Medical image lesion three-dimensional stripping method

Abstract

The invention relates to a lesion three-dimensional stripping method of a medical image. The method comprises the following steps: acquiring an ovary ultrasonic medical image of a patient in each direction, preprocessing the ovary ultrasonic medical image and inputting the preprocessed image into a three-dimensional image processing system; performing three-dimensional reconstruction on the ovary ultrasonic medical images in each direction to obtain three-dimensional images of the ovary and adjacent tissues; slicing processing is carried out along a certain direction of the three-dimensional image to obtain a plurality of first slice images, image feature extraction is carried out, features matched with preset lesion feature items are obtained through feature comparison and marked, lesion areas are reserved, and areas outside the lesion areas are deleted; and recombining to construct a first three-dimensional image of the focus, identifying the first three-dimensional image of the focus, when the first three-dimensional image belongs to a specified category or the size of the first three-dimensional image is greater than a preset threshold value, acquiring an ovary CT medical image, performing three-dimensional reconstruction of the focus again, and acquiring and outputting a second three-dimensional image of the focus and the category and size of the focus. The three-dimensional construction of the ovary lesion can accurately assist a doctor in accurately diagnosing the disease of a patient.

Description

Technical Field

[0001] The present invention belongs to the technical field of image data processing, and particularly relates to a method for three-dimensional stereoscopic peeling of lesions in medical images. Background Art

[0002] Since the ovary is intraperitoneal in position, and the blood flow inside the ovary is very rich and it is closely connected to multiple organs, in the medical images of the ovary, the ovarian lesions will be affected by other adjacent tissues, making it difficult for doctors to analyze the lesions, and resulting in a relatively low accuracy of the diagnosis results of ovarian lesions for patients by doctors.

[0003] Therefore, with the development of image processing technology, before doctors diagnose the condition based on the medical images of various parts of a patient, generally, an image recognition system will first perform image recognition on the medical images. In the prior art, the medical images are input into the image recognition system. The image recognition system has corresponding algorithms preset inside and various corresponding image data of lesions built in. When the medical images are input into the image recognition system, each region in the input images will be compared with the preset lesion image data to determine the type and size of the lesions, thereby assisting doctors in diagnosing the patient's condition.

[0004] However, due to the special position of the ovary, which is located in the ovarian fossa on the lateral pelvic wall and is blocked by adjacent other organs and has rich blood flow, it has a certain impact on the imaging equipment for taking medical images of the ovary. Therefore, it is necessary to improve the image processing process in the prior art to highlight the ovarian lesions with the existing medical images of the ovary and assist doctors in accurately diagnosing the patient's condition. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for three-dimensional stereoscopic peeling of lesions in medical images to improve the image processing process in the prior art, so as to highlight the ovarian lesions with the existing medical images of the ovary and assist doctors in accurately diagnosing the patient's condition.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is as follows: A method for three-dimensional stereoscopic peeling of lesions in medical images includes the following steps: S1: Obtain ovarian ultrasound medical images of a patient in all directions through an ultrasound device, perform image preprocessing on the ovarian ultrasound medical images, and input the preprocessed ovarian ultrasound medical images into a three-dimensional image processing system; S2: The three-dimensional image processing system performs three-dimensional reconstruction based on the ovarian ultrasound medical images in all directions to obtain three-dimensional images of the ovary and adjacent tissues in the medical images; S3: Perform image slicing on the constructed three-dimensional image of the ovary and adjacent tissues at a specified thickness and along a certain direction of the three-dimensional image to obtain multiple first slice images with the specified thickness; S4: Respectively perform image feature extraction on the multiple first slice images, compare the extracted image features with the preset lesion image features, obtain the features that match the preset lesion feature items among them, and mark the lesion areas in each first slice image based on the features that match the preset lesion feature items, retain the lesion areas, and delete the areas outside the lesion areas; S5: Recombine the retained areas of each first slice image in the original slice order to construct the first three-dimensional image of the lesion, identify the lesion category and size of the first three-dimensional image of the lesion. When it belongs to the specified category or its size is greater than the preset threshold, execute step S6, otherwise output the lesion three-dimensional data including the lesion category size; S6: Obtain the ovarian CT medical images of the patient in each direction through a CT device, and after preprocessing the ovarian CT medical images, repeat steps S2 - S5 to obtain the second three-dimensional image of the lesion, as well as the lesion category and size and output them. The doctor aids in diagnosing the patient's condition based on the output data.

