Three-dimensional ultrasonic super-resolution imaging method and system based on three-dimensional diffusion model

Abstract

The invention discloses a three-dimensional ultrasonic super-resolution imaging method and system based on a three-dimensional diffusion model, and belongs to the technical field of medical ultrasonic super-resolution imaging. According to the invention, the problems of global information loss and large memory occupation of the existing three-dimensional imaging method are solved. According to the method, the generative diffusion model based on the three-dimensional convolutional neural network is adopted, noise can be gradually added to the diffusion model in the forward diffusion process, the resolution and features of the image are gradually changed, and the trained diffusion model can improve the three-dimensional ultrasonic super-resolution imaging effect. The diffusion model is processed on the basis of sparse convolution, sparse pooling and sparse activation, and redundant operation on zero-value data can be avoided, so that the reasoning speed is increased, the calculation and storage efficiency of sparse data is optimized, and memory occupation is reduced to the maximum extent. According to the method, the image does not need to be subjected to block processing, so that global information is not lost. The method can be applied to medical ultrasonic super-resolution imaging.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical ultrasound super-resolution imaging, and particularly relates to a three-dimensional ultrasound super-resolution imaging method and system based on a three-dimensional diffusion model. BACKGROUND

[0002] Ultrasound localization microscopy (ULM) avoids interference between different microbubbles by sequentially observing them so that isolated sources can be detected in each image. When the point spread function (PSF) on the radio frequency channel data or the beamformed image is known, micron-precision localization of each microbubble can be obtained. Since these contrast agents are purely intravascular, accumulation of these subwavelength localizations will produce a super-resolution image of the microvessels.

[0003] Although satisfactory results have been achieved under ultrasound localization microscopy, deep learning-based methods are commonly used for 2D imaging. In fact, the third spatial dimension greatly increases the size of the intermediate feature maps and the high-resolution output. For example, implementing Deep-stULM directly in 3D images requires at least several orders of magnitude more memory than in 2D images. If 3D ultrasound super-resolution images are obtained indirectly through 2D ultrasound super-resolution reconstruction, the block processing of the images will lose the global information dependency between blocks, thereby affecting the accuracy and reliability of the reconstruction. SUMMARY

[0004] The purpose of the present application is to solve the problems of global information loss and large memory occupation in existing three-dimensional imaging methods, and a three-dimensional ultrasound super-resolution imaging method and system based on a three-dimensional diffusion model are proposed.

[0005] The technical solution adopted by the present application to solve the above technical problems is:

[0006] According to an aspect of the present application, a three-dimensional ultrasound super-resolution imaging method based on a three-dimensional diffusion model, the method specifically comprises the following steps:

[0007] Step S1, obtaining radio frequency data of microvessels through ultrasound imaging, extracting microbubble signals from the radio frequency data of microvessels, obtaining microbubble position information according to the extracted microbubble signals, and obtaining a three-dimensional B-mode image data set of the position of the microbubbles;

[0008] obtaining a three-dimensional ultrasound high-resolution image corresponding to the three-dimensional B-mode image data set;

[0009] Step S2, building a diffusion model based on a three-dimensional U-Net network, the diffusion model including a three-dimensional U-Net network, a feature extraction network, a first noise elimination module and a second noise elimination module;

[0010] The diffusion model is trained using a 3D B-mode image dataset, and the 3D ultrasound high-resolution images corresponding to the 3D B-mode images are used as output labels for the decoding part of the 3D U-Net network in the diffusion model.

[0011] Step S3: performing ultrasound imaging on the tissue to be identified to obtain radiofrequency data of the microvessels to be identified, and processing the radiofrequency data of the microvessels to be identified to obtain a three-dimensional B-mode image to be identified;

[0012] The three-dimensional B-mode image to be identified is used as the input of the encoding part of the trained diffusion model, and the three-dimensional B-mode image to be identified is used as the constraint of the three-dimensional U-Net network in the trained diffusion model to obtain the three-dimensional ultrasound super-resolution imaging results of the tissue to be identified.

