Ultrasonic prostate imaging method, imaging segmentation method and system based on sparse angle amplification

Abstract

The invention discloses an ultrasonic prostate imaging method based on sparse angle amplification, and the method comprises the steps: collecting sparse angle RF ultrasonic signal data, carrying out the preprocessing, obtaining a preliminary signal aggregation result, and enabling a third signal aggregation layer of a signal preprocessing network to comprise a row aggregation transformation layer and a column aggregation transformation layer; the initial signal aggregation result is input into an imaging sub-network, an imaging result is obtained, and the imaging sub-network is a deep neural network model of a coding and decoding structure and is provided with a multi-scale signal synthesis module, an imaging sub-network encoder, a middle layer, an imaging sub-network decoder and an output layer. The multi-scale signal synthesis module is used for aggregating a plurality of signals in different scale ranges in the aspect of image arbitrary point numerical value acquisition by utilizing multi-scale cavity convolution and corresponding arbitrary point maximum value operation based on the preliminary signal aggregation result, and taking the maximum value as a final numerical value; and the imaging sub-network encoder, the middle layer, the imaging sub-network decoder and the output layer output an imaging result based on the final numerical value.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical ultrasound imaging, and particularly relates to an ultrasound prostate imaging method, an imaging segmentation method, a system, a device and a storage medium based on sparse angle amplification. BACKGROUND

[0002] Early diagnosis and treatment of prostatitis is of great significance to human reproductive health. Compared with CT, MRI and puncture diagnosis methods, ultrasound imaging has the advantages of non-invasiveness, non-radiation, real-time and convenience, and is expected to realize diagnosis and treatment integration under the guidance of real-time ultrasound imaging. The existing ultrasound imaging (gray-scale ultrasound) outputs images of the overall macrostructure of the prostate, and the technology is mainly used for real-time guidance of prostate cancer puncture biopsy pathological diagnosis. However, the internal structure of prostatitis changes slightly, so the existing ultrasound imaging technology cannot represent the difference in internal tissue composition change, and it is difficult to diagnose prostatitis. Traditional ultrasound imaging is mainly based on focused scanning and multi-angle complex plane wave methods, which have long imaging time and low efficiency. Based on single-angle focused imaging or single-angle plane wave imaging, the imaging resolution is low and the contrast is low.

[0003] In addition, based on the existing ultrasound imaging technology, the physician needs to manually calibrate the prostatitis lesion area in the imaging result according to experience, and then further diagnose through CT, MRI, puncture and the like. The diagnosis cycle is long, and the CT and MRI imaging devices cannot realize real-time imaging guided puncture and treatment. SUMMARY

[0004] To solve the above problems, the purpose of the present application is to provide an ultrasound prostate imaging method, an imaging segmentation method and a system based on sparse angle amplification, which acts on the prostatitis lesion area, based on the received RF ultrasound echo signal, through the sparse angle amplification ultrasound prostatitis imaging network and segmentation network constructed by the present application, realizes high-resolution high-contrast imaging results, and outputs the prostatitis lesion area, which is suitable for early diagnosis of prostate diseases.

[0005] In the first aspect of the present application, the technical scheme provided is: an ultrasound prostate imaging method based on sparse angle amplification, characterized by comprising the following steps:

[0006] Receiving transducer to collect sparse angle RF ultrasound signal data;

[0007] Pretreating the collected sparse angle RF ultrasound signal data through a signal pretreatment network to obtain a preliminary signal aggregation result, wherein the signal pretreatment network has three signal aggregation layers, and the third signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer, so that the aggregated signal is aligned in dimension with the dimension of the final output image;

[0008] The pre-processed preliminary signal aggregation result is input into an imaging sub-network, and an imaging result is obtained, wherein the imaging sub-network is a codec structured deep neural network model, having a multi-scale signal synthesis module, an imaging sub-network encoder, an intermediate layer, an imaging sub-network decoder and an output layer, the multi-scale signal synthesis module uses multi-scale hole convolution and corresponding arbitrary point maximum value operation based on the preliminary signal aggregation result to aggregate signals of multiple different scale ranges in terms of image arbitrary point value acquisition, and takes the maximum value as the final value, and the imaging sub-network encoder, the intermediate layer, the imaging sub-network decoder and the output layer output the imaging result based on the final value.

[0009] Further, the first layer signal aggregation layer in the signal preprocessing network has a row aggregation transformation layer, a column interpolation transformation layer and a column aggregation transformation layer.

[0010] The row aggregation transformation layer of the first layer signal aggregation layer performs signal aggregation on the data collected by each channel in the Z direction, and two-stage one-dimensional convolution operation is used in signal aggregation, the first stage convolution operation uses a convolution kernel size of 79, a convolution kernel step of 3, and 0 signal padding numbers of 79 / / 2=39 at both ends, and each channel is provided with an independent convolution kernel, and there are a total of 128 convolution kernels with a size of 79, and the convolution operation of the ith channel is represented as follows:

[0011]

[0012] The second stage convolution operation uses a convolution kernel size of 56, a convolution kernel step of 1, and 0 signal padding numbers at both ends, and each channel is provided with an independent convolution kernel, and there are a total of 128 convolution kernels with a size of 56, and the convolution operation of the ith channel is represented as follows:

[0013]

[0014] The column interpolation transformation layer of the first layer signal aggregation layer performs signal up-sampling operation in the X direction in advance, and the up-sampling method uses linear interpolation method, and the operation function is defined as: making the signal dimension double in the X direction, and approaching the X direction dimension;

[0015] The column aggregation transformation layer of the first layer signal aggregation layer uses one-dimensional convolution operation in the X direction in signal aggregation, and the convolution kernel size is 128, the convolution kernel step is 1, and the 0 signal padding number at both ends is 66, and each column is provided with an independent convolution kernel, and there are a total of 713 convolution kernels with a size of 128, and the convolution operation of the jth row signal is represented as follows:

[0016]

[0017] After signal aggregation, the signal dimension is SR

[261]

[713] .

[0018] Further, the second layer signal aggregation layer in the signal preprocessing network has a row aggregation transformation layer, a column interpolation transformation layer, and a column aggregation transformation layer:

[0019] The row aggregation transformation layer of the second layer signal aggregation layer aggregates signals in the Z direction for data collected by each channel, and uses two-stage one-dimensional convolution operation for signal aggregation. The first stage convolution kernel size is 14, the convolution kernel step is 1, and the number of 0 signals filled at both ends is 0. An independent convolution kernel is set for each channel, and there are a total of 261 convolution kernels with a size of 14. The i-th column convolution operation is represented as follows:

[0020]

[0021] The second stage convolution operation uses a convolution kernel size of 7, a convolution kernel step of 1, and a number of 0 signals filled at both ends of 0. An independent convolution kernel is set for each channel, and there are a total of 700 convolution kernels with a size of 7. The i-th channel convolution operation is represented as follows:

[0022]

[0023] The column interpolation transformation layer of the second layer signal aggregation layer performs an upsampling operation on the signal in the X direction in advance. The upsampling method uses a linear interpolation method, and the operation function is defined as: Upsample liner () which makes the signal dimension in the X direction approach the size of I b ,

[0024] SR

[522]

[694] = Upsample liner (SR

[261]

[694] ) ;

[0025] The column aggregation transformation layer of the second layer signal aggregation layer further aggregates the upsampled signal in the X direction. A one-dimensional convolution operation is used in the X direction for signal aggregation. The convolution kernel size is 32, the convolution kernel step is 1, and the number of 0 signals filled at both ends is 18. An independent convolution kernel is set for each column, and there are a total of 694 convolution kernels with a size of 128. The j-th row signal convolution operation is represented as follows:

[0026]

[0027] After signal aggregation, the signal dimension is SR

[527]

[694] .

