Method for synthesizing CTPA blood vessel image based on plain-scan CT image

Abstract

The invention discloses a method for synthesizing a CTPA blood vessel image based on a plain-scan CT image, and the method comprises the steps: carrying out the registration and preprocessing of the plain-scan CT image and the CTPA image, taking three continuous plain-scan CT slice images at a selected depth as samples, taking three continuous CTPA slice images at corresponding positions in the CTPA image as labels, and carrying out the recognition of the CTPA blood vessel image. Determining a blood vessel tissue area of the label and the edge of the enhanced blood vessel; a U-Net model is constructed; constructing a loss function loss; training the U-Net model based on the sample; and sequentially taking three plain-scan CT slice images from a plain-scan CT image to be predicted according to the depth, and inputting the three plain-scan CT slice images into the trained U-Net model to obtain a predicted CTPA image. According to the method, the image effect similar to CTPA is simulated or reconstructed through the non-enhanced plain-scan CT image. This method has reduced patient radiation exposure, shortened examination time, and possibly reduced examination costs.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and computer tomography (CT) imaging, and in particular to a method for synthesizing a CTPA vascular image based on a plain scan CT image, which is suitable for efficiently synthesizing an image with a CTPA (CT Pulmonary Angiography) effect from a plain scan CT image. Background Art

[0002] CTPA, or Computed Tomography Pulmonary Angiography, is a medical imaging technique that is specifically used to evaluate the vascular conditions of the pulmonary arteries and their branches, and is particularly valuable in diagnosing pulmonary embolism. However, traditional CTPA examinations require injection of contrast agents, which increases the patient's radiation exposure and examination complexity, and may also cause adverse reactions such as contrast agent allergies. In addition, injection of contrast agents also increases the cost and time of the examination.

[0003] Plain CT, also known as non-enhanced CT or ordinary CT scan, is a commonly used medical imaging technique. It is widely used to examine various parts of the body, mainly for the detection of space-occupying lesions (such as tumors, cysts, hematomas, etc.), vascular diseases, etc. Plain CT does not require the injection of contrast agents (such as iodine), so it is relatively simple and quick, and patients do not need to bear the risks of allergies caused by contrast agents.

[0004] Therefore, researchers began to explore methods to synthesize CTPA images based on plain CT images, in order to improve the safety and efficiency of the examination without affecting the accuracy of diagnosis.

[0005] The method of synthesizing CTPA (CT pulmonary angiography) images based on plain CT images is an innovative technology in the field of medical image processing, which aims to simulate or reconstruct image effects similar to CTPA through non-enhanced plain CT images. This method has the advantages of reducing patient radiation exposure, shortening examination time, and possibly reducing examination costs. Summary of the invention

[0006] The purpose of the present invention is to provide a method for synthesizing CTPA images based on plain scan CT images to address the above problems in the prior art, aiming to simulate or reconstruct an image effect similar to CTPA through non-enhanced plain scan CT images, thereby obtaining image quality similar to CTPA without the need to inject contrast agents. This method has the advantages of reducing patient radiation exposure, shortening examination time, and reducing examination costs.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] A method for synthesizing a CTPA vascular image based on a plain scan CT image comprises the following steps:

[0009] Step 1: Register the plain scan CT image and CTPA image;

[0010] Step 2: Preprocess the registered plain scan CT image and CTPA image, take three consecutive plain scan CT slice images at the selected depth as samples, take three consecutive CTPA slice images at the corresponding position in the CTPA image as labels, and determine the vascular tissue area of ​​the label and the edge of the enhanced blood vessel;

[0011] Step 3: Build the U-Net model;

[0012] Step 4: Construct loss function loss;

[0013] Step 5: Train the U-Net model based on the samples;

[0014] Step 6: Take three plain scan CT slice images to be predicted in sequence according to the depth and input them into the trained U-Net model to obtain the predicted CTPA image.

[0015] As described above, the registration of the plain scan CT image and the CTPA image includes the following steps:

[0016] Step 1.1: Segment the plain CT image and CTPA image of the pulmonary blood vessels of the same experimental subject using a threshold segmentation algorithm to obtain a plain CT segmentation image and a CTPA segmentation image;

[0017] Step 1.2: Register the plain scan CT segmentation image with the CTPA segmentation image to obtain the corresponding deformation parameters;

[0018] Step 1.3: Apply the deformation parameters to the plain scan CT segmentation image to obtain the registered plain scan CT segmentation image, and reconstruct each registered plain scan CT segmentation image into a registered plain scan CT image to achieve registration of the plain scan CT image and the CTPA image.