[0007] Preferably, the specific process of the image preprocessing in step S1 is as follows: S11: After performing image denoising processing on the ovarian ultrasound medical images in each direction, perform image scaling processing on the ovarian ultrasound medical images in each direction to scale the ovarian ultrasound medical images in each direction to a matching size; S12: Respectively perform image cutting on the ovarian ultrasound medical images in each direction after the scaling processing, and obtain multiple image blocks for each ovarian ultrasound medical image; S13: Remap the pixel values of each pixel point inside each image block according to the specified rule, calculate the gray values of each pixel point in each image block, and count the number of pixel points exceeding the preset threshold in each image block, and evenly disperse the part exceeding the threshold to the specified gray value.

[0008] Preferably, the specific process of the image scaling processing in step S11 is as follows: S111: Create an image coordinate system, and obtain the coordinates of the target pixel point in the ovarian ultrasound medical image under the image coordinate system M ( x , y ); S112: Obtain the coordinates of 4 pixel points within the M neighborhood of the coordinates of the target pixel point, Q 11 ( x 1, y 1), Q 12( x 1, y 2), Q 21 ( x 2, y 1 ) , Q 22 ( x 2, y 2); S113: Calculate the x coordinates of the intermediate pixel points R1 and R2: R1 = x 1 + ( x - x 1) / ( x 2 - x 1) * ( x 2 - x ), R2 = x 1 + ( x - x 1) / ( x 2 - x 1) * ( y 2 - x ); S114: Perform interpolation in the y direction: For R1, calculate f (R1, y ) = (Q11 * ( y 2 - y ) + Q21 * ( y - y 1)) / ( y 2 - y 1); For R2, calculate f (R2, y ) = (Q12 * ( y 2 - y ) + Q22 * ( y - y 1)) / ( y 2 - y 1); S115: Calculate the final target point pixel value f ( x , y ) = (f(R1, y ) * ( y 2 - y ) + f(R2, y ) * ( y - y 1)) / ( y 2 - y 1), f where f is the interpolation function.

[0009] Preferably, the specific process of three-dimensional reconstruction based on the ovarian ultrasound medical images in each direction in step S2 is as follows: S21: Based on the gray values of each pixel in the ovarian ultrasound medical images obtained in step S12, transform the gray value of each pixel according to a specified rule to obtain the depth value of each pixel; S22: Generate a set of data points representing the 3D shape based on the depth value of each pixel. The information of each data point in the set of data points includes the coordinates in three-dimensional space; S23: Simplify the set of data points to obtain a mesh composed of a specified shape, and map each ovarian ultrasound medical image to the mesh surface according to the correspondence between the position of the pixel points and the mesh, so as to realize the three-dimensional image of the tissue in the ovarian ultrasound medical image.

[0010] Preferably, the specific process of extracting image features from the multiple first slice images in step S3 is as follows: S31: Create a convolutional neural network model for extracting image features, train the convolutional neural network model with the labeled ovarian ultrasound medical images, and copy the multiple first slice images into multiple groups. The number of groups of slice images is the same as the number of categories of image features to be extracted; S32: The convolutional neural network model is set with multiple input channels, and the number of input channels matches the number of groups of slices. Input different groups of slice images into different input channels of the convolutional neural network model respectively according to the slice order; S33: Different input channels correspond to different combinations of convolutional layers and pooling layers. Set sliding windows of different sizes in each convolutional layer. Each sliding window is set as a matrix of a specified size. The matrix of each convolutional layer slides on each slice image. Different convolutional layers extract different local feature maps of the image through different dot product operations. The pooling layer performs downsampling processing on the output of the convolutional layer, and calculates the maximum value as the output of the pooling layer through the window set in the pooling layer for the output of the convolutional layer; S34: Set a fully connected layer after the combination layer of the convolutional layer and the pooling layer to perform feature fusion on the outputs of the combination layers of the convolutional layer and the pooling layer in different channels, and transmit them to the output layer for output.