[0028] Further, the third layer signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer:

[0029] The row aggregation transform layer of the third layer signal aggregation layer aggregates the data collected by each channel in the Z direction, and two-stage one-dimensional convolution operation is used for signal aggregation, the first stage convolution kernel size is 5, convolution kernel step is 1, and the number of signal padding at both ends is 0, and an independent convolution kernel is set for each channel, and there are a total of 527 convolution kernels with a size of 5, and the i-th column convolution operation is represented as follows:

[0030]

[0031] The second stage convolution operation uses a convolution kernel with a size of 3, a convolution kernel step of 1, and a signal padding number of 0 at both ends, and an independent convolution kernel is set for each channel, so there are a total of 527 convolution kernels with a size of 3, and the i-th channel convolution operation is represented as follows:

[0032]

[0033] The column aggregation transform layer of the third layer signal aggregation layer further aggregates the up-sampled signal in the X direction, and one-dimensional convolution operation is used in the X direction during signal aggregation, the convolution kernel size is 16, the convolution kernel step is 1, and the number of signal padding at both ends is 8, and an independent convolution kernel is set for each column, and there are a total of 688 convolution kernels with a size of 16, and the j-th row signal convolution operation is represented as follows:

[0034]

[0035] After signal aggregation, the signal dimension is SR

[528]

[688] The final image I b has the same dimension size.

[0036] Further, the acquisition of the numerical value V i,j of any point by the multi-scale signal synthesis module can be obtained by the following formula:

[0037] V i,j

[0038] = Max(Conv [3*3][1] (SR i,j ), Conv [3*3][2] (SR i,j ), Conv [3*3][3] (SR i,j ), Conv [3*3][4] (SR i,j ))

[0039] Where, Conv [3*3][1] (SR i,j ), Conv [3*3][2] (SR i,j ), Conv [3*3][3] (SRi,j ), Conv [3*3][4] (SR i,j ) respectively represent convolution operation in the range of 3*3 centered on SR i,j , there are 9 point values in the range of 3*3, and the intervals between points are 1, 2, 3 and 4 respectively;

[0040] The above four convolution operations are used to further aggregate signals at different scales, and V i,j is obtained by Max() maximum value operation, so as to effectively realize the function of two-dimensional multi-scale envelope detection.

[0041] Further, the imaging sub-network encoder is composed of four down-sampling modules, which are D1, D2, D3 and D4 modules respectively, the structures of D1, D2, D3 and D4 are the same, the width, height and channel number of the input image and the output image are different, and each module is embedded with an MSSC module, which can be expressed by mathematical formula as:

[0042] Output

[0043] =MaxPool2d [2*2] (Conv [1*1] (Cat(Input,Relu(MSSC [3*3] (Relu(

[0044] MSSC [5*5] (Input)))))))

[0045] Wherein, MSSC [3*3] represents a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5*5] represents a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() represents a Relu activation function, Cat() represents channel merging, Conv [1*1] represents 1*1 convolution operation, MaxPool2d [2*2] represents two-dimensional maximum pooling operation and down-sampling operation, and the down-sampling size is 2*2.

[0046] The input size of D1 module is 1*528*688, and the output size is 32*264*344, which is expressed as Output D1 .

[0047] The input size of D2 module is 32*264*344, and the output size is 64*132*172, which is expressed as Output D2 .

[0048] The input size of the D3 module is 64*132*172, and the output size is 128*66*86, denoted as Output D3 ;

[0049] The input size of the D4 module is 128*66*86, and the output size is 256*33*43, denoted as Output D4 ;

[0050] The intermediate layer operation can be expressed by a mathematical formula as follows:

[0051] Output = Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5]

[0052] (Input))))))

[0053] The input size of the intermediate layer is 256*33*43, and the output size is 256*33*43, denoted as Output C ;

[0054] The imaging subnetwork decoder is composed of four up-sampling modules, namely U4, U3, U2, and U1 modules. The U4, U3, U2, and U1 modules have the same structure, different input image and output image width, height, and channel number, and each module is embedded with an MSSC module, which can be expressed by a mathematical formula as follows:

[0055] Output

[0056] = Relu(MSSC [3*3] (Relu(MSSC [5*5] (Cat(ConvT [2*2] (Input1),

[0057] ConvT [2*2] (Input2))))))

[0058] Wherein, MSSC [3*3] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 3*3. MSSC [5*5] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 5*5. Relu() represents a Relu activation function. Cat() represents channel merging. ConvT [2*2] represents a 2*2 deconvolution operation.

[0059] The input of the U4 module is Output D4 , and the output is Output CThe size is 256*33*43, and the output size is 128*66*86, denoted as Output U4 ;

[0060] The U3 module input is Output D3 , Output U4 The size is 128*66*86, and the output size is 64*132*172, denoted as Output U3 ;

[0061] The U2 module input is Output D2 , Output U3 The size is 64*132*172, and the output size is 32*264*344, denoted as Output U2 ;

[0062] The U1 module input is Output D1 , Output U2 , the size is 32*264*344, and the output size is 1*528*688, denoted as Output U1 .

[0063] The output layer operation can be expressed by a mathematical formula as follows:

[0064] Output = Relu (Conv [1*1] (Cat (Input, Relu (MSSC [3*3] (Relu (MSSC [5*5]

[0065] (Input)))))))

[0066] The output layer input is Output U1 The size is 1*528*688, and the output size is 1*528*688, denoted as Output F ;

[0067] Through the design of the above coding and decoding network, the 75-degree imaging result is used as a label, and the imaging sub-network is trained to realize the functions of two-dimensional envelope detection, denoising and high-resolution reconstruction on the basis of input aggregated signals SR

[528]

[688] , so as to obtain a high-contrast and high-resolution image.

[0068] In the second aspect of the present application, the technical scheme provided is: an ultrasonic prostate imaging segmentation method based on sparse angle amplification, which is based on the imaging result of the imaging method in any one of the above aspects and includes the following steps:

[0069] standardized to a standard normal distribution by a Batch Normalization operation;

[0070] Data encoding is performed by a segmentation subnetwork encoder composed of four down-sampling modules, namely D1, D2, D3 and D4 modules, which have the same structure, different width, height and channel numbers of input images and output images, and each module is embedded with an MSSC module, which can be expressed by a mathematical formula as follows:

[0071] Output

[0072] =MaxPool2d [2*2] (Conv [1*1] (Cat(Input,Relu(MSSC [3*3] (Relu(

[0073] MSSC [5*5] (Input)))))))

[0074] wherein, MSSC [3*3] represents a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5*5] represents a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() represents a Relu activation function, Cat() represents channel merging, Conv [1*1] represents a 1*1 convolution operation, MaxPool2d [2*2] represents a two-dimensional maximum pooling operation, and a down-sampling operation with a down-sampling size of 2*2;

[0075] The input size of the D1 module is 1*528*688, and the output size is 32*264*344, which is represented as Output D1 .

