Medical image reconstruction method and system for radiology department

Abstract

The invention provides a medical image reconstruction method and system for radiology department, and the method comprises the following steps: inputting a medical image, and carrying out the preprocessing of the medical image; inputting the preprocessed medical image as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model comprises a deep convolutional neural network and a reconstruction module based on a virtual single pixel; starting a deep learning model, and obtaining an initial reconstructed image feature; constructing an image statistical model; wherein the deep learning model comprises a deep convolutional neural network and a Non-Local model, and the deep convolutional neural network and the Non-Local model are fused; and obtaining a reconstructed image. The medical image reconstruction method and system for the radiology department have the advantages that the extracted image multi-scale features are high in accuracy and fineness, and error control is accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and more specifically, relates to a medical image reconstruction method and system for radiology. Background Art

[0002] Medical imaging is an important means of diagnosing diseases. Improving the imaging resolution of medical imaging equipment is extremely meaningful for disease diagnosis and treatment plan formulation. Moreover, medical images with higher resolution are of great help to the software in automatically segmenting and reconstructing the details of three-dimensional models. There are several major differences between medical imaging equipment and general imaging equipment. First, the radiation energy of the equipment is highly limited, reducing the time the patient is irradiated. Second, the imaging rate is fast. Shorten the patient's discomfort time. Third, the equipment is designed to examine and diagnose diseases. Many artifacts will be generated during the imaging process.

[0003] The existing deep learning convolutional neural network framework has the problem that it cannot effectively improve the clarity of the reconstructed image, and the image information is easily lost during the convolution operation. Summary of the invention

[0004] The purpose of the present invention is to provide a medical image reconstruction method and system for radiology, aiming to solve the technical problem of poor reconstruction effect existing in the prior art.

[0005] To achieve the above object, the technical solution adopted by the present invention is: to provide a medical image reconstruction method for radiology, comprising the following steps:

[0006] Input a medical image and pre-process the medical image;

[0007] The preprocessed medical image is input as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model includes: a deep convolutional neural network and a reconstruction module based on a virtual single pixel, and a reconstruction module based on a virtual single pixel is introduced into the deep convolutional neural network;

[0008] Start the deep learning model to obtain the initial reconstructed image features;

[0009] Constructing an image statistical model; wherein the deep learning model includes: a deep convolutional neural network and a Non-Local model, and the deep convolutional neural network and the Non-Local model are fused; the adaptive processing module of the Non-Local model is used to realize image depth reconstruction quality control;

[0010] Obtain a reconstructed image.

[0011] Preferably, the preprocessing of the medical image includes one or more of image sampling, compression encoding and denoising.

[0012] Preferably, the calculation formula of the reconstruction module based on the virtual single pixel includes:

[0013] I(x,y)=∫∫S(x′,y′)h(xx′,yy′)dx′dy′

[0014] M i (x,y)=I(x,y)P i (x,y),i=1,2,3....,m

[0015] AX=b

[0016] P i ′(x′,y′)=∫∫P i (x,y)h(xx′,yy′)dxdy

[0017] Where x, y represent the coordinates of two directions on the image plane; x′, y′ represent the coordinates of two directions on the sample plane; S represents the sample structure distribution; h represents the point spread function of the optical system; I represents the collected image; M i Represents the value relative to the i-th digital template P i Modulated image; m represents the number of digital templates; A∈R m×n represents the adjustment matrix (m P i ′, each template consists of n pixels); X∈R 1×n represents the structural sample distribution S of the solution; b∈R 1×m Represents a vector consisting of m integral values.

[0018] Preferably, the operation process of the Non-Local model includes:

[0019]

[0020] Among them, the dimension of the image is M×M, i∈M, j∈M; X1(i,j) and X2(i,j) are two pixel features that have no adjacent relationship in the image; Represents the feature mean of the complete image, and the image features are represented by a vector.