[0019] As mentioned above, preprocessing includes image normalization and data augmentation.

[0020] As mentioned above, the loss function loss is based on the following formula:

[0021] loss = mean(w*|y pred -y|)

[0022] w=I+49*y roi +50*y edge

[0023] Among them, yroi represents the vascular tissue area, y edge represents the edge of the enhanced blood vessel, w represents the weight matrix, y is the label, and y pred represents the predicted label generated by the U-Net model, mean represents the mean calculated on the matrix, and I is the identity matrix.

[0024] As mentioned above, the U-Net model includes an encoder and a decoder. The encoder includes four encoding layers and an intermediate encoding layer, and the decoder includes four decoding layers.

[0025] After the sample is input into the U-Net model, it will pass through each encoding layer of the encoder in sequence for feature extraction, then pass through the intermediate encoding layer to further calculate the features, and finally pass through each decoding layer of the decoder in sequence to decode the features.

[0026] As mentioned above, the sample is input into the first encoding layer, the output features of the first encoding layer are input into the fourth decoding layer and the second encoding layer, the output features of the second encoding layer are input into the third decoding layer and the third encoding layer, the output features of the third encoding layer are input into the second decoding layer and the fourth encoding layer, the output features of the fourth encoding layer are input into the first decoding layer and the middle encoding layer, and the output of the middle encoding layer is input into the first decoding layer.

[0027] As mentioned above, both the encoding layer and the decoding layer include three groups of convolutional networks and attention mechanism networks. Each group of convolutional networks includes two layers of convolutional networks.

[0028] The input feature is convolved twice through the first group of convolutional networks, and the input feature is added to the first convolution output feature after convolution of the first group of convolutional networks to obtain the first added feature. The first added feature is input into the second group of convolutional networks for calculation to obtain the second convolution output feature. The first added feature and the second convolution output feature are added to obtain the second added feature. The second added feature is input as the input feature into the attention mechanism network to calculate the self-attention matrix Q, matrix K, and matrix V, and generate the attention result. Finally, the attention result is input into the third group of convolutional networks to obtain the third convolution output feature. The third convolution output feature is added to the attention result to obtain the final output feature.

[0029] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and wherein the processor executes the steps of the method.

[0030] A computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

[0031] A computer program product comprises a computer program, which implements the steps of the method when executed by a processor.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. No need to inject contrast agent, reducing the risk of allergies and renal damage in patients.

[0034] 2. Reduced medical costs and alleviated the financial burden on patients.

[0035] 3. Use non-enhanced plain scan CT images to simulate or reconstruct images similar to CTPA. This method can reduce patient radiation exposure, shorten examination time, and potentially reduce examination costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the network structure diagram of the CTPA task generated by plain scan CT.

[0037] Figure 2 It is a network diagram of convolution and attention mechanism. DETAILED DESCRIPTION

[0038] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0039] Embodiment 1:

[0040] A method for synthesizing a CTPA vascular image based on a plain scan CT image comprises the following steps:

[0041] Step 1: The plain scan CT image and CTPA image of the pulmonary vessels are registered based on the Symmetric Normalization registration and threshold algorithm.

[0042] For the plain scan CT images and CTPA images of the same experimental object, the Symmetric Normalization (SyN) algorithm is first used to align the plain scan CT images and CTPA images. Symmetric Normalization is a registration technology based on symmetric normalization. It uses a nonlinear deformation field to describe the deformation between images and can capture very complex structural changes and deformations. The core idea is to regard the registration between two images as calculating a deformation field, so that one image (moving image) can be deformed to the coordinate system of another image (fixed image) through the deformation field.

[0043] As a preferred solution, in order to improve the accuracy of registration, step 1 includes the following steps:

[0044] Step 1.1: Use the threshold segmentation algorithm to segment the plain scan CT image and CTPA image of the pulmonary vessels of the same experimental subject. During the threshold segmentation, set the hu value to -500, and extract the areas with hu values ​​greater than -500 in the plain scan CT image and CTPA image, that is, the body surface, all body tissues, and pulmonary vessels after removing the background. Obtain the plain scan CT segmentation image and CTPA segmentation image.