[0076] The input size of the D2 module is 32*264*344, and the output size is 64*132*172, which is represented as Output D2 .

[0077] The input size of the D3 module is 64*132*172, and the output size is 128*66*86, which is represented as Output D3 .

[0078] The input size of the D4 module is 128*66*86, and the output size is 256*33*43, which is represented as Output D4 .

[0079] Operation is performed through the middle layer of the segmentation subnetwork:

[0080] Output = Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Input)))))), [5*5]

[0081] (Input)))))),

[0082] The input size of the middle layer of the split subnetwork is 256*33*43, and the output size is 256*33*43, which is denoted as Output C ;

[0083] Data decoding is performed by the split subnetwork decoder, which is composed of four up-sampling modules, namely U4, U3, U2, and U1 modules. The U4, U3, U2, and U1 modules have the same structure, different input image and output image width, height, and channel number, and each module is embedded with a MSSC module, which can be expressed by a mathematical formula as follows:

[0084] Output

[0085] = Relu(MSSC [3*3] (Relu(MSSC [5*5] (Cat(ConvT [2*2] (Input1),

[0086] ConvT [2*2] (Input2))))))

[0087] Wherein, MSSC [3*3] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 3*3. MSSC [5*5] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 5*5. Relu() represents a Relu activation function. Cat() represents channel merging. ConvT [2*2] represents a 2*2 deconvolution operation.

[0088] The input of the U4 module is Output D4 , and the size of Output C is 256*33*43. The output size is 128*66*86, denoted as Output U4 .

[0089] The input of the U3 module is Output D3 , and the size of Output U4 is 128*66*86. The output size is 64*132*172, denoted as Output U3 .

[0090] The U2 module input is Output D2 , Output U3 The size is 64*132*172, and the output size is 32*264*344, represented as Output U2 ;

[0091] The U1 module input is Output D1 , Output U2 , the size is 32*264*344, and the output size is 1*528*688, represented as Output U1 ;

[0092] Through the segmentation sub-network output layer operation:

[0093] Output = Relu(Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5]

[0094] (Input)))))))

[0095] The segmentation sub-network output layer input is Output U1 The size is 1*528*688, and the output size is 1*528*688, represented as Output F ;

[0096] Through the above coding and decoding network design, the lesion segmentation label is utilized, and through training, the segmentation sub-network can realize the function of lesion segmentation on the basis of inputting the aggregated signal SR

[528]

[688] .

[0097] In a third aspect, the present application provides a technical solution of an ultrasonic prostate imaging system based on sparse angle amplification, comprising:

[0098] A signal receiving module is configured to collect sparse angle RF ultrasonic signal data through a transducer;

[0099] A preprocessing module is configured to preprocess the collected sparse angle RF ultrasonic signal data through a signal preprocessing network to obtain a preliminary signal aggregation result, wherein the signal preprocessing network has three signal aggregation layers, and the third signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer, so that the aggregated signal is aligned with the dimension of the final output image in the dimension.

[0100] An imaging module is configured to input the preprocessed preliminary signal aggregation result into an imaging subnetwork and obtain an imaging result, wherein the imaging subnetwork is a codec structured deep neural network model, which comprises a multi-scale signal synthesis module, an imaging subnetwork encoder, an intermediate layer, an imaging subnetwork decoder and an output layer.

[0101] In a fourth aspect, the present application provides an ultrasonic prostate imaging and segmentation system based on sparse angle amplification, which comprises the imaging system and further comprises:

[0102] A segmentation module is configured to normalize the imaging data into a standard normal distribution through a Batch Normalization operation, and then realize lesion segmentation based on the preliminary signal aggregation result by using a lesion segmentation label through a segmentation subnetwork.

[0103] In a fifth aspect, the present application provides an electronic device, which comprises:

[0104] A memory is configured to store a processing program.

[0105] A processor is configured to realize the imaging method or the imaging segmentation method according to any one of the above aspects when executing the processing program.

[0106] Compared with the prior art, the present application has the following advantages and positive effects:

[0107] Based on a small amount of sparse angle RF signals, high-resolution and high-contrast imaging results can be amplified and recovered, compared with the traditional ultrasonic imaging mainly based on the focused scanning mode and the multi-angle complex plane wave mode, the present application can significantly reduce the imaging time by more than 80%, and has higher contrast definition. High-resolution, high-contrast and high-speed imaging can be realized. Through the prostate inflammation perception segmentation network constructed by the present application, high-resolution and high-contrast output of the prostate inflammation lesion area can be realized, compared with the existing artificial marking technology, the present application can automatically output the marking result, and the efficiency is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0108] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings, in which:

[0109] Figure 1The schematic diagram of the sparse angle expansion ultrasonic prostate inflammation imaging and perception segmentation network architecture of the application;

[0110] Figure 2 The schematic diagram of the signal preprocessing network of the application;

[0111] Figure 3 The internal architecture of the imaging subnetwork of the application;

[0112] Figure 4 The internal architecture of the segmentation subnetwork of the application;

[0113] Figure 5 The imaging implementation process of the sparse angle expansion ultrasonic prostate inflammation imaging and perception segmentation network of an embodiment of the application;

[0114] Figure 6 The imaging effect comparison of the sparse angle expansion ultrasonic prostate inflammation imaging and perception segmentation. DETAILED DESCRIPTION

[0115] The application will be further described below in conjunction with the drawings and specific embodiments. The advantages and features of the application will be more apparent according to the following description and claims. It should be noted that the drawings are very simplified and all use non-precise ratios, only to facilitate, clearly assist in explaining the purpose of the embodiments of the application.

[0116] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the application are only used to explain the relative position relationship, motion condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications will also change accordingly.

[0117] Traditional ultrasonic imaging is mainly based on focused scanning and multi-angle composite plane wave, and the imaging time is long and the efficiency is low. Based on single-angle focused imaging or single-angle plane wave imaging, the imaging resolution is low and the contrast is low.

[0118] Referring to Figure 1 The sparse angle expansion ultrasonic prostate inflammation imaging and perception segmentation network architecture constructed by the application is based on a small amount of sparse angle RF signals, which can expand and restore high-resolution and high-contrast imaging results. Compared with traditional ultrasonic imaging mainly based on focused scanning and multi-angle composite plane wave imaging, the imaging time of the application can be reduced by more than 80%, and the contrast is higher. High-resolution, high-contrast and high-speed imaging can be achieved. And through the prostate inflammation perception segmentation network constructed by the application, high-resolution and high-contrast output of prostate inflammation lesion area can be achieved. Compared with the existing artificial marking technology, the application can automatically output the marking results, and significantly improve the efficiency.