[0021] Preferably, the calculation process of the Non-Local model further includes:

[0022]

[0023] f(X1,X2)=f(X1(i,j),X2(i,j))

[0024]

[0025] Among them, G(X1(i,j),X2(i,j)) represents the spatial position relationship function of X1(i,j) and X2(i,j). f(X1(i,j),X2(i,j)) represents the feature similarity evaluation function of X1(i,j) and X2(i,j); C(X1(i,j),X2(i,j)) represents the control factor of the non-local mean operation; it is determined for the dot product operation of X1(i,j) and X2(i,j); Y(i,j) represents the image feature output result of the model; Represents the operation based on the l2 norm in the vector matrix; min represents the minimum value operation using the gradient descent method; They are respectively represented as the output parameters of the ideal image features learned and trained in the model for the two input features X1(i,j) and X2(i,j); represents the average features of the complete image; represents the activation function of the model; a represents the feature mapping weighted control factor of the image pixel feature X1(i,j); β represents the feature mapping weighted control factor of the image pixel feature X2(i,j); b represents the bias value of the basic processing unit of the model; Represents the Gaussian function represents the Gaussian function; Represents the output result of image feature classification.

[0026] Preferably, the operation process of the Non-Local model further includes constructing a loss function formula, and the expression of the loss function formula is:

[0027]

[0028] Preferably, the deep learning model includes noise control processing; and / or the image statistics model includes noise control processing.

[0029] A medical image reconstruction system for a radiology department, characterized by being used to implement the steps in a medical image reconstruction method for a radiology department as described in any one of the above items.

[0030] Preferably, the preprocessing module is used to input a medical image and preprocess the medical image;

[0031] A deep learning module, used for inputting the preprocessed medical image as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model comprises: a deep convolutional neural network and a reconstruction module based on a virtual single pixel, and the reconstruction module based on a virtual single pixel is introduced into the deep convolutional neural network;

[0032] An image acquisition module is used to start the deep learning model and obtain the initial reconstructed image features;

[0033] An image statistics module is used to construct an image statistics model; wherein the deep learning model includes: a deep convolutional neural network and a Non-Local model, and the deep convolutional neural network and the Non-Local model are fused; the adaptive processing module of the Non-Local model is used to realize image depth reconstruction quality control;

[0034] The reconstructed image acquisition module is used to obtain the reconstructed image.

[0035] The beneficial effect of the medical image reconstruction method and system for radiology provided by the present invention is that: compared with the prior art, the medical image reconstruction method and system for radiology provided by the present invention can extract multi-scale features of images with high accuracy and refinement, and can achieve excellent reconstruction effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A flowchart of a medical image reconstruction method system for radiology provided by an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of the structure of a medical image reconstruction system for a radiology department provided by an embodiment of the present invention.

[0039] In the figure: 1. Preprocessing module; 2. Deep learning module; 3. Image acquisition module; 4. Image statistics module; 5. Reconstructed image acquisition module. DETAILED DESCRIPTION

[0040] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0041] Please also read Figure 1 to Figure 2 Now, a medical image reconstruction method for radiology provided by the present invention is described. The medical image reconstruction method for radiology includes the following steps:

[0042] S1. Input a medical image and preprocess the medical image;

[0043] In this step, preprocessing includes: image sampling, compression encoding operation and denoising processing.

[0044] Specifically, the preprocessing process is to sequentially undergo image sampling, compression coding operation and denoising. More specifically, the compression coding operation preferably uses a wavelet decomposition method. The denoising process preferably uses a filter to denoise the compression coding and sampling results. More preferably, the power density function of the filter is:

[0045]

[0046] S2. Input the preprocessed medical image as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model includes: a deep convolutional neural network and a reconstruction module based on a virtual single pixel, and a reconstruction module based on a virtual single pixel is introduced into the deep convolutional neural network; the accuracy and refinement of image feature extraction can be improved.

[0047] The deep convolutional neural network extracts features from the encoded image, performs dimension reduction through multi-scale pooling, and performs feature classification. It can improve the multi-scale representation capability of image features. The image shape is reconstructed using a virtual single pixel-based image reconstruction method.