[0011] Preferably, the specific process of the fully connected layer performing feature fusion on the outputs of the combination layers of the convolutional layer and the pooling layer in different channels in step S34 is as follows: S341: The fully connected layer converts the local features extracted from different channels into local feature vectors; S342: Calculate a final global feature vector according to a preset rule for the feature vectors corresponding to the local features extracted from different channels; S343: Map the global feature vector to probabilities corresponding to features of different categories through a preset activation function.

[0012] Preferably, in step S4, the specific process of comparing the extracted image features with the preset lesion image features to obtain the features that match the preset lesion feature items is as follows: S41: Convert the preset lesion image features into corresponding feature vectors, and match the global feature vector in step S343 with the feature vectors corresponding to the conversion of the preset lesion image features; S42: Verify the matching result based on the probabilities corresponding to features of different categories, obtain the lesion features in the patch image, and obtain the region corresponding to the lesion features in the slice image.

[0013] Preferably, when there are multiple types of lesions shown by the lesion features in the slice image obtained in step S42, color them with different colors in the corresponding regions.

[0014] Preferably, when recombining according to the original slice order based on the regions retained in each first slice image in step S5, perform different three-dimensional reconstructions on different lesions according to different colors.

[0015] The beneficial effects of the present invention include: The method for three-dimensional stereoscopic peeling of lesions in medical images provided by the present invention obtains ovarian ultrasound medical images of a patient in all directions, preprocesses them and inputs them into a three-dimensional image processing system; performs three-dimensional reconstruction on the ovarian ultrasound medical images in each direction to obtain three-dimensional images of the ovary and adjacent tissues; performs slicing processing along a certain direction of the three-dimensional image to obtain multiple first slice images for image feature extraction, obtains and marks the features that match the preset lesion feature items through feature comparison, retains the lesion area, and deletes the area outside the lesion area; recombines to construct the first three-dimensional image of the lesion, identifies the first three-dimensional image of the lesion, and when it belongs to a specified category or its size is greater than a preset threshold, obtains the ovarian CT medical image and performs three-dimensional reconstruction of the lesion again to obtain the second three-dimensional image of the lesion, as well as the lesion category and size and outputs them. The three-dimensional construction of ovarian lesions can accurately assist doctors in the accurate diagnosis of patients' diseases.

[0016] First, by constructing a three-dimensional image of the ovary and adjacent tissues based on ovarian ultrasonic medical imaging, when there is a lesion in the image, a combined three-dimensional image between the lesion and the ovarian tissue can be obtained, but because the lesion and the tissue are interlaced, even if a three-dimensional image is constructed, it is impossible to clearly obtain the relevant data of the lesion. Therefore, the present invention slices the constructed combined three-dimensional image between the lesion and the ovarian tissue, and performs feature extraction and other processing based on the sliced ​​image, removes the image area except the lesion area, retains only the lesion area, and combines the sliced ​​images based on each lesion area to obtain the three-dimensional data of the lesion, providing a data basis for subsequent lesion identification and size calculation, and improving the efficiency and accuracy of auxiliary diagnosis.

[0017] Secondly, for some of the critical and severe lesions, CT images are acquired through CT imaging equipment, and then three-dimensional reconstruction of the lesions is performed based on the CT images. This avoids the limitations of lesion display in ovarian ultrasound medical images acquired by ultrasound equipment, and can further improve the accuracy of the three-dimensional image of the lesion, and further improve the efficiency and accuracy of auxiliary diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the architecture of the neural network model of the present invention.

[0019] Figure 2 It is a schematic flow chart of the method for three-dimensional lesion peeling of medical images of the present invention. DETAILED DESCRIPTION

[0020] The following is combined with Figures 1 - 2 The present invention is further described in detail: Example 1 See attached Figure 2 As shown, a method for three-dimensional lesion peeling of a medical image comprises the following steps: S1: Obtain ovarian ultrasound medical images of the patient in all directions through ultrasound equipment, perform image preprocessing on the ovarian ultrasound medical images, and input the preprocessed ovarian ultrasound medical images into the three-dimensional image processing system. The image preprocessing process includes image denoising to prevent the noise data from affecting the subsequent three-dimensional peeling of the lesion to achieve the three-dimensional reconstruction of the lesion. The subsequent image scaling process makes all images in the same image size, providing a standard data basis for the subsequent three-dimensional reconstruction of the ovary and adjacent tissues, and avoiding the problem of inability to match in the three-dimensional reconstruction process.