[0119] First embodiment

[0120] The embodiment of the application is an ultrasound prostate imaging method based on sparse angle amplification, comprising the following steps:

[0121] Sparse angle RF ultrasound signal data is collected by receiving transducers;

[0122] The collected sparse angle RF ultrasound signal data is preprocessed by a signal preprocessing network to obtain a preliminary signal aggregation result, wherein the signal preprocessing network has three signal aggregation layers, and the third signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer, so that the aggregated signal is aligned in dimension with the dimension of the final output image.

[0123] The preprocessed preliminary signal aggregation result is input into an imaging subnetwork to obtain an imaging result, wherein the imaging subnetwork is a deep neural network model with a codec structure, having a multi-scale signal synthesis module, an imaging subnetwork encoder, an intermediate layer, an imaging subnetwork decoder, and an output layer. The multi-scale signal synthesis module uses multi-scale hole convolution and corresponding maximum value operation at any point to aggregate signals in multiple different scale ranges and take the maximum value as the final value in terms of image value at any point. The imaging subnetwork encoder, the intermediate layer, the imaging subnetwork decoder, and the output layer output the imaging result based on the final value.

[0124] In the technical solution of the embodiment, the signal preprocessing network mainly has three signal aggregation layers, wherein the first and second signal aggregation layers have the same basic structure but perform different operations. The third signal aggregation layer does not have an intermediate column interpolation transformation layer compared with the first two layers. The purpose of this layer design is to preliminarily aggregate the collected RF signals, and secondly to align the aggregated signal in dimension with the dimension of the final output image, and finally to output an image representing the size of W*H. Through this layer design, the signal loss problem of dimension alignment using downsampling in existing deep learning algorithms is avoided.

[0125] Taking 128 elements, a center frequency of 7.24 MHz, a sampling frequency of 27.78 MHz, and an imaging depth of 5.0 cm as an example, the collected original RF signal is represented as F, and the final output image size is 528*688. The following layers are described.

[0126] Referring to Figure 2 , the first layer of the signal preprocessing network has a row aggregation transformation layer, a column interpolation transformation layer, and a column aggregation transformation layer:

[0127] The row aggregation transform layer of the first layer signal aggregation layer aggregates the data collected by each channel in the Z direction (longitudinal direction) by using two-stage one-dimensional convolution operation. The first-stage convolution operation uses a convolution kernel with a size of 79, a convolution kernel step of 3, and 39 zeros filled at both ends of the signal. Each channel is provided with an independent convolution kernel, and there are 128 convolution kernels with a size of 79. The convolution operation of the ith channel is represented as follows:

[0128]

[0129] The second-stage convolution operation uses a convolution kernel with a size of 56, a convolution kernel step of 1, and 0 zeros filled at both ends of the signal. Each channel is provided with an independent convolution kernel, and there are 128 convolution kernels with a size of 56. The convolution operation of the ith channel is represented as follows:

[0130]

[0131] The column interpolation transform layer of the first layer signal aggregation layer performs signal up-sampling operation in the X direction in advance. The up-sampling method uses linear interpolation, and the operation function is defined as: the signal dimension in the X direction is doubled, and the X direction dimension is approximated.

[0132] The column aggregation transform layer of the first layer signal aggregation layer performs one-dimensional convolution operation in the X direction during signal aggregation. The convolution kernel has a size of 128, a convolution kernel step of 1, and 66 zeros filled at both ends of the signal. Each column is provided with an independent convolution kernel, and there are 713 convolution kernels with a size of 128. The convolution operation of the jth row signal is represented as follows:

[0133]

[0134] After signal aggregation, the signal dimension is SR

[261]

[713] .

[0135] Further, the second layer signal aggregation layer in the signal preprocessing network has a row aggregation transform layer, a column interpolation transform layer, and a column aggregation transform layer.

[0136] The row aggregation transform layer of the second layer signal aggregation layer aggregates the data collected by each channel in the Z direction by using two-stage one-dimensional convolution operation. The first-stage convolution kernel has a size of 14, a convolution kernel step of 1, and 0 zeros filled at both ends of the signal. Each channel is provided with an independent convolution kernel, and there are 261 convolution kernels with a size of 14. The convolution operation of the ith column is represented as follows:

[0137]

[0138] The second-level convolution operation uses a convolution kernel size of 7, a convolution kernel step of 1, and 0 number of 0s for signal padding at both ends. An independent convolution kernel is set for each channel, and there are a total of 700 convolution kernels with a size of 7. The i-th channel convolution operation is represented as follows:

[0139]

[0140] The column interpolation transformation layer of the second-layer signal aggregation layer performs an upsampling operation on the signal in the X direction. The upsampling method uses a linear interpolation method, and the operation function is defined as: Upsample liner , which makes the signal dimension in the X direction approach the size of I b .

[0141] SR

[522]

[694] = Upsample liner (SR

[261]

[694] ).

[0142] The column aggregation transformation layer of the second-layer signal aggregation layer further aggregates the upsampled signal in the X direction. A one-dimensional convolution operation is used in the X direction during signal aggregation. The convolution kernel size is 32, the convolution kernel step is 1, and the number of 0s for signal padding at both ends is 18. An independent convolution kernel is set for each column, and there are a total of 694 convolution kernels with a size of 128. The j-th row signal convolution operation is represented as follows:

[0143]

[0144] After signal aggregation, the signal dimension is SR

[527]

[694] .

[0145] Further, the third-layer signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer.

[0146] The row aggregation transformation layer of the third-layer signal aggregation layer aggregates the data collected by each channel in the Z direction. A two-stage one-dimensional convolution operation is used during signal aggregation. The first-stage convolution kernel size is 5, the convolution kernel step is 1, and the number of 0s for signal padding at both ends is 0. An independent convolution kernel is set for each channel, and there are a total of 527 convolution kernels with a size of 5. The i-th column convolution operation is represented as follows:

[0147]

[0148] The second-stage convolution operation uses a convolution kernel size of 3, a convolution kernel step of 1, and 0 number of 0s for signal padding at both ends. An independent convolution kernel is set for each channel, so there are a total of 527 convolution kernels with a size of 3 in this layer. The i-th channel convolution operation is represented as follows:

[0149]

[0150] The column aggregation transform layer of the third layer signal aggregation layer further aggregates the up-sampled signals in the X direction. A one-dimensional convolution operation is used in the X direction during signal aggregation. The convolution kernel size is 16, the convolution kernel step is 1, and the number of 0s added to the ends of the signal is 8. An independent convolution kernel is set for each column. There are a total of 688 convolution kernels with a size of 16. The jth row signal convolution operation is represented as follows:

[0151]

[0152] After signal aggregation is completed, the signal dimension is SR

[528]

[688] The final image I b has the same dimension size.

[0153] In the design of the imaging sub-network, further improvements are made on the basis of the deep neural network encoding-decoding structure. The improvements are mainly made through the comparison of sparse angle ultrasound imaging quality and 75 angle ultrasound imaging quality and the characteristics of the signal output of the signal preprocessing network.

[0154] From the comparison of sparse angle ultrasound imaging quality and 75 angle ultrasound imaging quality, it can be seen that the contrast and resolution of the image of 75 angle imaging quality are greatly improved. This improvement in imaging quality can be achieved by image denoising technology and super-resolution reconstruction technology. The encoding-decoding structure can effectively realize image denoising and resolution reconstruction through training.