[0013] Furthermore, the diffusion model is trained using the three-dimensional B-mode image dataset, and the specific process is as follows:

[0014] Step S21: gradually adding noise to the 3D ultrasound high-resolution image corresponding to the 3D B-mode image using a forward diffusion process until the 3D ultrasound high-resolution image is transformed into a Gaussian noise image;

[0015] Step S22: Using the Gaussian noise image obtained in step S21 as input to the diffusion model. Within the diffusion model, the Gaussian noise image serves as input to the encoding portion of the 3D U-Net network. Each 3D B-mode image serves as input to the feature extraction network of the diffusion model. The extracted features serve as constraints of the 3D U-Net network, i.e., the features of each 3D B-mode image are integrated into the input of the decoding portion of the 3D U-Net network.

[0016] Step S23: Use the noise of each step output by the decoding part of the three-dimensional U-Net network, calculate the loss function value using the noise of each step output by the decoding part and the noise added in each step of the forward diffusion process, reversely adjust the model parameters according to the loss function value, and stop training when the loss function value converges to obtain a trained diffusion model.

[0017] Furthermore, the specific process of step S21 is as follows:

[0018] In the forward diffusion process, the time step Image At time step Image The probability distribution on for:

[0019]

[0020] in, is a noise variance, is an identity matrix, is a normal distribution;

[0021] At a time step , the three-dimensional ultrasound high-resolution image is completely turned into Gaussian noise.

[0022] Further, the three-dimensional B-mode image sequentially passes through four down-sampling layers in the feature extraction network, and the output of the last down-sampling layer is taken as the feature of the three-dimensional B-mode image.

[0023] Further, in the three-dimensional U-Net network, the encoding part includes a first down-sampling unit to a fourth down-sampling unit, and the decoding part includes a first up-sampling unit to a fourth up-sampling unit.

[0024] The input of the decoding part of the three-dimensional U-Net network is the feature of each three-dimensional B-mode image, and the input of the decoding part of the three-dimensional U-Net network is the feature of each three-dimensional B-mode image.

[0025] Step S221, the output of the first down-sampling unit of the encoding part is passed through the first noise elimination module, and the output of the first noise elimination module and the feature of the three-dimensional B-mode image are calculated by the class attention mechanism to obtain the class attention mechanism calculation result a.

[0026] Step S222, the output of the second down-sampling unit of the encoding part is passed through the second noise elimination module, and the output of the second noise elimination module and a are calculated by the class attention mechanism to obtain the class attention mechanism calculation result b.

[0027] Step S223, add b and the output of the fourth down-sampling unit, and take the addition result as the input of the first up-sampling unit of the decoding part.

[0028] Further, the first noise elimination module is FF-Parser.

[0029] Further, in step S221, the specific process of the class attention mechanism calculation is as follows:

[0030] Step 1, layer normalization processing is performed on the feature of the three-dimensional B-mode image and the output of the first noise elimination module, and then the two layer normalization processing results are multiplied to obtain a multiplication result c.

[0031] Step 2, multiply c and the output of the first noise elimination module to obtain a multiplication result a, and a is the result of the class attention mechanism calculation.

[0032] Further, the loss function value is calculated by using each step noise output by the decoding part and noise added in each step of the forward diffusion process, and the loss function is specifically mean square error loss :

[0033]

[0034] wherein, represents noise added in the forward process; represents the output of the decoding part, i.e., noise of the predicted output; represents 2-norm; represents expectation; represents a three-dimensional ultrasonic high-resolution image after denoising by the model.

[0035] Further, the convolution layer in the three-dimensional U-Net network only traverses non-zero elements of the input tensor, i.e., only performs convolution calculation on the non-zero element positions; the ReLU activation function in the three-dimensional U-Net network only traverses non-zero elements of the input tensor, i.e., only applies the activation function to the non-zero element positions; the maximum pooling layer in the three-dimensional U-Net network matches the pooling window with the non-zero elements of the input sparse tensor; and the loss function only traverses non-zero elements of the input tensor, i.e., only calculates the loss at the non-zero element positions.

[0036] Based on another aspect of the present application, a three-dimensional ultrasonic super-resolution imaging system based on a three-dimensional diffusion model is used to perform a three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model.