[0021] S2: The three-dimensional image processing system performs three-dimensional reconstruction based on the ovarian ultrasound medical images in each direction to obtain three-dimensional images of the ovaries and adjacent tissues within the medical images. Since different tissues and the lesions present in the tissues will exhibit different states in the images, and due to the mutual influence of the adjacent tissues, the ovarian ultrasound medical image in a certain direction is relatively one-sided. Therefore, it is necessary to obtain ovarian ultrasound medical images in each direction to make the subsequent three-dimensional imaging results more accurate.

[0022] S3: Perform image slicing processing on the constructed three-dimensional images of the ovaries and adjacent tissues according to a specified thickness and along a certain direction of the three-dimensional images to obtain multiple first slice images with the specified thickness. During the image slicing process, it is necessary to preset the slicing direction in advance. The slicing direction is set to the direction perpendicular to a certain axis direction of the tissue, that is, set the cutting plane according to the direction perpendicular to a certain axis direction of the tissue, and realize the slicing processing of the three-dimensional image according to the specified length through the set cutting plane.

[0023] S4: Respectively perform image feature extraction on the multiple first slice images, compare the extracted image features with the preset lesion image features, obtain the features that match the preset lesion feature items, and mark the lesion areas in each first slice image based on the features that match the preset lesion feature items, retain the lesion areas, and delete the areas outside the lesion areas. The way to delete the areas outside the lesion areas is to perform image excision processing, that is, excise the image areas outside the lesion areas along the edges of the lesion areas.

[0024] S5: Recombine the retained regions based on each first slice image in the original slice order to construct the first three-dimensional image of the lesion. Identify the first three-dimensional image of the lesion to determine the lesion category and size. When it belongs to the specified category or its size is greater than the preset threshold, execute step S6; otherwise, output the three-dimensional data of the lesion including the lesion category and size. Since only the lesion regions are retained in each image, directly obtaining the three-dimensional image of the lesion can be achieved when recombining the images based on the retained lesion regions. When obtaining a separate three-dimensional image of the lesion, it can provide great convenience for subsequent identification of the lesion type and calculation of the lesion size. However, due to the nature of the ultrasound imaging device itself, there are certain limitations in the obtained ovarian ultrasound medical images, such as low imaging quality or unclear and inaccurate display of the lesion, resulting in certain errors in the finally constructed three-dimensional image of the lesion. Therefore, when the first three-dimensional image constructed based on the ovarian ultrasound medical image shows that the patient has a specified lesion or the size exceeds the preset threshold, such as showing critical lesions such as tumors in the patient's ovary or the lesion size is too large, it is necessary to further obtain the ovarian CT image through a CT device and perform three-dimensional imaging on the lesion again to analyze various data of the lesion, so as to improve the accuracy of assisting in the diagnosis of the patient's condition. For those with slightly shown lesion types and small sizes, there is no need to obtain subsequent ovarian CT images.

[0025] S6: Obtain the ovarian CT medical images of the patient in all directions through a CT device. After preprocessing the ovarian CT medical images, repeat steps S2 - S5 to obtain the second three-dimensional image of the lesion, as well as the lesion category and size, and output them. The doctor assists in diagnosing the patient's condition based on the output data.

[0026] Although with the development of image processing technology, in the prior art, before a doctor diagnoses a patient's condition based on medical images of various parts of the patient, image recognition is first performed on the medical images through an image recognition system. The medical images are input into the image recognition system, which has corresponding algorithms preset inside and contains the corresponding image data of various lesions. When the medical images are input into the image recognition system, each region in the input image is compared with the preset lesion image data to determine the type and size of the lesion, thereby assisting the doctor in diagnosing the patient's condition. However, due to the special location of the ovary, in the ovarian fossa on the pelvic side wall, and being blocked by adjacent organs and having rich blood flow, it has a certain impact on the imaging device for taking ovarian medical images. Therefore, it is necessary to improve the image processing process in the prior art to highlight ovarian lesions with the existing ovarian medical images and assist the doctor in accurately diagnosing the patient's condition.