[0155] The original RF signal is processed by the signal preprocessing network, and the output can obtain the preliminary signal aggregation result, which is similar to the traditional DAS. However, the network cannot completely realize the effect of DAS beamforming. Therefore, in the imaging sub-network, we propose a multi-scale signal synthesis module. Based on the aggregated signal, the module uses multi-scale hollow convolution and corresponding maximum value operation at any point to aggregate signals in multiple different scale ranges and take the maximum value as the final value at the image point. Its function is similar to the two-dimensional envelope detection in the traditional algorithm.

[0156] Further, the multi-scale signal synthesis module can obtain the value V i,j of any point by the following formula:

[0157] V i,j

[0158] = Max(Conv [3*3][1] (SR i,j ), Conv [3*3][2] (SR i,j ), Conv [3*3][3] (SR i,j ), Conv [3*3][4] (SRi,j ))

[0159] wherein, Conv [3*3][1] (SR i,j ), Conv [3*3][2] (SR i,j ), Conv [3*3][3] (SR i,j ), Conv [3*3][4] (SR i,j ) represent the convolution operation in the range of 3*3 centered on SR i,j , there are 9 point values in the range of 3*3, and the intervals between points are 1, 2, 3 and 4 respectively;

[0160] The above four convolution operations are used to further aggregate signals at different scales, and V i,j is obtained through Max() maximum value operation, so as to effectively realize the function of two-dimensional multi-scale envelope detection. At the same time, the module can also effectively realize the function of two-dimensional multi-scale feature extraction in the image segmentation sub-network.

[0161] Further, referring to Figure 3 , the imaging sub-network encoder is composed of four down-sampling modules, which are D1, D2, D3 and D4 modules. The structures of D1, D2, D3 and D4 are the same, and the width, height and channel number of the input image and the output image are different. Each module is embedded with an MSSC module, which can be expressed by mathematical formula as follows:

[0162] Output

[0163] = MaxPool2d [2*2] (Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(

[0164] MSSC [5*5] (Input)))))))

[0165] wherein, MSSC [3*3] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 3*3. MSSC [5*5] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 5*5. Relu() represents a Relu activation function. Cat() represents channel merging. Conv [1*1] represents a 1*1 convolution operation. MaxPool2d [2*2] represents a two-dimensional maximum pooling operation and a down-sampling operation, and the down-sampling size is 2*2.

[0166] The input size of the D1 module is 1*528*688, and the output size is 32*264*344, denoted as Output D1 ;

[0167] The input size of the D2 module is 32*264*344, and the output size is 64*132*172, denoted as Output D2 ;

[0168] The input size of the D3 module is 64*132*172, and the output size is 128*66*86, denoted as Output D3 ;

[0169] The input size of the D4 module is 128*66*86, and the output size is 256*33*43, denoted as Output D4 ;

[0170] The intermediate layer operation can be expressed by a mathematical formula as follows:

[0171] Output = Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5]

[0172] (Input))))))

[0173] The input size of the intermediate layer is 256*33*43, and the output size is 256*33*43, denoted as Output C ;

[0174] The imaging subnetwork decoder is composed of four up-sampling modules, namely U4, U3, U2, and U1 modules. The U4, U3, U2, and U1 modules have the same structure, different input image and output image width, height, and channel number, and each module is embedded with an MSSC module, which can be expressed by a mathematical formula as follows:

[0175] Output

[0176] = Relu(MSSC [3*3] (Relu(MSSC [5*5] (Cat(ConvT [2*2] (Input1),

[0177] ConvT [2*2] (Input2))))))

[0178] Wherein, MSSC [3*3] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 3*3. MSSC [5*5]denotes a multi-scale signal synthesis module, the size of the convolution kernel used in the module is 5*5, Relu() denotes a Relu activation function, Cat() denotes channel merging, ConvT [2*2] denotes a 2*2 deconvolution operation;

[0179] The input of the U4 module is Output D4 , Output C The size is 256*33*43, and the output size is 128*66*86, denoted as Output U4 ;

[0180] The input of the U3 module is Output D3 , Output U4 The size is 128*66*86, and the output size is 64*132*172, denoted as Output U3 ;

[0181] The input of the U2 module is Output D2 , Output U3 The size is 64*132*172, and the output size is 32*264*344, denoted as Output U2 ;

[0182] The input of the U1 module is Output D1 , Output U2 , the size is 32*264*344, and the output size is 1*528*688, denoted as Output U1 .

[0183] The output layer operation can be expressed by a mathematical formula as follows:

[0184] Output = Relu(Conv [1-1] (Cat(Input, Relu(MSSC [3-3] (Relu(MSSC [5-5]

[0185] (Input)))))))

[0186] The input of the output layer is Output U1 The size is 1*528*688, and the output size is 1*528*688, denoted as Output F ;

[0187] Through the design of the above coding and decoding network, the 75-degree imaging result is used as a label, and the imaging sub-network is trained to make the imaging sub-network in the input aggregated signal SR

[528]

[688] On the basis of the above, the function of two-dimensional envelope detection, denoising and high-resolution reconstruction is realized, so as to obtain an image with high contrast and high resolution.

[0188] Based on the same inventive concept, the embodiment also provides an ultrasonic prostate imaging system based on sparse angle amplification, comprising:

[0189] The signal receiving module is configured to collect sparse angle RF ultrasonic signal data through the transducer.

[0190] The preprocessing module is configured to preprocess the collected sparse angle RF ultrasonic signal data through a signal preprocessing network to obtain a preliminary signal aggregation result, wherein the signal preprocessing network has three signal aggregation layers, and the third signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer, so that the aggregated signal is aligned in dimension with the dimension of the final output image.

[0191] The imaging module is configured to input the preprocessed preliminary signal aggregation result into an imaging subnetwork and obtain an imaging result, wherein the imaging subnetwork is a deep neural network model with a codec structure, having a multi-scale signal synthesis module, an imaging subnetwork encoder, an intermediate layer, an imaging subnetwork decoder and an output layer.

[0192] Second embodiment

[0193] In the design of the segmentation subnetwork, we further improved the deep neural network codec structure, integrated BN and MSSC modules, to enhance the ability to segment the lesion area from the aggregated signal.

[0194] Based on the same inventive concept, the embodiment provides an ultrasonic prostate imaging segmentation method based on sparse angle amplification, which is based on the imaging result of the imaging method according to any one of the first embodiment, see Figure 4 , comprising the following steps:

[0195] The imaging data is standardized to a standard normal distribution through Batch Normalization operation;

[0196] Data encoding is performed by a split sub-network encoder composed of four down-sampling modules, namely D1, D2, D3 and D4 modules. The D1, D2, D3 and D4 modules have the same structure, different input image and output image width, height and channel number, and each module is embedded with an MSSC module, which can be expressed by a mathematical formula as follows:

[0197] Output

[0198] =MaxPool2d [2*2] (Conv [1*1] (Cat(Input,Relu(MSSC [3*3] (Relu(

[0199] MSSC [5*5] (Input)))))))

[0200] Wherein, MSSC [3*3] represents a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5*5] represents a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() represents a Relu activation function, Cat() represents channel merging, Conv [1*1] represents a 1*1 convolution operation, MaxPool2d [2*2] represents a two-dimensional maximum pooling operation, and a down-sampling operation, the down-sampling size is 2*2.