[0037] The present application has the following advantages:

[0038] The present application adopts a generative diffusion model based on a three-dimensional convolutional neural network, and the diffusion model gradually adds noise in the forward diffusion process, so that the resolution and features of the image gradually change. The trained diffusion model can be used for three-dimensional ultrasonic super-resolution imaging and improve the effect of three-dimensional ultrasonic super-resolution imaging. The diffusion model of the present application is processed based on sparse convolution, sparse pooling and sparse activation, which can avoid redundant operations on zero-value data, so as to speed up the inference, optimize the calculation and storage efficiency of sparse data, and minimize the memory occupation. Moreover, the method of the present application does not need to perform block processing on the image, and can realize three-dimensional ultrasonic super-resolution imaging without losing global information. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of a three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model of the present application;

[0040] Figure 2 is a schematic diagram of the reverse denoising process based on the diffusion model. DETAILED DESCRIPTION

[0041] DETAILED DESCRIPTION Figure 1 The present embodiment is described. The three-dimensional ultrasound super-resolution imaging method based on a three-dimensional diffusion model described in the present embodiment specifically comprises the following steps:

[0042] Step S1, obtain radio frequency data of microvessels by ultrasound imaging, then extract microbubble signals from the radio frequency data of microvessels (eliminate background signals), obtain microbubble position information according to the extracted microbubble signals, and obtain a three-dimensional B-mode image data set of the position of the microbubbles. The three-dimensional B-mode image is an image obtained by a three-dimensional imaging method based on a B-mode (brightness mode) technology extended in medical ultrasound imaging;

[0043] Then obtain a three-dimensional ultrasound high-resolution image corresponding to the three-dimensional B-mode image data set. The three-dimensional ultrasound high-resolution image in the present application refers to an image with a spatial resolution of tens of microns;

[0044] Step S2, build a diffusion model based on a three-dimensional U-Net network, which includes a three-dimensional U-Net network, a feature extraction network, a first noise elimination module, and a second noise elimination module;

[0045] Train the diffusion model using the three-dimensional B-mode image data set, and use the three-dimensional ultrasound high-resolution image corresponding to the three-dimensional B-mode image as the output label of the decoding part of the three-dimensional U-Net network in the diffusion model;

[0046] Step S3, obtain radio frequency data of microvessels to be identified by ultrasound imaging of the tissue to be identified, and process the radio frequency data of the microvessels to be identified to obtain a three-dimensional B-mode image to be identified;

[0047] Use the three-dimensional B-mode image to be identified as the input of the encoding part of the trained diffusion model, and use the three-dimensional B-mode image to be identified as the constraint of the three-dimensional U-Net network in the trained diffusion model, to obtain a three-dimensional ultrasound super-resolution imaging result of the tissue to be identified. Through the three-dimensional ultrasound super-resolution imaging technology, a high-resolution image with a spatial resolution of tens of microns can be obtained.

[0048] DETAILED DESCRIPTION Figure 2 The present embodiment is described. The present embodiment is a further limitation of the first embodiment. The diffusion model is trained using the three-dimensional B-mode image data set, and the specific process is as follows:

[0049] Step S21, gradually add noise to the three-dimensional ultrasound high-resolution image corresponding to the three-dimensional B-mode image by using the forward diffusion process until the three-dimensional ultrasound high-resolution image is converted into a Gaussian noise image;

[0050] Step S22, take the Gaussian noise image obtained in step S21 as the input of the diffusion model, and take each three-dimensional B-mode image as the input of the feature extraction network of the diffusion model, and take the extracted features as the constraint of the three-dimensional U-Net network, that is, the features of each three-dimensional B-mode image are integrated into the input of the decoding part of the three-dimensional U-Net network.

[0051] Step S23, output the noise of each step by using the decoding part of the three-dimensional U-Net network, calculate the loss function value by using the noise output by the decoding part and the noise added by the forward diffusion process, and adjust the model parameters in the reverse direction according to the loss function value until the loss function value converges, and stop training, and obtain the trained diffusion model.

[0052] The other steps and parameters are the same as those in the first embodiment.

[0053] The working principle of the three-dimensional U-Net network of the present application is further described in detail as follows: the working process of the three-dimensional U-Net network is a continuous iteration process. In the first iteration, the Gaussian noise image is taken as the input of the encoding part, and the noise added in the Tth step can be predicted by the decoding part. According to the Gaussian noise image and the noise added in the Tth step, the image after adding noise in the T-1th step can be obtained. In the second iteration, the image after adding noise in the T-1th step is taken as the input of the encoding part, and the noise added in the T-1th step can be obtained by the decoding part. According to the image after adding noise in the T-1th step and the noise added in the T-1th step, the image after adding noise in the T-2th step can be obtained. In this way, the original image without adding any noise can be restored. And in the iteration process, the constraint of the three-dimensional U-Net network is unchanged.