[0027] Therefore, in this application, ovarian ultrasound medical images in various directions of the patient are obtained and preprocessed, and then input into a three-dimensional image processing system; the ovarian ultrasound medical images in each direction are three-dimensionally reconstructed to obtain three-dimensional images of the ovary and adjacent tissues; slicing is performed along a certain direction of the three-dimensional image to obtain multiple first slice images for image feature extraction. Features matching the preset lesion feature items are obtained through feature comparison and marked, the lesion area is retained, and the areas outside the lesion area are deleted; the first three-dimensional image of the lesion is reconstructed by recombination, and the first three-dimensional image of the lesion is identified. When it belongs to a specified category or its size is greater than a preset threshold, ovarian CT medical images are obtained and the lesion is three-dimensionally reconstructed again to obtain the second three-dimensional image of the lesion, as well as the lesion category and size, and then output. In this process, the three-dimensional construction of ovarian lesions can accurately assist doctors in the accurate diagnosis of the patient's condition. On the one hand, it can improve the efficiency of assisting doctors in diagnosis, and on the other hand, it can effectively improve the accuracy of assisting doctors in diagnosis.

[0028] Embodiment 2 On the basis of Embodiment 1, the specific process of image preprocessing in step S1 is as follows: S11: After performing image denoising processing on the ovarian ultrasound medical images in each direction, perform image scaling processing on the ovarian ultrasound medical images in each direction, and scale the ovarian ultrasound medical images in each direction to a matching size. The image denoising process is as follows: Deploy a filter, and set the size of the filter to 5x5. Place the 5x5 filter on each pixel of the image: Take the current pixel as the center, cover the filter on the image, and calculate the average value of the surrounding pixels: Take the average of all pixel values within the area covered by the filter to obtain a new pixel value, and assign the new pixel value to the pixel value at the current position, and replace the original pixel value with the calculated new pixel value.

[0029] S12: Perform image cutting on the ovarian ultrasound medical images in each direction after scaling processing, and obtain multiple image blocks for each ovarian ultrasound medical image; S13: Remap the pixel values of each pixel point inside each image block according to a specified rule, calculate the gray values of each pixel point within each image block, and count the number of pixel points exceeding the preset threshold within each image block, and evenly disperse the part exceeding the threshold to a specified gray value.

[0030] In this embodiment, the specific process of image scaling processing in step S11 is as follows: S111: Create an image coordinate system, and obtain the coordinates of the target pixel point in the ovarian ultrasound medical image under the image coordinate system M ( x , y ) S112: Obtain the coordinates of 4 pixel points within the M neighborhood of the coordinates of the target pixel point, Q 11 ( x 1, y 1), Q 12 ( x 1, y 2), Q 21 ( x 2, y 1 ) 、Q 22 ( x 2, y 2); S113: Calculate the x coordinates of the intermediate pixel points R1 and R2: R1 = x 1 + ( x - x 1) / ( x 2 - x 1) * ( x 2 - x ), R2 = x 1 + ( x - x 1) / ( x 2 - x 1) * ( y 2 - x ); S114: Perform interpolation in the y direction: For R1, calculate f (R1, y ) = (Q11 * ( y 2 - y ) + Q21 * ( y - y 1)) / ( y 2 - y 1); For R2, calculate f (R2, y ) = (Q12 * ( y 2 - y ) + Q22 * ( y - y 1)) / ( y 2 - y 1); S115: Calculate the final target point pixel value f ( x , y ) = (f(R1, y ) * ( y 2 - y ) + f(R2, y ) * (y - y 1)) / ( y 2 - y 1), f is an interpolation function.

[0031] The specific process of three - dimensional reconstruction based on ovarian ultrasound medical images in each direction in step S2 is as follows: S21: Based on the gray - scale values of each pixel point in the ovarian ultrasound medical images obtained in step S12, transform the gray - scale value of each pixel point according to a specified rule to obtain the depth value of each pixel point; S22: Generate a set of data points representing the 3D shape based on the depth value of each pixel point. The information of each data point in the set of data points includes coordinates in three - dimensional space; S23: Simplify the set of data points to obtain a mesh composed of a specified shape, and map each ovarian ultrasound medical image to the mesh surface according to the correspondence between the position of pixel points and the mesh, to realize the three - dimensional image of the tissue in the ovarian ultrasound medical image.