[0201] The input size of the D1 module is 1*528*688, and the output size is 32*264*344, which is represented as Output D1 .

[0202] The input size of the D2 module is 32*264*344, and the output size is 64*132*172, which is represented as Output D2 .

[0203] The input size of the D3 module is 64*132*172, and the output size is 128*66*86, which is represented as Output D3 .

[0204] The input size of the D4 module is 128*66*86, and the output size is 256*33*43, which is represented as Output D4 .

[0205] Through the operation of the split sub-network middle layer:

[0206] Output=Conv [1*1] (Cat(Input,Relu(MSSC [3*3](Relu(MSSC [5*5]

[0207] (Input)))))),

[0208] The input size of the middle layer of the segmentation subnetwork is 256*33*43, and the output size is 256*33*43, which is denoted as Output C ;

[0209] Data decoding is performed by the segmentation subnetwork decoder, which is composed of four up-sampling modules, namely U4, U3, U2, and U1 modules. The U4, U3, U2, and U1 modules have the same structure, different input image and output image width, height, and channel number, and each module is embedded with an MSSC module. The mathematical formula can be expressed as:

[0210] Output

[0211] = Relu(MSSC [3*3] (Relu(MSSC [5*5] (Cat(ConvT [2*2] (Input1),

[0212] ConvT [2-2] (Input2))))))

[0213] Wherein, MSSC [3-3] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 3*3. MSSC [5-5] represents a multi-scale signal synthesis module, and the convolution kernel size used in the module is 5*5. Relu() represents a Relu activation function. Cat() represents channel merging. ConvT [2-2] represents a 2*2 deconvolution operation.

[0214] The input of the U4 module is Output D4 , and the output size is 128*66*86, which is denoted as Output C . U4 ;

[0215] The input of the U3 module is Output D3 , and the output size is 64*132*172, which is denoted as Output U4 . U3 ;

[0216] The input of the U2 module is Output D2 , and the output size is 32*264*364, which is denoted as Output U3 .The size is 64*132*172, and the output size is 32*264*344, denoted as Output U2 ;

[0217] The U1 module input is Output D1 , Output U2 , the size is 32*264*344, and the output size is 1*528*688, denoted as Output U1 ;

[0218] Through the segmentation sub-network output layer operation:

[0219] Output = Relu(Conv [1-1] (Cat(Input, Relu(MSSC [3-3] (Relu(MSSC [5-5]

[0220] (Input)))))))

[0221] The segmentation sub-network output layer input is Output U1 The size is 1*528*688, and the output size is 1*528*688, denoted as Output F ;

[0222] Through the above coding and decoding network design, the lesion segmentation label is utilized, and through training, the segmentation sub-network can realize the function of lesion segmentation on the basis of inputting the aggregated signal SR

[528]

[688] .

[0223] Based on the same inventive concept, the embodiment also provides an ultrasound prostate imaging segmentation system based on sparse angle amplification, comprising the imaging system of the first embodiment, further comprising:

[0224] The segmentation module is used for normalizing the imaging data to a standard normal distribution through Batch Normalization operation, and then realizing lesion segmentation through the segmentation sub-network on the basis of the preliminary signal aggregation result by using the lesion segmentation label.

[0225] For example, referring to Figure 5 and Figure 6 , the following process is performed:

[0226] S1: Sparse acquisition of n=1 angle plane wave RF data << full angle=75 RF data

[0227] Start RF signal collection by professional ultrasonic signal device, define full-angle excitation as 75 angle excitation ultrasonic, that is, m=75; the application only collects RF of 1 angle in m angles randomly, that is, n=1; at this time, it meets the sparse angle described in the application, that is, n << 75.

[0228] S2: signal filtering, denoising and normalization preprocessing

[0229] Input the 1 angle RF collected in process S2 into the signal preprocessing module, and the internal preprocessing network performs signal filtering, denoising, normalization and aggregation; load the aggregated signal to the heterogeneous twin sub-network, which includes the sparse angle augmented ultrasonic prostatitis imaging sub-network and the prostatitis inflammation lesion area perception segmentation sub-network;

[0230] S3: load the sparse angle augmented ultrasonic prostatitis imaging sub-network

[0231] Through the trained sparse angle augmented ultrasonic prostatitis imaging sub-network, based on the collected 1 angle RF data, the imaging effect is augmented, and the sparse angle imaging effect is augmented to m=75 imaging effect;

[0232] S4: load the prostatitis inflammation lesion area perception segmentation sub-network and output the segmentation result

[0233] Through the trained inflammation lesion area perception segmentation sub-network, based on the collected 1 angle RF data, the lesion area is segmented, and the prostatitis inflammation lesion area is accurately segmented;

[0234] S5: output the imaging result

[0235] Through process S3, the augmented high-resolution and high-contrast imaging result is obtained, denoted as I m×n , wherein m and n are the row number and column number of image pixels respectively.

[0236] S6: output the segmentation result

[0237] Through process S4, the result after lesion area segmentation is obtained, denoted as S m×n , and S m×n is equal in size, wherein S m×n Each pixel value in S m×n is 0 or 1.

[0238] S7: pixel fusion

[0239] The pixel-level fusion result of the results obtained in processes S5 and S6 is denoted as O m×n , and the fusion formula is as follows:

[0240] O m×n =(2-S m×n)-0.5-I m×n

[0241] S8: high-speed output of imaging and segmentation results with high resolution and high contrast

[0242] S8: high-speed output of imaging and segmentation results with high resolution and high contrast m×n .

[0243] Based on the same concept, the present application also provides an electronic device, comprising: a memory for storing a processing program; a processor for implementing the imaging method or the imaging segmentation method according to any one of the above embodiments when executing the processing program.

[0244] Based on the same concept, the present application also provides a readable storage medium, which stores a processing program, and the processing program is executed by a processor to implement the imaging method or the imaging segmentation method according to any one of the above embodiments.

[0245] The imaging method or the imaging segmentation method, if implemented in the form of program instructions and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the embodiments can be embodied in the form of software, and the computer software is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present disclosure. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0246] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific implementation of the above-described system and device can be referred to the corresponding process in the foregoing method embodiments.

[0247] The embodiments of the present application are described in detail above in combination with the accompanying drawings, but the present application is not limited to the above-described embodiments. Even if various changes are made to the present application, as long as the changes fall within the scope of the claims of the present application and equivalent technologies, they still fall within the protection scope of the present application.

Claims

1. A method of ultrasound prostate imaging based on sparse angle augmentation, characterized by, The method comprises the following steps: acquiring sparse angle RF ultrasound signal data through a receiving transducer; preprocessing the acquired sparse angle RF ultrasound signal data through a signal preprocessing network to obtain a preliminary signal aggregation result, wherein the signal preprocessing network has three signal aggregation layers, and the third signal aggregation layer comprises a row aggregation transformation layer and a column aggregation transformation layer, so that the aggregated signal is aligned with the dimension of the final output image in the dimension; inputting the preprocessed preliminary signal aggregation result into an imaging subnetwork and obtaining an imaging result, wherein the imaging subnetwork is a deep neural network model with a codec structure, comprising a multi-scale signal synthesis module, an imaging subnetwork encoder, an intermediate layer, an imaging subnetwork decoder and an output layer, the multi-scale signal synthesis module uses multi-scale hole convolution and corresponding maximum value operation at any point to aggregate signals in multiple different scale ranges and take the maximum value as the final value in the aspect of image value at any point, and the imaging subnetwork encoder, the intermediate layer, the imaging subnetwork decoder and the output layer output the imaging result based on the final value.