[0045] Step 1.2: Using the Symmetric Normalization algorithm, the plain scan CT segmentation image is registered with the CTPA segmentation image to obtain the corresponding deformation parameters.

[0046] Step 1.3: Apply the deformation parameters to the plain scan CT segmentation image to obtain the registered plain scan CT segmentation image, and reconstruct each registered plain scan CT segmentation image into a registered plain scan CT image to achieve registration of the plain scan CT image and the CTPA image.

[0047] Step 2: Preprocess the registered plain scan CT image and CTPA image. The dimension of the plain scan CT image is depth × height × width D × H × W. Each depth is a plain scan CT slice image. Select any plain scan CT slice image in the depth dimension, and take out the plain scan CT slice images of the previous depth and the next depth at the same time as three consecutive plain scan CT slice images at the selected depth. The dimension of the three plain scan CT slice images is 3 × H × W. The three consecutive plain scan CT slice images at the selected depth are used as samples and input into the network. The three consecutive CTPA slice images at the corresponding position in the CTPA image are used as labels, and the vascular tissue area of ​​the label and the edge of the enhanced blood vessel are determined. This process ensures that the input and label are completely aligned in space by matching the slices at the same position, so that the learning of the U-Net model is more accurate. This selection method can effectively utilize the context information of the image and improve the prediction accuracy of the U-Net model.

[0048] The preprocessing in step 2 includes:

[0049] Step 2.1: Image normalization:

[0050] First, the window width and window level of the registered plain CT images and CTPA images of each experimental subject were normalized. The window width was set to 600 and the window level was set to 0. Through this process, the pixel values ​​of the registered plain CT images and CTPA images were scaled to between 0 and 1. After normalization, the effective information in the registered plain CT images and CTPA images can be concentrated in the numerical range that is easy for the network to process, which improves the convergence speed and prediction accuracy of the model.

[0051] Step 2.2: Data Augmentation:

[0052] When processing registered plain scan CT images, data enhancement methods are used to increase data diversity and prevent model overfitting. Specific enhancement methods include:

[0053] Noise addition: Random noise is added to the registered plain scan CT images to simulate image noise under different scanning conditions and increase the robustness of the U-Net model.

[0054] Gaussian smoothing: Gaussian filtering is performed on the registered plain scan CT images to reduce noise and smooth the images, so that the U-Net model can better capture large-scale features.

[0055] Rotation: Randomly rotate the registered plain scan CT images and CTPA images by a certain angle to enhance the U-Net model’s ability to learn directional invariance.

[0056] Translation: Randomly translate the registered plain scan CT images and CTPA images to simulate position changes and enhance the spatial invariance of the U-Net model.

[0057] Scaling: The registered plain scan CT images are first scaled down and then scaled up to simulate images of different resolutions and enhance the adaptability of the U-Net model to scale changes.

[0058] Step 3: Build a U-Net model. The U-Net model is a deep learning network widely used in medical image segmentation. Its structure includes an encoder, a decoder, and a jump connection. Input the sample into the U-Net model, and the U-Net model outputs the predicted label and compares it with the corresponding label.

[0059] Step 4: Construct a loss function. In order to calculate the difference between the predicted CTPA image obtained by the U-Net model and the corresponding label, this application uses the L1 loss function (i.e., mean absolute error) to measure the prediction accuracy of the U-Net model. Since the present invention mainly focuses on the pulmonary vascular area, that is, the target area, and the pulmonary vascular area accounts for a small proportion of the entire CT image, the target area is weighted in the loss function loss, and the formula is as follows:

[0060] loss = mean(w*|y pred -y|)

[0061] w=I+49*y roi +50*y edge

[0062] Among them, y roi represents the vascular tissue area, y edge represents the edge of the enhanced blood vessel, w represents the weight matrix, y is the label, and y pred represents the predicted label generated by the U-Net model, mean represents the mean calculated for the matrix, I is the unit matrix, and vascular tissue area y roi and enhance the edge of the blood vessels edge It is extracted based on the label, and its dimension is the same as the label. The vascular tissue area y roi and enhance the edge of the blood vessels edge All are target areas, and the value of non-target areas is 0.