[0032] Embodiment 3 Based on Embodiment 1 or Embodiment 2, the specific process of extracting image features from the multiple first slice images in step S3 is as follows: S31: Create a convolutional neural network model for extracting image features, train the convolutional neural network model with the labeled ovarian ultrasound medical images, copy the multiple first slice images into multiple groups, and the number of groups of slice images is the same as the number of categories of image features to be extracted; S32: The convolutional neural network model is set with multiple input channels, and the number of input channels matches the number of groups of slices. Input different groups of slice images into different input channels of the convolutional neural network model respectively according to the slice order; S33: Different input channels correspond to different combinations of convolutional layers and pooling layers. Set sliding windows of different sizes in each convolutional layer. Each sliding window is set as a matrix of a specified size. The matrix of each convolutional layer slides on each slice image. Different convolutional layers extract different local feature maps of the image through different dot - product operations. The pooling layer performs down - sampling processing on the output of the convolutional layer, and calculates the maximum value as the output of the pooling layer through the window set in the pooling layer for the output of the convolutional layer; S34: Set a fully - connected layer after the combination layer of convolutional layers and pooling layers to perform feature fusion on the outputs of the combination layers of convolutional layers and pooling layers in different channels, and transmit it to the output layer for output.

[0033] In this embodiment, the specific process of the fully connected layer performing feature fusion on the outputs of the combined layers of convolutional layers and pooling layers in different channels in step S34 is as follows: S341: The fully connected layer converts the local features extracted from different channels into local feature vectors; S342: Calculate the feature vectors corresponding to the local features extracted from different channels according to a preset rule to obtain a final global feature vector; S343: Map the global feature vector to the probabilities corresponding to features of different categories through a preset activation function.

[0034] See Figure 1 As shown, the convolutional neural network model for extracting image features of the present invention is provided with an input layer, a combined layer of convolutional layers and pooling layers, a fully connected layer, and an output layer. The input layer is provided with a plurality of input channels, and the number of the input channels is the same as the number of categories of image features to be extracted. The image features include edge features, corner features, texture features, etc. The input channels are matched with the combined layer of convolutional layers and pooling layers, that is, each input channel is matched with a specified group of combined layers of convolutional layers and pooling layers for extracting one type of image feature. That is to say, the edge feature is input through one input channel and undergoes feature extraction through a specific combined layer of convolutional layers and pooling layers. Other types of corner features, texture features, etc. are also respectively input through different input channels and undergo feature extraction through specific combined layers of convolutional layers and pooling layers. Therefore, after the output of the combined layer of convolutional layers and pooling layers, there are different types of image features, which are then integrated and processed by the fully connected layer and output through the output layer.

[0035] In step S4, the extracted image features are compared with the preset lesion image features to obtain the features that match the preset lesion feature items. The specific process is as follows: S41: Convert the preset lesion image features into corresponding feature vectors, and match the global feature vector in step S343 with the feature vectors corresponding to the converted preset lesion image features; S42: Verify the matching result based on the probabilities corresponding to features of different categories, obtain the lesion features in the patch image, and obtain the regions corresponding to the lesion features in the slice image.

[0036] When there are multiple types of lesions shown by the lesion features in the slice image obtained in step S42, color them with different colors in the corresponding regions. When recombining according to the original slice order based on the regions retained in each first slice image in step S5, different three-dimensional reconstructions are performed on different lesions according to different colors.

[0037] In summary, the three-dimensional stereoscopic peeling method of the lesion of the medical image provided by the present invention obtains the ovarian ultrasound medical images of the patient in all directions and inputs them into the three-dimensional image processing system after preprocessing; the ovarian ultrasound medical images in all directions are three-dimensionally reconstructed to obtain the three-dimensional images of the ovary and adjacent tissues; slice processing is performed along a certain direction of the three-dimensional image to obtain multiple first slice images for image feature extraction, and features matching the preset lesion feature items are obtained and marked through feature comparison, the lesion area is retained, and the area outside the lesion area is deleted; the first three-dimensional image of the lesion is recombined and constructed, and the first three-dimensional image of the lesion is identified. When it belongs to the specified category or its size is greater than the preset threshold, the ovarian CT medical image is obtained and the lesion is reconstructed in three dimensions again, and the second three-dimensional image of the lesion and the lesion category and size are obtained and output. The three-dimensional construction of ovarian lesions can accurately assist doctors in making accurate diagnosis of patients' symptoms.