2. The sparse-angle augmented-based ultrasound prostate imaging method of claim 1, wherein, The first signal aggregation layer in the signal preprocessing network has a row aggregation transformation layer, a column interpolation transformation layer and a column aggregation transformation layer: The row aggregation transformation layer of the first signal aggregation layer performs signal aggregation on the data collected by each channel in the Z direction, and two-stage one-dimensional convolution operation is adopted in the signal aggregation, the first-stage convolution operation adopts a convolution kernel size of 79, a convolution kernel step of 3 and 0 padding of 79 / / 2=39 at both ends, an independent convolution kernel is set for each channel, and there are a total of 128 convolution kernels with a size of 79, and the convolution operation of the i-th channel is represented as follows: SR [i][768] = RF [i][2304] K 3 [i][79][39] ; The second-stage convolution operation adopts a convolution kernel size of 56, a convolution kernel step of 1 and 0 padding at both ends, an independent convolution kernel is set for each channel, and there are a total of 128 convolution kernels with a size of 56, and the convolution operation of the i-th channel is represented as follows: The column interpolation transformation layer of the first signal aggregation layer performs signal up-sampling operation in the X direction in advance, and the up-sampling method uses linear interpolation, and the operation function is defined as: so that the signal dimension is doubled in the X direction and the approximated X direction dimension is obtained; The column aggregation transformation layer of the first signal aggregation layer adopts one-dimensional convolution operation in the X direction in the signal aggregation, the convolution kernel size is 128, the convolution kernel step is 1, and the 0 padding at both ends is 66, an independent convolution kernel is set for each column, and there are a total of 713 convolution kernels with a size of 128, and the convolution operation of the j-th row signal is represented as follows: The signal dimension is SR after signal aggregation is completed [261][713] .

3. The sparse-angle augmented-based ultrasound prostate imaging method of claim 1, wherein, The second signal aggregation layer in the signal preprocessing network has a row aggregation transformation layer, a column interpolation transformation layer and a column aggregation transformation layer: The row aggregation transformation layer of the second signal aggregation layer performs signal aggregation on the data collected by each channel in the Z direction, and two-stage one-dimensional convolution operation is adopted in the signal aggregation, the first-stage convolution kernel size is 14, the convolution kernel step is 1, and the 0 padding at both ends is 0, an independent convolution kernel is set for each channel, and there are a total of 261 convolution kernels with a size of 14, and the convolution operation of the i-th column is represented as follows: The second-level convolution operation adopts a convolution kernel size of 7, a convolution kernel step of 1, and 0 signal padding at both ends, and each channel is provided with an independent convolution kernel, and there are a total of 700 convolution kernels with a size of 7, and the i-th channel convolution operation is represented as follows: The column interpolation transform layer of the second layer signal aggregation layer performs upsampling operation of the signal in the X direction in advance, and the upsampling mode uses a linear interpolation mode, and the operation function is defined as: Upsample liner , which makes the signal dimension in the X direction approach the size of I b , SR [522][694] = Upsample liner (SR [261][694] ); The column aggregation transformation layer of the second layer of signal aggregation layers further aggregates the up-sampled signals in the X direction, and a one-dimensional convolution operation is adopted in the X direction during signal aggregation, the convolution kernel size is 32, the convolution kernel step is 1, the number of signal padding at both ends is 18, an independent convolution kernel is arranged for each column, there are a total of 694 convolution kernels with a size of 128, and the j-th row signal convolution operation is represented as follows: The signal dimension is SR after signal aggregation is completed [527][694] .

4. The sparse-angle-based augmented ultrasound prostate imaging method of claim 1, wherein, The third layer of signal aggregation layers includes row aggregation transformation layers and column aggregation transformation layers: The row aggregation transformation layer of the third layer of signal aggregation layers aggregates the data collected by each channel in the Z direction, and a two-stage one-dimensional convolution operation is adopted during signal aggregation, the first-stage convolution kernel size is 5, the convolution kernel step is 1, the number of signal padding at both ends is 0, an independent convolution kernel is arranged for each channel, and there are a total of 527 convolution kernels with a size of 5, and the i-th column convolution operation is represented as follows: The second-level convolution operation adopts a convolution kernel size of 3, a convolution kernel step of 1, and 0 signal padding at both ends, and each channel is provided with an independent convolution kernel, so there are a total of 527 convolution kernels with a size of 3 in this layer, and the i-th channel convolution operation is represented as follows: The column aggregation transformation layer of the third layer of signal aggregation layers further aggregates the up-sampled signals in the X direction, and a one-dimensional convolution operation is adopted in the X direction during signal aggregation, the convolution kernel size is 16, the convolution kernel step is 1, the number of signal padding at both ends is 8, an independent convolution kernel is arranged for each column, and there are a total of 688 convolution kernels with a size of 16, and the j-th row signal convolution operation is represented as follows: After the signal aggregation is completed, the signal dimension is SR [528][688] The final image I b The dimension sizes are consistent.

5. The sparse-angle-based augmented ultrasound prostate imaging method of claim 1, wherein, The multi-scale signal synthesis module obtains the value V of any point i,j The acquisition can be obtained by the following formula: V i,j = Max(Conv [3*3][1] (SR i,j ), Conv [3*3][2] (SR i,j ), Conv [3*3][3] (SR i,j ), Conv [3*3][4] (SR i,j )) wherein, Conv [3*3][1] (SR i,j ), Conv [3*3][2] (SR i,j ), Conv [3*3][3] (SR i,j ), Conv [3*3][4] (SR i,j ) respectively represent a convolution operation in a range of 3*3 centered on SR i,j , there are 9 point values in the range of 3*3, and the intervals between points are 1, 2, 3, and 4, respectively. The above four convolution operations are used to further aggregate signals at different scales, and V is obtained by Max() maximum operation i,j The envelope detection function of two dimensions and multiple scales is effectively realized.

