Brain anatomical structure localization method for two-dimensional brain image data

Abstract

The present invention discloses a method for locating brain anatomical structures suitable for two-dimensional brain image data, which obtains a two-dimensional brain image template by retrieval, obtains a feature area by segmentation, and uses a registration method to match the two-dimensional brain image to be located with the two-dimensional brain image template, thereby obtaining the position information of the brain anatomical structure. The entire positioning process does not require human participation, and can achieve rapid brain anatomical structure positioning. The positioning target is all brain anatomical structures on the two-dimensional brain image, and the positioning accuracy is improved compared with the existing method.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a brain anatomical structure positioning method applicable to two-dimensional brain image data. Background Art

[0002] With the advancement of imaging technology, people can obtain a large amount of brain image data through different imaging technologies, and use these data to study the structure and function of the brain. When studying different brain anatomical structures, it is very important to locate the brain anatomical structure of interest in the brain image data. When locating the brain anatomical structure of interest, professionals usually learn a lot of anatomical knowledge through long-term training to determine the position of the brain anatomical structure in the image. However, this method is very time-consuming, and for brain anatomical structures with unclear texture features, manual positioning often produces large deviations. In order to meet the needs of brain anatomical structure positioning of massive brain image data, a large number of studies have been conducted to digitize the brain anatomical structure positioning process of three-dimensional brain images through image registration, such as using manual-assisted methods to register three-dimensional brain images with three-dimensional brain maps, and using automatically extracted features to match three-dimensional brain images with three-dimensional brain maps. However, for two-dimensional brain images, due to the lack of matching templates and the inability to directly match two-dimensional brain images with three-dimensional brain maps, the brain anatomical structure positioning method of three-dimensional brain images cannot be directly transferred to two-dimensional brain images, and corresponding brain anatomical structure positioning methods need to be developed for two-dimensional images.

[0003] Existing methods for locating brain anatomical structures in two-dimensional images are mainly divided into two types: image segmentation methods and image registration methods. Image segmentation methods can be further divided into image segmentation based on traditional methods and segmentation methods based on deep learning. Image segmentation based on traditional methods mainly includes region division, k-means clustering, watershed, active contour, graph-based segmentation, Markov random field and other methods. The deep learning-based segmentation method mainly divides the brain anatomical structure by learning the brain anatomical structure features in the two-dimensional brain image. For example, the Chinese patent name is "A Brain Tumor Medical Image Segmentation Method Based on Deep Learning", and the application number is: 201910158251.5. The deep learning-based brain tumor medical image segmentation method performs deep learning training on the segmentation model. The deep learning trained segmentation model then performs segmentation processing on the received brain tumor medical image data information to be segmented. The segmentation result is accurate and fast. It can be applied to two-dimensional brain images of different modalities, solving the problem of time-consuming and labor-intensive traditional manual segmentation methods. However, due to hardware limitations, this method can only locate limited brain anatomical structures. In addition, due to the lack of a large amount of accurate labeling data, the accuracy of the segmentation result is insufficient. The image registration method mainly divides the brain anatomical structure by finding a two-dimensional template and then registering the two-dimensional brain image with the two-dimensional template. This method can locate all brain anatomical structures, so it is the mainstream method currently studied. Existing image registration methods mainly learn the correspondence between three-dimensional template images and two-dimensional brain images through deep learning, slice the three-dimensional template images, and then obtain the two-dimensional template for registration. However, since there are often only three-dimensional template images of a single modality, when applied to images of different modalities, the similarity between the two-dimensional template and the two-dimensional brain image will drop sharply, thereby seriously reducing the positioning accuracy of the brain anatomical structure, thus greatly limiting the widespread application of such methods. Summary of the invention

[0004] Therefore, in order to solve the above problems, the present invention provides a brain anatomical structure positioning method applicable to two-dimensional brain image data.

[0005] The present invention is achieved through the following technical solutions:

[0006] A method for locating brain anatomical structures applicable to two-dimensional brain image data comprises the following steps:

[0007] S1: Collect annotated three-dimensional brain images, enhance the three-dimensional brain image data and slice them to obtain a two-dimensional brain image database, extract the brain anatomical structure distribution features in the two-dimensional brain image database, and obtain a feature database;

[0008] S2: Using the two-dimensional brain image database and the feature database to obtain a feature extraction network, the most similar features between the image to be located and the feature database and their corresponding detected images are retrieved through the feature extraction network;

[0009] S3: Using the data in the two-dimensional brain image database to train a fully convolutional feature region segmentation network, the feature region of the image to be located and the feature region of the detected image are segmented through the fully convolutional feature region segmentation network;

[0010] S4: The image to be located and the detected image are respectively fused with their corresponding feature regions to obtain a fused image to be located and a detected fused image, and the fused image to be located and the detected fused image are registered to obtain a brain anatomical structure positioning result of the image to be located.

[0011] Preferably, the step S1 comprises:

[0012] S11: Collect annotated three-dimensional brain images to form a three-dimensional brain image database, where each three-dimensional brain image in the database corresponds to a three-dimensional annotated image;

[0013] S12: performing data enhancement processing on each 3D brain image and its 3D annotated image by using affine transformation and grayscale stretching, and storing the 3D brain image and 3D annotated image data after the data enhancement processing into a 3D brain image database;

[0014] S13: Slicing the three-dimensional brain image and the annotated image after data enhancement processing in the coronal direction to obtain a two-dimensional brain image database formed by the corresponding two-dimensional brain image and the two-dimensional annotated image;

[0015] S14: Divide the number of pixels occupied by each brain anatomical structure in the two-dimensional annotated image corresponding to each two-dimensional brain image by the number of pixels occupied by the entire brain contour, calculate the area ratio occupied by each brain anatomical structure in the entire brain contour, and obtain a feature database of the two-dimensional brain image.