[0038] By constructing a three-dimensional image of the ovary and adjacent tissues based on ovarian ultrasound medical images, when there is a lesion in the image, a combined three-dimensional image between the lesion and the ovarian tissue can be obtained. However, since the lesion and the tissue are interlaced, even if the three-dimensional image is constructed, the relevant data of the lesion cannot be clearly obtained. The constructed combined three-dimensional image between the lesion and the ovarian tissue is further sliced, and feature extraction and other processing are performed based on the sliced ​​image. The image area other than the lesion area is removed, and only the lesion area is retained. The sliced ​​images of each lesion area are combined to obtain the three-dimensional data of the lesion, providing a data basis for subsequent lesion identification and size calculation, and improving the efficiency and accuracy of auxiliary diagnosis. Due to some of the critical and serious lesions, in order to avoid the limitations of the ultrasound equipment in obtaining the lesion display in the ovarian ultrasound medical image, it is necessary to obtain the CT image again through the CT imaging equipment, and then reconstruct the lesion in three dimensions based on the CT image, so as to further improve the accuracy of the three-dimensional image of the lesion, and further improve the efficiency and accuracy of auxiliary diagnosis.

Claims

1. A method for three-dimensional stereoscopic peeling of lesions in medical images, characterized in that, It includes the following steps: S1: Obtain ovarian ultrasound medical images of the patient in various directions through an ultrasound device, perform image preprocessing on the ovarian ultrasound medical images, and input the preprocessed ovarian ultrasound medical images into a three-dimensional image processing system; S2: The three-dimensional image processing system performs three-dimensional reconstruction based on the ovarian ultrasound medical images in various directions to obtain three-dimensional images of the ovaries and adjacent tissues within the medical images; S3: Perform image slicing processing on the constructed three-dimensional images of the ovaries and adjacent tissues at a specified thickness and along a certain direction of the three-dimensional image to obtain multiple first slice images of the specified thickness; S4: Perform image feature extraction on each of the multiple first slice images respectively, compare the extracted image features with the preset lesion image features, obtain the features that match the preset lesion feature items, and mark the lesion areas in each first slice image based on the features that match the preset lesion feature items, retain the lesion areas, and delete the areas outside the lesion areas; S5: Recombine the retained areas of each first slice image in the original slice order to construct a first three-dimensional image of the lesion, identify the lesion category and size of the first three-dimensional image of the lesion. When it belongs to a specified category or its size is greater than a preset threshold, execute step S6; otherwise, output the lesion three-dimensional data including the lesion category size; S6: Obtain ovarian CT medical images of the patient in various directions through a CT device, and after preprocessing the ovarian CT medical images, repeat steps S2 - S5 to obtain a second three-dimensional image of the lesion, as well as the lesion category and size and output them. The doctor aids in diagnosing the patient's condition based on the output data.

2. The three-dimensional stereoscopic peeling method for lesions of medical images according to claim 1, wherein, The specific process of the image preprocessing in step S1 is as follows: S11: After performing image denoising processing on the ovarian ultrasound medical images in various directions, perform image scaling processing on the ovarian ultrasound medical images in various directions to scale the ovarian ultrasound medical images in various directions to a matching size; S12: Perform image cutting on the scaled ovarian ultrasound medical images in various directions respectively, and each ovarian ultrasound medical image obtains multiple image blocks; S13: Remap the pixel values of each pixel point inside each image block according to a specified rule, calculate the gray values of each pixel point within each image block, and count the number of pixel points exceeding the preset threshold within each image block, and evenly disperse the part exceeding the threshold to a specified gray value.

3. The three-dimensional stereoscopic peeling method for lesions of medical images according to claim 2, characterized in that, The specific process of the image scaling processing in step S11 is as follows: S111: Create an image coordinate system and obtain the coordinates of the target pixel points in the ovarian ultrasound medical image under the corresponding image coordinate system M ( x , y ); S112: Obtain the coordinates of 4 pixel points within the M neighborhood of the coordinates of the target pixel point, Q 11 ( x 1, y 1), Q 12 ( x 1, y 2), Q 21 ( x 2, y 1 ) 、Q 22 ( x 2, y 2); S113: Calculate the x coordinates of the intermediate pixels R1 and R2: R1 = x 1 + ( x - x 1) / ( x 2 - x 1) * ( x 2 - x ), R2 = x 1 + ( x - x 1) / ( x 2 - x 1) * ( y 2 - x ); S114: Interpolate in the y direction: For R1, calculate f (R1, y ) = (Q11 * ( y 2 - y ) + Q21 * ( y - y 1)) / ( y 2 - y 1); For R2, calculate f (R2, y ) = (Q12 * ( y 2- y ) + Q22 * ( y - y 1)) / ( y 2- y 1); S115: Calculate the pixel value of the final target point f ( x , y ) = (f(R1, y ) * ( y 2 - y ) + f(R2, y ) * ( y - y 1)) / ( y 2 - y 1), f where f is the interpolation function.