[0063] Step 5: Train the U-Net model based on the samples;

[0064] Step 6: Take three plain scan CT slice images to be predicted in sequence according to the depth and input them into the trained U-Net model to obtain the predicted CTPA image.

[0065] The U-Net model of this scheme includes an encoder and a decoder (as shown in Figure 1). The encoder includes four encoding layers (in order from top to bottom) and an intermediate encoding layer, and the decoder includes four decoding layers (in order from bottom to top). After the sample is input into the U-Net model, it will pass through each encoding layer of the encoder in turn for feature extraction, and then further calculate the features through the intermediate encoding layer, and finally pass through each decoding layer of the decoder in turn to decode the features. It should be noted that in the decoding process, the input of the decoding layer is the output feature of the corresponding encoding layer, that is, the jump connection (for example, the sample is input into the first encoding layer, the output feature of the first encoding layer will be input into the fourth decoding layer and the second encoding layer, the output feature of the second encoding layer will be input into the third decoding layer and the third encoding layer, the output feature of the third encoding layer will be input into the second decoding layer and the fourth encoding layer, the output feature of the fourth encoding layer will be input into the first decoding layer and the intermediate encoding layer, and the output of the intermediate encoding layer is input into the first decoding layer) and the features of the previous decoding layer (for example, the output feature of the first decoding layer will be input into the second decoding layer).

[0066] The coding layer and decoding layer of this scheme have the same structure (such as Figure 2As shown), that is, they are all composed of attention mechanism networks. Both the encoding layer and the decoding layer include three groups of convolutional networks and attention mechanism networks, and each group of convolutional networks includes two layers of convolutional networks. Take the encoding layer of a certain one as an example. First, the input features are convolved twice through the first group of convolutional networks, and the input features are added to the first convolutional output features after the convolution of the first group of convolutional networks to obtain the first added features. Then, the first added features are input into the second group of convolutional networks again for calculation to obtain the second convolutional output features, and the first added features and the second convolutional output features are added again to obtain the second added features. Subsequently, based on the second added features as input features, the attention mechanism network is input to calculate the self-attention matrix Q, matrix K, and matrix V, and the attention results are generated. Finally, the attention results are input into the third group of convolutional networks for further processing to obtain the third convolutional output features, and the third convolutional output features are added to the attention results. The final output features are obtained.

[0067] In the attention mechanism network, the input features are used to obtain the matrix Q, matrix K, and matrix V using the convolutional network. Then the dot product between the matrix Q and the matrix K is calculated to obtain the attention weight matrix. Finally, the matrix V is weighted and summed using the attention weight matrix to obtain the final output attention result.

[0068] Suppose the size of the input feature currently input to the attention mechanism network is b×c×h×w, where b is the number of features input to the attention mechanism network at one time, c is the number of channels, h is the feature height, and w is the feature width. Through three convolutional networks, three matrices Q, K, and V with sizes of b×c1×h×w are obtained, where c1 is the number of channels. Resize matrices Q, K, and V, that is, split the c1 dimension into n head ×c2,n head is the attention parameter, that is, c1 = n head *c2, and merge the h and w dimensions to get hw, so the final size of the matrix Q, matrix K, and matrix V are all b×n head ×c2×hw. Here, the h and w dimensions are merged and the c1 dimension is split, just to facilitate the calculation of multiple attentions. This can be simply understood as the shape transformation of the matrix. Then matrix Q and matrix K are matrix multiplied, that is, Q T K gets the attention value A, whose size is b×n head ×hw×hw, here is the result of matrix multiplication. It represents the relationship between hw pixels in each two-dimensional image, thus forming a matrix of hw*hw, and normalizing the attention value. Finally, the attention value A and the matrix V are matrix multiplied, that is, AV T , whose size is b×n head×hw×c2, transform its size, that is, merge the c2 dimension into n head dimension, we get c1=n head *c2 dimension, and split the hw dimension into two dimensions of h×w, so the final output is b×c1×h×w.

[0069] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0070] Embodiment 2:

[0071] In this embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in each method embodiment of Embodiment 1 when executing the computer program.