[0054] The encoding part can gradually down-sample the input image or noise to extract features of different scales, and the decoding part can use these features to gradually restore the image in the up-sampling process. The present application fuses the features of the corresponding layers in the encoding part and the decoding part through the skip connection, so that the model can fully utilize the feature information at different scales. In this way, the decoding part can utilize both the low-level features extracted by the encoding part and the high-level features generated in the up-sampling process when generating the image.

[0055] Specific embodiment three: the present embodiment is a further limitation of the second embodiment, and the specific process of step S21 is as follows:

[0056] In the forward diffusion process, the time step The image At time step The image The probability distribution on the image Is:

[0057]

[0058] Where, is the noise variance, is the identity matrix, is a normal distribution;

[0059] At time step , the three-dimensional ultrasound high-resolution image is completely changed into Gaussian noise.

[0060] Other steps and parameters are the same as embodiment two.

[0061] In the reverse denoising process of the traditional method, the data is gradually recovered from the noise through a series of time steps:

[0062]

[0063] Where, and are the mean and variance predicted by the neural network respectively. At time step 0, the final data is generated.

[0064] The present application uses a three-dimensional U-Net network model to simulate the reverse denoising process to predict the noise or data distribution.

[0065] Embodiment four: this embodiment is a further limitation of embodiment three, wherein the three-dimensional B-mode image sequentially passes through four down-sampling layers in the feature extraction network, and the output of the last down-sampling layer is taken as the feature of the three-dimensional B-mode image.

[0066] Other steps and parameters are the same as embodiment three.

[0067] Embodiment five: in combination with Figure 2 to illustrate this embodiment. This embodiment is a further limitation of embodiment three, wherein the encoding part of the three-dimensional U-Net network includes the first to fourth down-sampling units, and the decoding part includes the first to fourth up-sampling units.

[0068] The input of the decoding part of the three-dimensional U-Net network into which the features of each three-dimensional B-mode image are fused is specifically:

[0069] Step S221, the output of the first down-sampling unit of the encoding part is passed through a first noise cancellation module, the output of the first noise cancellation module and the features of the three-dimensional B-mode image are calculated by the similar attention mechanism to obtain the calculation result a of the similar attention mechanism;

[0070] Step S222, the output of the second down-sampling unit of the encoding part is passed through a second noise cancellation module, the output of the second noise cancellation module and a are calculated by the similar attention mechanism to obtain the calculation result b of the similar attention mechanism;

[0071] Step S223, b and the output of the fourth down-sampling unit are added, and the addition result is taken as the input of the first up-sampling unit of the decoding part.

[0072] The other steps and parameters are the same as those in the third embodiment.

[0073] The sixth embodiment is a further limitation of the fifth embodiment, and the first noise cancellation module is an FF-Parser (Feature Frequency Parser).

[0074] The other steps and parameters are the same as those in the fifth embodiment.

[0075] The second noise cancellation module is the same as the first noise cancellation module. The processing process of the noise cancellation module is introduced as follows: the input of the noise cancellation module is denoted as m, 3D FFT (3D Fast Fourier Transform) is performed on the input m, then the 3D FFT result is multiplied by a complex tensor A, wherein A is a tensor that can be automatically learned and changed by a neural network, and finally 3D IFFT (3D Fast Fourier Transform) is performed on the multiplication result, and the 3D IFFT result is taken as the output of the noise cancellation module.

[0076] The seventh embodiment is a further limitation of the sixth embodiment, and in step S221, the specific process of the similar attention mechanism calculation is as follows:

[0077] Step 1, layer normalization is performed on the features of the three-dimensional B-mode image and the output of the first noise cancellation module, then the two layer normalization results are multiplied to obtain a multiplication result c;

[0078] Step 2, c is multiplied by the output of the first noise cancellation module to obtain a multiplication result a, and a is the result of the similar attention mechanism calculation.

[0079] The other steps and parameters are the same as those in the sixth embodiment.