4. A three-dimensional stereoscopic peeling method for lesions of medical images according to claim 2, wherein, The specific process of the three-dimensional reconstruction based on the ovarian ultrasound medical images in various directions in step S2 is as follows: S21: Based on the gray values of each pixel point in each ovarian ultrasound medical image obtained in step S12, transform the gray value of each pixel point according to a specified rule to obtain the depth value of each pixel point; S22: Generate a set of data points representing the 3D shape based on the depth value of each pixel point. The information of each data point in the set of data points includes the coordinates in three-dimensional space; S23: Simplify the set of data points to obtain a grid composed of specified shapes, and map each ovarian ultrasound medical image to the surface of the grid according to the correspondence between the positions of pixel points and the grid, so as to realize the three-dimensional image of the tissues in the ovarian ultrasound medical image.

5. A three-dimensional stereoscopic peeling method for lesions in medical images according to claim 1, characterized in that, The specific process of extracting image features from the multiple first slice images in step S3 is as follows: S31: Create a convolutional neural network model for extracting image features, train the convolutional neural network model with the labeled ovarian ultrasound medical images, and copy the multiple first slice images into multiple groups, where the number of groups of slice images is the same as the number of categories of image features to be extracted; S32: Set the convolutional neural network model to have multiple input channels, where the number of input channels matches the number of groups of slices, and input different groups of slice images into different input channels of the convolutional neural network model respectively according to the slice order; S33: Different input channels correspond to a combination layer of different convolutional layers and pooling layers. Set sliding windows of different sizes in each convolutional layer, where each sliding window is set as a matrix of a specified size. The matrix of each convolutional layer slides on each slice image. Different convolutional layers extract different local feature maps of the image through different dot product operations. The pooling layer performs downsampling on the output of the convolutional layer, and calculates the maximum value as the output of the pooling layer through the window set in the pooling layer for the output of the convolutional layer; S34: Set a fully connected layer after the combination layer of the convolutional layer and the pooling layer to perform feature fusion on the outputs of the combination layers of the convolutional layer and the pooling layer in different channels, and transmit it to the output layer for output.

6. A three-dimensional stereoscopic peeling method for lesions in medical images according to claim 5, characterized in that, The specific process of the fully connected layer in step S34 performing feature fusion on the outputs of the combination layers of the convolutional layer and the pooling layer in different channels is as follows: S341: The fully connected layer converts the local features extracted from different channels into local feature vectors; S342: Calculate the feature vectors corresponding to the local features extracted from different channels according to a preset rule to obtain a final global feature vector; S343: Map the global feature vector to the probabilities corresponding to the features of different categories through a preset activation function.

7. A three-dimensional stereoscopic peeling method for lesions of medical images according to claim 6, characterized in that, The specific process of comparing the extracted image features with the preset lesion image features in step S4 and obtaining the features that match the preset lesion feature items is as follows: S41: Convert the preset lesion image features into corresponding feature vectors, and match the global feature vector in step S343 with the feature vectors corresponding to the conversion of the preset lesion image features; S42: Verify the matching result based on the probabilities corresponding to the features of different categories, obtain the lesion features in the slice image, and obtain the region corresponding to the lesion features in the slice image.

8. A three-dimensional stereoscopic peeling method for lesions in medical images according to claim 7, characterized in that, When there are multiple types of lesions shown in the lesion features in the slice image obtained in step S42, perform coloring processing with different colors on different types of lesions in the corresponding regions.

9. A three-dimensional stereoscopic peeling method for lesions in medical images according to claim 8, characterized in that, When recombining based on the regions retained in each first slice image in step S5 in the original slice order, perform different three-dimensional reconstructions on different lesions according to different colors.

Images