples,

extracting at least one sample first 2-dimensional (2D) image from the sample initial image of the training sample; and

extracting at least one sample second 2D image from the sample gradient image of the training sample; and

generating the image reconstruction model by training the preliminary model using the at least one sample first 2D image, the at least one sample second 2D image, and the sample target image of each of the plurality of training samples.

13 . The system of claim 12 , wherein for each of plurality of the training samples,

the at least one sample first 2D image includes at least one sample first axial image, at least one sample first sagittal image, and at least one sample first coronary image extracted from the sample initial image of the training sample; and

the at least one sample second 2D image includes at least one sample second axial image, at least one sample second sagittal image, and at least one sample second coronary image extracted from the sample gradient image of the training sample.

14 . The system of claim 13 , wherein the training the preliminary model includes:

for each of the plurality of training samples,

generating a sample first concatenated image by concatenating the at least one sample first axial image and the at least one sample second axial image of the training sample;

generating a sample second concatenated image by concatenating the at least one sample first sagittal image and the at least one sample second sagittal image of the training sample; and

generating a sample third concatenated image by concatenating the at least one sample first coronary image and the at least one sample second coronary image of the training sample; and

training the preliminary model using the sample first concatenated image, the sample second concatenated image, the sample third concatenated image, and the sample target image of each of the plurality of training samples.

15 . The system of claim 14 , wherein the preliminary model includes:

a first component configured to generate a sample first feature map by processing the sample first concatenated image of each of the plurality of training samples;

a second component configured to generate a sample second feature map by processing the sample second concatenated image of each of the plurality of training samples;

a third component configured to generate a sample third feature map by processing the sample third concatenated image of each of the plurality of training samples; and

a fourth component configured to process the sample first feature map, the sample second feature map, and the sample third feature map of each of the plurality of training samples.

16 . The system of claim 12 , wherein the preliminary model is a cascaded neural network including a plurality of models that are sequentially trained, the plurality of models include a first model and one or more second models downstream to the first model, and the training the preliminary model includes:

for each of the plurality of training samples, training the first model using the at least one sample first 2D image, the at least one sample second 2D image, and the sample target image of the training sample; and

for each of the one or more second models,

for each of the plurality of training samples,

obtaining a sample output image based on the at least one sample first 2D image, the at least one sample second 2D image of the training sample, and the one or more trained models that are generated before training the second model;

extracting at least one sample third 2D image from the sample output image corresponding to the training sample; and

training the second model using the at least one sample first 2D image, the at least one sample second 2D image, the at least one sample third 2D image, and the sample target image of the training sample.

17 . The system of claim 11 , wherein for each of the plurality of training samples,

the corresponding sample initial image is generated based on first scan data that is collected by a sample imaging device and corresponds to a first radiation dose associated with the sample subject,

the corresponding sample target image is generated based on second scan data that is collected by the sample imaging device and corresponds to a second radiation dose associated with the sample subject, and

the second radiation dose is higher than the first radiation dose.

18 . The system of claim 1 , wherein each of the initial image and the gradient image is a three-dimensional (3D) image.

19 . A method for image reconstruction implemented on a computing device having at least one processor and at least one storage device, the method comprising:

obtaining an initial image of a subject, the initial image being generated based on scan data of the subject that is collected by an imaging device;

generating a gradient image associated with the initial image;

extracting at least one first 2-dimensional (2D) image from the initial image;

extracting at least one second 2D image from the gradient image;

generating one or more concatenated images by concatenating the at least one first 2D image and the at least one second 2D image, wherein the at least one first 2D image and the at least one second 2D image include a first dimension and a second dimension, the at least one first 2D image and the at least one second 2D image are concatenated along a third dimension different from the first dimension and the second dimension to generate the one or more concatenated images including the first dimension, the second dimension, and the third dimension; and

generating, based on the one or more concatenated images, a target image of the subject by applying an image reconstruction model, the target image having a higher image quality than the initial image.

20 . A non-transitory computer-readable storage medium including instructions for image reconstruction that, when accessed by at least one processor of a system, causes the system to perform a method, the method comprising:

obtaining an initial image of a subject, the initial image being generated based on scan data of the subject that is collected by an imaging device;

generating a gradient image associated with the initial image;

extracting at least one first 2-dimensional (2D) image from the initial image;

extracting at least one second 2D image from the gradient image;

generating one or more concatenated images by concatenating the at least one first 2D image and the at least one second 2D image, wherein the at least one first 2D image and the at least one second 2D image include a first dimension and a second dimension, the at least one first 2D image and the at least one second 2D image are concatenated along a third dimension different from the first dimension and the second dimension to generate the one or more concatenated images including the first dimension, the second dimension, and the third dimension; and

generating, based on the one or more concatenated images, a target image of the subject by applying an image reconstruction model, the target image having a higher image quality than the initial image.

Assignments (3)

Priority Claims (1)

Continuity (2)

References Cited (34)