[0080] Specific implementation eight: this implementation is a further limitation of specific implementation seven, the loss function value is calculated by using the noise of each step output by the decoding part and the noise added in each step of the forward diffusion process, and the loss function is mean square error (MSE) loss That is, the average of the noise deviation predicted by each step is:

[0081]

[0082] Wherein, represents the noise added in the forward process; represents the output of the decoding part, that is, the predicted noise output; represents the 2 norm; represents the expectation; represents the three-dimensional ultrasonic high-resolution image after denoising by the model.

[0083] The other steps and parameters are the same as those in specific implementation seven.

[0084] The present application takes Gaussian noise as the input of the three-dimensional U-net model, and under the constraint of the original image, the model will predict the noise added in the first step, and the image after denoising in the first step is obtained according to the noise added in the first step , and the image is taken as the input of the three-dimensional U-net model, the model will predict the noise added in the second step, and the image after denoising in the second step is obtained according to the noise added in the second step , and the image is taken as the input of the three-dimensional U-net model, the model will predict the noise added in the second step, and the image after denoising in the second step is obtained according to the noise added in the second step , and so on, so as to obtain the three-dimensional ultrasonic super-resolution imaging result. Wherein,

[0085] , .

[0086]

[0087] Wherein, , .

[0088] ​​​​Specific embodiment nine: this embodiment is a further limitation of specific embodiment one, the convolution layer in the three-dimensional U-Net network only traverses the non-zero elements of the input tensor, that is, only the non-zero element positions are convolved; the ReLU activation function in the three-dimensional U-Net network only traverses the non-zero elements of the input tensor, that is, only the non-zero element positions are applied to the activation function; the maximum pooling layer in the three-dimensional U-Net network matches the pooling window with the non-zero elements of the input sparse tensor; the loss function only traverses the non-zero elements of the input tensor, that is, only the non-zero element positions are calculated for the loss.

[0089] Other steps and parameters are the same as those in specific embodiment one.

[0090] The present application can perform efficient calculation and storage optimization through sparse convolution, sparse pooling and sparse activation, so as to minimize the memory occupation, thereby meeting the memory requirements of the three-dimensional diffusion model under the condition of hardware limitation.

[0091] Specific embodiment ten, the three-dimensional ultrasonic super-resolution imaging system based on a three-dimensional diffusion model, the system is used for executing the three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model in one of specific embodiments one to nine.

[0092] The above examples of the present application are only to illustrate the calculation model and calculation process of the present application, and are not a limitation on the embodiments of the present application. For those skilled in the art, other different forms of changes or variations can be made on the basis of the above description, and it is impossible to enumerate all the embodiments here, and any obvious changes or variations derived from the technical solutions of the present application still fall within the protection scope of the present application.

Claims

1. A three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model, characterized in that: The method specifically comprises the following steps: Step S1, obtaining radiofrequency data of microvessels through ultrasound imaging, extracting microbubble signals from the radiofrequency data of microvessels, obtaining microbubble location information based on the extracted microbubble signals, and obtaining a three-dimensional B-mode image dataset of the microbubble locations; Then obtain a three-dimensional ultrasound high-resolution image corresponding to the three-dimensional B-mode image data set; Step S2: building a diffusion model based on a three-dimensional U-Net network, wherein the diffusion model includes a three-dimensional U-Net network, a feature extraction network, a first noise elimination module, and a second noise elimination module; The diffusion model is trained using a 3D B-mode image dataset, and the 3D ultrasound high-resolution images corresponding to the 3D B-mode images are used as output labels for the decoding part of the 3D U-Net network in the diffusion model. Step S3: performing ultrasound imaging on the tissue to be identified to obtain radiofrequency data of the microvessels to be identified, and processing the radiofrequency data of the microvessels to be identified to obtain a three-dimensional B-mode image to be identified; The three-dimensional B-mode image to be identified is used as the input of the encoding part of the trained diffusion model, and the three-dimensional B-mode image to be identified is used as the constraint of the three-dimensional U-Net network in the trained diffusion model to obtain the three-dimensional ultrasound super-resolution imaging results of the tissue to be identified.