[0048] This step also includes using a filter to perform denoising. The use of the filter can achieve noise control during the processing.

[0049] The power density function of the filter is:

[0050]

[0051] In this step, a reconstruction module based on a virtual single pixel is used. This reconstruction module combines the introduced integral value equivalence relationship with the single pixel imaging algorithm and can restore the fine structure of the image by solving the linear matrix in the spatial domain.

[0052] Specifically, the image is post-modulated using a digital template, and an equivalent relationship is obtained, that is, the integral value of the post-modulated image is equal to the integral value of the product of the sample distribution and the convolution template.

[0053] The deconvolution problem is transformed into solving a linear matrix using the VSPI algorithm. In a linear matrix, given the convolution template and the corresponding integral value, the sample distribution can be obtained. Schematic diagram of the equivalent relationship of the integral value.

[0054] The calculation formula of the reconstruction module based on virtual single pixel includes:

[0055] I(x,y)=∫∫S(x′,y′)h(xx′,yy′)dx′dy′

[0056] Where x, y represent the coordinates of two directions on the image plane; x′, y′ represent the coordinates of two directions on the sample plane; S represents the sample structure distribution. h represents the point spread function of the optical system. I represents the collected image. I is the convolution result of S and h.

[0057] M i (x,y)=I(x,y)P i (x,y),i=1,2,3....,m

[0058] Among them, M i Represents the value relative to the i-th digital template P i Modulated image; m represents the number of digital templates.

[0059] Modulation image M i The integral value of is expressed as:

[0060]

[0061] Convolution template P i The formula for ′ is:

[0062] P i ′(x′,y′)=∫∫P i (x,y)h(xx′,yy′)dxdy

[0063] b i =∫∫P i ′(x′,y′)S(x′,y′)dx′dy′

[0064] Modulation image M i The integral value is equal to the sample structure distribution S and the convolution template P i The integral value of ′ is as follows:

[0065]

[0066] The formula for sample structure distribution S is:

[0067] AX=b

[0068] Among them, A∈R m×n represents the adjustment matrix (m P i ′, each template consists of n pixels); X∈R 1×n represents the structural sample distribution S of the solution; b∈R 1×m Represents a vector consisting of m integral values.

[0069] S3, start the deep learning model to obtain the initial reconstructed image features. Specifically, start the deep convolutional neural network to recognize the reconstructed initial image and obtain the initial reconstructed image features.

[0070] S4. Construct an image statistical model; wherein the deep learning model includes: a deep convolutional neural network and a Non-Local model. The fusion of the deep convolutional neural network and the Non-Local model can improve the autonomous learning and training capabilities of the key control parameters of image depth reconstruction. Among them, the adaptive processing module of the Non-Local model is used to achieve image depth reconstruction quality control. The Non-Local model is based on non-local average operation. In the feature mapping of image processing, it can implement weighted convolution operations of image feature spatial state and temporal state. The Non-Local model is used for accurate classification of image features in image reconstruction. Specifically, a forward feedback mechanism is used for non-local mean filtering, and by designing a combination module of image feature mean filtering and forward feedback, accurate classification and detection of image features are achieved. The implementation method of forward feedback is mainly to perform a one-dimensional convolution operation. A deep convolutional neural network is used for accurate extraction of image features.

[0071] This step also includes using a filter to perform denoising. The use of the filter can achieve noise control during the processing.

[0072] The power density function of the filter is:

[0073]

[0074] In this step, the operation process of the Non-Local model is:

[0075]

[0076] Among them, the dimension of the image is M×M, i∈M, j∈M; X1(i,j) and X2(i,j) are two pixel features that have no adjacent relationship in the image; Represents the feature mean of the complete image, and the image features are represented by a vector.