6. The sparse angle amplification based ultrasonic prostate imaging method of claim 5, characterized in that, The imaging sub-network encoder is composed of four down-sampling modules, namely D1, D2, D3 and D4 modules, the input image and output image have different width, height and channel number in the D1, D2, D3 and D4 modules which have the same structure, and each module is embedded with a MSSC module, which can be expressed by a mathematical formula as follows: Output = MaxPool2d [2*2] (Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5] (Input))))))) wherein, MSSC [3*3] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5*5] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() denotes a Relu activation function, Cat() denotes channel merging, Conv [1*1] denotes a 1*1 convolution operation, MaxPool2d [2*2] denotes a two-dimensional maximum pooling operation and a down-sampling operation, and the down-sampling size is 2*2; The input size of the D1 module is 1*528*688, and the output size is 32*264*344, which is represented as Output D1 ; The D2 module input size is 32*264*344, and the output size is 64*132*172 represented as Output D2 ; The D3 module input size is 64*132*172, and the output size is 128*66*86 represented as Output D3 ; The D4 module input size is 128*66*86, and the output size is 256*33*43, which is represented as Output D4 ; The intermediate layer operation can be expressed by a mathematical formula as follows: Output = Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5] (Input)))))) The intermediate layer input size is 256*33*43, and the output size is 256*33*43, denoted as Output C ; The imaging sub-network decoder is composed of four up-sampling modules, namely U4, U3, U2 and U1 modules, the input image and output image have different width, height and channel number in the U4, U3, U2 and U1 modules which have the same structure, and each module is embedded with a MSSC module, which can be expressed by a mathematical formula as follows: Output = Relu(MSSC [3*3] (Relu(MSSC [5*5] (Cat(ConvT [2*2] (Input1), ConvT [2-2] (Input2)))))) wherein, MSSC [3-3] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5-5] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() denotes a Relu activation function, Cat() denotes channel merging, ConvT [2-2] denotes a 2*2 deconvolution operation; U4 module input is Output D4 , Output C Size is 256*33*43, output size is 128*66*86 represented as Output U4 ; U3 module input is Output D3 ,Output U4 Size is 128*66*86, output size is 64*132*172 represented as Output U3 ; U2 module input is Output D2 ,Output U3 Size is 64*132*172, output size is 32*264*344 represented as Output U2 ; The U1 module input is Output D1 ,Output U2 , with a size of 32*264*344, and the output size is 1*528*688, denoted as Output U1 . The output layer operation can be expressed by a mathematical formula as follows: Output = Relu(Conv [1-1] (Cat(Input, Relu(MSSC [3-3] (Relu(MSSC [5-5] (Input))))))) The output layer input is Output U1 with a size of 1*528*688 and an output size of 1*528*688 denoted as Output F ; By the design of the above coding network, the 75-degree imaging result is used as a label, and the imaging sub-network is trained to realize the functions of two-dimensional envelope detection, denoising and high-resolution reconstruction on the basis of the input aggregated signal SR [528][688] , so as to obtain a high-contrast and high-resolution image.

7. An ultrasound prostate imaging segmentation method based on sparse angle augmentation, which is based on the imaging result of the imaging method according to any one of claims 1-6, characterized in that, The method comprises the following steps: The imaging data is standardized to a standard normal distribution through a Batch Normalization operation; The data is encoded by a segmentation sub-network encoder, the segmentation sub-network encoder is composed of four down-sampling modules, namely D1, D2, D3 and D4 modules, the input image and output image have different width, height and channel number in the D1, D2, D3 and D4 modules which have the same structure, and each module is embedded with a MSSC module, which can be expressed by a mathematical formula as follows: Output = MaxPool2d [2*2] (Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5] (Input))))))) wherein, MSSC [3-3] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5*5] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() denotes a Relu activation function, Cat() denotes channel merging, Conv [1*1] denotes a 1*1 convolution operation, MaxPool2d [2*2] denotes a two-dimensional maximum pooling operation, and a down-sampling operation, the down-sampling size is 2*2; The input size of the D1 module is 1*528*688, and the output size is 32*264*344, which is represented as Output D1 ; The D2 module input size is 32*264*344, and the output size is 64*132*172 represented as Output D2 ; The D3 module input size is 64*132*172, and the output size is 128*66*86 represented as Output D3 ; The D4 module input size is 128*66*86, and the output size is 256*33*43, represented as Output D4 ; The operation is performed by splitting the intermediate layer of the sub-network: Output = Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5] (Input)))))) The input size of the middle layer of the segmentation sub-network is 256*33*43, and the output size is 256*33*43, which is represented as Output C ; Data decoding is performed by a split sub-network decoder composed of four up-sampling modules, namely U4, U3, U2, and U1 modules, which are the same in structure, but different in width, height, and channel number of input images and output images. Each module is embedded with an MSSC module, which can be expressed in mathematical formula as: Output = Relu(MSSC [3*3] (Relu(MSSC [5*5] (Cat(ConvT [2*2] (Input1), ConvT [2*2] (Input2)))))) wherein, MSSC [3*3] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 3*3, MSSC [5*5] denotes a multi-scale signal synthesis module, the convolution kernel size used in the module is 5*5, Relu() denotes a Relu activation function, Cat() denotes channel merging, ConvT [2*2] denotes a 2*2 deconvolution operation; U4 module input is Output D4 , Output C Size is 256*33*43, output size is 128*66*86 represented as Output U4 ; U3 module input is Output D3 ,Output U4 Size is 128*66*86, output size is 64*132*172 represented as Output U3 ; U2 module input is Output D2 ,Output U3 Size is 64*132*172, output size is 32*264*344 represented as Output U2 ; The U1 module input is Output D1 ,Output U2 , with a size of 32*264*344, and the output size is 1*528*688, denoted as Output U1 ; The operation is performed by splitting the output layer of the sub-network: Output = Relu(Conv [1*1] (Cat(Input, Relu(MSSC [3*3] (Relu(MSSC [5*5] (Input))))))) The input of the segmentation subnetwork output layer is Output U1 with a size of 1*528*688, and an output size of 1*528*688 is denoted as Output F ; Through the design of the above coding network, the lesion segmentation label is utilized, and through training, the segmentation sub-network can realize the function of lesion segmentation on the basis of inputting the aggregated signal SR [528][688] .

8. An ultrasound prostate imaging system based on sparse angle augmentation, characterized by, It includes: A signal receiving module for receiving sparse angle RF ultrasound signal data collected by a transducer; A preprocessing module for preprocessing the collected sparse angle RF ultrasound signal data to obtain a preliminary signal aggregation result through a signal preprocessing network, wherein the signal preprocessing network has three signal aggregation layers, and the third signal aggregation layer includes a row aggregation transformation layer and a column aggregation transformation layer to align the dimensions of the aggregated signal with the dimensions of the final output image; An imaging module for inputting the preprocessed preliminary signal aggregation result into an imaging sub-network and obtaining an imaging result, wherein the imaging sub-network is a deep neural network model with a coding and decoding structure, having a multi-scale signal synthesis module, an imaging sub-network encoder, an intermediate layer, an imaging sub-network decoder, and an output layer. The multi-scale signal synthesis module uses multi-scale dilated convolution and corresponding maximum value operation at any point to aggregate signals in multiple different scale ranges and take the maximum value as the final value in terms of image value at any point. The imaging sub-network encoder, intermediate layer, imaging sub-network decoder, and output layer output the imaging result based on the final value.

9. A sparse-angle repopulation based ultrasound prostate imaging segmentation system, comprising the imaging system of claim 8, characterized in that, It also includes: A segmentation module for normalizing imaging data to a standard normal distribution through a Batch Normalization operation, and then implementing lesion segmentation based on the preliminary signal aggregation result using lesion segmentation labels through a segmentation sub-network.

10. An electronic device, comprising: It includes: A memory for storing a processing program; A processor for implementing the imaging method or imaging segmentation method of any one of claims 1 to 7 when executing the processing program.

Images