2. The three-dimensional ultrasound super-resolution imaging method based on a three-dimensional diffusion model according to claim 1, characterized in that: The diffusion model is trained using a three-dimensional B-mode image dataset. The specific process is as follows: Step S21: gradually adding noise to the 3D ultrasound high-resolution image corresponding to the 3D B-mode image using a forward diffusion process until the 3D ultrasound high-resolution image is transformed into a Gaussian noise image; Step S22: Using the Gaussian noise image obtained in step S21 as input to the diffusion model. Within the diffusion model, the Gaussian noise image serves as input to the encoding portion of the 3D U-Net network. Each 3D B-mode image serves as input to the feature extraction network of the diffusion model. The extracted features serve as constraints of the 3D U-Net network, i.e., the features of each 3D B-mode image are integrated into the input of the decoding portion of the 3D U-Net network. Step S23: Use the noise of each step output by the decoding part of the three-dimensional U-Net network, calculate the loss function value using the noise of each step output by the decoding part and the noise added in each step of the forward diffusion process, reversely adjust the model parameters according to the loss function value, and stop training when the loss function value converges to obtain a trained diffusion model.

3. The three-dimensional ultrasound super-resolution imaging method based on a three-dimensional diffusion model according to claim 2, characterized in that: The specific process of step S21 is as follows: In the forward diffusion process, the time step Image At time step Image The probability distribution on for: in, is the noise variance, is the identity matrix, is a normal distribution; In the time step At this point, the high-resolution 3D ultrasound image completely turns into Gaussian noise.

4. The three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model according to claim 3, characterized in that: The three-dimensional B-mode image passes through four downsampling layers in sequence in the feature extraction network, and the output of the last downsampling layer is used as the feature of the three-dimensional B-mode image.

5. The three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model according to claim 3, characterized in that: In the three-dimensional U-Net network, the encoding part includes the first downsampling unit to the fourth downsampling unit, and the decoding part includes the first upsampling unit to the fourth upsampling unit; The features of each 3D B-mode image are integrated into the input of the decoding part of the 3D U-Net network, specifically: Step S221: Pass the output of the first downsampling unit of the encoding part through the first noise elimination module, perform a similar attention mechanism calculation on the output of the first noise elimination module and the features of the three-dimensional B-mode image, and obtain a similar attention mechanism calculation result a; Step S222: Pass the output of the second downsampling unit of the encoded part through the second noise elimination module, perform a class attention mechanism calculation on the output of the second noise elimination module and a, and obtain a class attention mechanism calculation result b; Step S223: Add b and the output of the fourth down-sampling unit, and use the added result as the input of the first up-sampling unit of the decoding part.

6. The three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model according to claim 5, characterized in that: The first noise elimination module is FF-Parser.

7. The three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model according to claim 6, characterized in that: In step S221, the specific process of the attention-like mechanism calculation is as follows: Step 1: perform layer normalization processing on the features of the 3D B-mode image, perform layer normalization processing on the output of the first noise removal module, and then multiply the two layer normalization processing results to obtain a multiplication result c; Step 2: Multiply c by the output of the first noise elimination module to obtain the multiplication result a, which is the result calculated by the class attention mechanism.

8. The three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model according to claim 7, characterized in that: The noise at each step output by the decoding part and the noise added at each step in the forward diffusion process are used to calculate the loss function value. The loss function is specifically the mean square error loss : in, represents the noise added by the forward process; Represents the output of the decoding part, that is, the noise of the predicted output; represents the 2-norm; Expressing hope; Represents a high-resolution three-dimensional ultrasound image after model denoising.

9. The three-dimensional ultrasonic super-resolution imaging method based on a three-dimensional diffusion model according to claim 1, characterized in that: The convolution layer in the three-dimensional U-Net network only traverses the non-zero elements of the input tensor, that is, convolution calculation is only performed on the non-zero element positions; the ReLU activation function in the three-dimensional U-Net network only traverses the non-zero elements of the input tensor, that is, the activation function is only applied to the non-zero element positions; the maximum pooling layer in the three-dimensional U-Net network matches the pooling window with the non-zero elements of the input sparse tensor; the loss function only traverses the non-zero elements of the input tensor, that is, the loss is only calculated for the non-zero element positions.

10. A three-dimensional ultrasonic super-resolution imaging system based on a three-dimensional diffusion model, characterized in that: The system is used to execute the three-dimensional ultrasound super-resolution imaging method based on a three-dimensional diffusion model as described in any one of claims 1 to 9.

Images