[0077]

[0078]

[0079] f(X1,X2)=f(X1(i,j),X2(i,j))

[0080]

[0081] Among them, G(X1(i,j),X2(i,j)) represents the spatial position relationship function of X1(i,j) and X2(i,j). f(X1(i,j),X2(i,j)) represents the feature similarity evaluation function of X1(i,j) and X2(i,j); C(X1(i,j),X2(i,j)) represents the control factor of the non-local mean operation. It is determined for the dot product operation of X1(i,j) and X2(i,j). Y(i,j) represents the image feature output result of the model. Indicates the operation based on the l2 norm in the vector matrix. min indicates the minimum value operation using the gradient descent method. They are respectively represented as the output parameters of the ideal image features learned and trained in the model for the two input features X1(i,j) and X2(i,j); represents the average feature of the complete image. represents the activation function of the model; a represents the feature mapping weighted control factor of the image pixel feature X1(i,j); β represents the feature mapping weighted control factor of the image pixel feature X2(i,j); b represents the bias value of the basic processing unit of the model; Represents the Gaussian function represents a Gaussian function. Represents the output result of image feature classification.

[0082] Loss function of the non-local model for:

[0083]

[0084] S5. Obtain a reconstructed image.

[0085] The present invention also provides a medical image reconstruction system for radiology department, please refer to Figure 1 to Figure 2, used to implement the steps in a medical image reconstruction method for radiology as described in any of the above. Specifically, a medical image reconstruction system for radiology includes a preprocessing module 1, a deep learning module 2, an image acquisition module 3, an image statistics module 4 and a reconstructed image acquisition module 5, wherein the preprocessing module 1 is used to input a medical image and preprocess the medical image; the deep learning module 2 is used to input the preprocessed medical image as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model includes: a deep convolutional neural network and a reconstruction module based on a virtual single pixel, and a reconstruction module based on a virtual single pixel is introduced into the deep convolutional neural network; the image acquisition module 3 is used to start the deep learning model and obtain the initial reconstructed image features; the image statistics module 4 is used to construct an image statistics model; wherein the deep learning model includes: a deep convolutional neural network and a Non-Local model, and the deep convolutional neural network and the Non-Local model are fused; the adaptive processing module of the Non-Local model is used to realize image deep reconstruction quality control; the reconstructed image acquisition module 5 is used to obtain a reconstructed image.

[0086] The present invention provides a medical image reconstruction method and system for radiology. Compared with the prior art, the edge feature map extracted by the method contains less noise and non-high-frequency information, the image edge clarity is high, and the deep reconstruction image has almost no difference in visual perception from the input initial image. The multi-scale features of the image extracted by the method and system are highly accurate and refined, and can achieve excellent reconstruction effects.

[0087] The present invention also provides an electronic device, comprising a bus, a transceiver, a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the transceiver, the memory and the processor are connected via the bus, and wherein the computer program, when executed by the processor, implements the steps in the above-mentioned medical image reconstruction method for the radiology department.

[0088] Compared with the prior art, the beneficial effects of the electronic device provided by the present invention are the same as the beneficial effects of the medical image reconstruction method for the radiology department described in the above technical solution, and will not be elaborated here.

[0089] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps in the above-mentioned medical image reconstruction method for the radiology department are implemented.

[0090] Compared with the prior art, the beneficial effects of a computer-readable storage medium provided by the present invention are the same as the beneficial effects of a medical image reconstruction method for radiology described in the above technical solution, and are not elaborated here.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A medical image reconstruction method for radiology, characterized in that: The following steps are involved: Input a medical image and pre-process the medical image; The preprocessed medical image is input as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model includes: a deep convolutional neural network and a reconstruction module based on a virtual single pixel, and a reconstruction module based on a virtual single pixel is introduced into the deep convolutional neural network; Start the deep learning model to obtain the initial reconstructed image features; Constructing an image statistical model; wherein the deep learning model includes: a deep convolutional neural network and a Non-Local model, and the deep convolutional neural network and the Non-Local model are fused; the adaptive processing module of the Non-Local model is used to realize image depth reconstruction quality control; Obtain a reconstructed image.