[0072] Embodiment 3:

[0073] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in each method embodiment of Embodiment 1 are implemented.

[0074] Embodiment 4:

[0075] In this embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in each method embodiment of Embodiment 1 are implemented.

[0076] It should be noted that the embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the described embodiments or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A method for synthesizing CTPA vascular images based on plain scan CT images, characterized in that: The following steps are involved: Step 1: Register the plain scan CT image and CTPA image; Step 2: Preprocess the registered plain scan CT image and CTPA image, take three consecutive plain scan CT slice images at the selected depth as samples, take three consecutive CTPA slice images at the corresponding position in the CTPA image as labels, and determine the vascular tissue area of ​​the label and the edge of the enhanced blood vessel; Step 3: Build the U-Net model; Step 4: Construct loss function loss; Step 5: Train the U-Net model based on the samples; Step 6: Take three plain scan CT slice images to be predicted in sequence according to the depth and input them into the trained U-Net model to obtain the predicted CTPA image.

2. According to claim 1, a method for synthesizing CTPA vascular images based on plain scan CT images is characterized in that: The registration of the plain scan CT image and the CTPA image comprises the following steps: Step 1.1: Segment the plain scan CT image and CTPA image of the pulmonary blood vessels of the same experimental subject using a threshold segmentation algorithm to obtain a plain scan CT segmentation image and a CTPA segmentation image; Step 1.2: Register the plain scan CT segmentation image with the CTPA segmentation image to obtain the corresponding deformation parameters; Step 1.3: Apply the deformation parameters to the plain scan CT segmentation image to obtain the registered plain scan CT segmentation image, and reconstruct each registered plain scan CT segmentation image into a registered plain scan CT image to achieve registration of the plain scan CT image and the CTPA image.

3. The method for synthesizing CTPA vascular images based on plain scan CT images according to claim 1, characterized in that: The preprocessing includes image normalization and data enhancement.

4. The method for synthesizing CTPA vascular images based on plain scan CT images according to claim 1, characterized in that: The loss function loss is based on the following formula: loss=mean(w*|y pred -y|) w=I+49*y roi +50*y edge Among them, y roi represents the vascular tissue area, y edge represents the edge of the enhanced blood vessel, w represents the weight matrix, y is the label, and y pred Represents the predicted label generated by the U-Net model, mean represents the mean calculated on the matrix, and I is the identity matrix.

5. The method for synthesizing CTPA vascular images based on plain scan CT images according to claim 1, characterized in that: The U-Net model includes an encoder and a decoder. The encoder includes four encoding layers and an intermediate encoding layer, and the decoder includes four decoding layers. After the sample is input into the U-Net model, it will pass through each encoding layer of the encoder in sequence for feature extraction, then pass through the intermediate encoding layer to further calculate the features, and finally pass through each decoding layer of the decoder in sequence to decode the features.

6. The method for synthesizing CTPA vascular images based on plain scan CT images according to claim 5, characterized in that: The sample is input into the first coding layer, the output features of the first coding layer are input into the fourth decoding layer and the second coding layer, the output features of the second coding layer are input into the third decoding layer and the third coding layer, the output features of the third coding layer are input into the second decoding layer and the fourth coding layer, the output features of the fourth coding layer are input into the first decoding layer and the middle coding layer, and the output of the middle coding layer is input into the first decoding layer.

7. The method for synthesizing CTPA vascular images based on plain scan CT images according to claim 6, characterized in that: The encoding layer and the decoding layer each include three groups of convolutional networks and attention mechanism networks, each group of convolutional networks includes two layers of convolutional networks. The input feature is convolved twice through the first group of convolutional networks, and the input feature is added to the first convolution output feature after convolution of the first group of convolutional networks to obtain the first added feature. The first added feature is input into the second group of convolutional networks for calculation to obtain the second convolution output feature. The first added feature and the second convolution output feature are added to obtain the second added feature. The second added feature is input as the input feature into the attention mechanism network to calculate the self-attention matrix Q, matrix K, and matrix V, and generate the attention result. Finally, the attention result is input into the third group of convolutional networks to obtain the third convolution output feature. The third convolution output feature is added to the attention result to obtain the final output feature.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. 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 of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images