2. A medical image reconstruction method for radiology as claimed in claim 1, characterized in that: The preprocessing of the medical image includes one or more of image sampling, compression encoding and denoising.

3. A medical image reconstruction method for radiology as claimed in claim 1, characterized in that: The calculation formula of the reconstruction module based on virtual single pixel includes: I(x,y)=∫∫S(x′,y′)h(xx′,yy′)dx′dy′ M i (x,y)=I(x,y)P i (x,y),i=1,2,3....,m AX=b P i ′(x′,y′)=∫∫P i (x,y)h(xx′,yy′)dxdy Where x, y represent the coordinates of the two directions on the image plane; x′, y′ represent the coordinates of the two directions on the sample plane; S represents the sample structure distribution; h represents the point spread function of the optical system; I represents the collected image; M i Represents the value relative to the i-th digital template P i Modulated image; m represents the number of digital templates; A∈R m×n represents the adjustment matrix (m P i ′, each template consists of n pixels); X∈R 1×n represents the structural sample distribution S of the solution; b∈R 1×m Represents a vector consisting of m integral values.

4. A medical image reconstruction method for radiology as claimed in claim 3, characterized in that: The operation process of the Non-Local model includes: Among them, the dimension of the image is M×M, i∈M, j∈M; X1(i,j) and X2(i,j) are two pixel features that have no adjacent relationship in the image; Represents the feature mean of the complete image, and the image features are represented by a vector.

5. A medical image reconstruction method and system for radiology as claimed in claim 4, characterized in that: The operation process of the Non-Local model also includes: f(X1,X2)=f(X1(i,j),X2(i,j)) Among them, G(X1(i,j),X2(i,j)) represents the spatial position relationship function of X1(i,j) and X2(i,j). f(X1(i,j),X2(i,j)) represents the feature similarity evaluation function of X1(i,j) and X2(i,j); C(X1(i,j),X2(i,j)) represents the control factor of the non-local mean operation; it is determined for the dot product operation of X1(i,j) and X2(i,j); Y(i,j) represents the image feature output result of the model; Represents the operation based on the l2 norm in the vector matrix; min represents the minimum value operation using the gradient descent method; They are respectively represented as the output parameters of the ideal image features learned and trained in the model for the two input features X1(i,j) and X2(i,j); represents the average features of the complete image; represents the activation function of the model; a represents the feature mapping weighted control factor of the image pixel feature X1(i,j); β represents the feature mapping weighted control factor of the image pixel feature X2(i,j); b represents the bias value of the basic processing unit of the model; Represents the Gaussian function represents the Gaussian function; Represents the output result of image feature classification.

6. A medical image reconstruction method for radiology as claimed in claim 5, characterized in that: The operation process of the Non-Local model also includes constructing a loss function formula, and the expression of the loss function formula is:

7. A medical image reconstruction method for radiology as claimed in claim 1, characterized in that: The deep learning model includes noise control processing; and / or the image statistics model includes noise control processing.

8. A medical image reconstruction system for radiology, characterized in that: Used to implement the steps in a medical image reconstruction method for radiology as described in any one of claims 1-7.

9. A medical image reconstruction system for radiology department as claimed in claim 8, characterized in that: include: A preprocessing module, used for inputting medical images and preprocessing the medical images; A deep learning module is used to input the preprocessed medical image as a sample into a deep learning model for training to obtain an image recognition model; wherein the deep learning model includes: a deep convolutional neural network and a reconstruction module based on a virtual single pixel, and a reconstruction module based on a virtual single pixel is introduced into the deep convolutional neural network; an image acquisition module is used to start the deep learning model and obtain initial reconstructed image features; an image statistics module is used to construct an image statistics model; wherein the deep learning model includes: a deep convolutional neural network and a Non-Local model, and the deep convolutional neural network and the Non-Local model are fused; an adaptive processing module of the Non-Local model is used to achieve image deep reconstruction quality control; a reconstructed image acquisition module is used to obtain a reconstructed image.

Images