[0016] Preferably, each three-dimensional brain image corresponds to a three-dimensional annotated image marked with 672 brain anatomical structures, the pixels in the three-dimensional annotated image correspond to the grayscale values ​​in the brain anatomical structure grayscale mapping table, and the grayscale value of each pixel represents the brain anatomical structure to which the pixel belongs.

[0017] Preferably, the affine transformation refers to: rotating in the x, y, and z directions respectively, with the rotation angle not exceeding 15°, the step size of the angle change being 1°, using bilinear interpolation when performing affine transformation on the three-dimensional brain image, and using nearest neighbor interpolation when performing affine transformation on the three-dimensional annotated image, so as to obtain the three-dimensional brain image and the three-dimensional annotated image after affine transformation; the grayscale stretching refers to: using adaptive histogram equalization to perform grayscale stretching on the three-dimensional brain image, wherein the contrast threshold of the equalization is set to 1.2, and the grid size of the equalization is 4*4.

[0018] Preferably, step S2 comprises:

[0019] S21: Use convolutional neural network to train the two-dimensional brain image database and feature database to obtain feature extraction network;

[0020] S22: Extracting features of the image to be located using a feature extraction network;

[0021] S23: Compare the features of the image to be located with the distribution features of the brain anatomical structure in the feature database to obtain the most similar features and their corresponding detected images.

[0022] Preferably, the most similar feature refers to: a brain anatomical structure distribution feature in a two-dimensional brain image database that has the smallest Euclidean distance to the feature of the image to be located.

[0023] Preferably, step S3 comprises:

[0024] S31: marking the characteristic region of the two-dimensional brain image to obtain a marked image;

[0025] S32: normalizing each two-dimensional brain image and deleting images without brain anatomical structures, inputting the normalized two-dimensional brain images into a fully convolutional neural network, using the labeled images as labels, adopting multi-category cross entropy as a loss function, training the network, and constructing a fully convolutional feature region segmentation network;

[0026] S33: inputting the image to be located into a fully convolutional feature region segmentation network to extract the feature region of the image to be located;

[0027] S34: Input the detected image into the fully convolutional feature region segmentation network to extract the feature region of the detected image.

[0028] Preferably, in step S31, there are 12 characteristic regions of the two-dimensional brain image, and their grayscales are 1-12 respectively. The characteristic regions are marked by grayscale extraction results, and the marked results are integrated on the marked image.

[0029] Preferably, the two-dimensional brain image is sampled using bilinear interpolation before being stored in the two-dimensional brain image database, the two-dimensional annotated image is sampled using nearest neighbor interpolation before being stored in the two-dimensional brain image database, and the image to be located is sampled using bilinear interpolation before being input into the full convolution feature region segmentation network, and the size of the sampled images is 512 * 512 pixel.

[0030] Preferably, step S4 comprises:

[0031] S41: averaging the grayscale of each pixel on the image to be located and the feature area of ​​the image to be located, and making a one-to-one correspondence between the pixels on the two images to obtain a fused image to be located;

[0032] S42: averaging the grayscale of each pixel in the detected image and the feature area of ​​the detected image, and making a one-to-one correspondence between the pixels in the two images to obtain a detected fused image;

[0033] S43: using mutual information as a similarity measure, rigidly registering the positioning fusion image and the detection fusion image to obtain a transformation matrix;

[0034] S44: Using mutual information as a similarity measure, the results after rigid registration are subjected to nonlinear registration of symmetric diffeomorphism to obtain the deformation field;

[0035] S45: applying the transformation matrix and the deformation field to the annotated data corresponding to the detected fused image, and finally obtaining the brain anatomical structure positioning result of the fused image to be positioned.

[0036] The beneficial effects of the technical solution of the present invention are mainly reflected in:

[0037] 1. Combining image segmentation and image registration, the segmentation method is used to obtain the feature area, and the registration method is used to match the two-dimensional brain image to be located with the two-dimensional brain image template, so as to obtain the position information of the brain anatomical structure. It can not only realize the rapid positioning of the brain anatomical structure of the two-dimensional brain image data, but also ensure the positioning accuracy of the brain anatomical structure when applied to images of different modalities.

[0038] 2. Each 3D brain image in the 3D brain image database corresponds to an annotated image with 672 brain anatomical structures marked. The samples are rich, and the annotated image corresponds to the grayscale mapping table of the brain anatomical structure. The grayscale value of each pixel represents the brain anatomical structure to which the pixel belongs, and the classification is clear. In addition, before slicing, the 3D brain images and 3D annotated images are data enhanced to obtain more samples to ensure the accuracy of the calculation.

[0039] 3. The convolutional neural network consists of 16 convolutional layers, pooling layers and fully connected layers. The front end of the network learns the features of the two-dimensional brain image through multiple convolutional layers, and then passes through an average pooling layer and a fully connected layer to convert the features learned by the convolutional layer into brain anatomical structure distribution features. The feature extraction network is trained by the convolutional neural network and can achieve rapid positioning based on deep learning of the brain anatomical structure features in the two-dimensional brain image.

[0040] 4. Use grayscale to extract and mark the feature areas of two-dimensional brain images, and integrate the marking results into one image for easy extraction. A fully convolutional feature area segmentation network is constructed based on the fully convolutional neural network to extract the feature areas of two-dimensional brain images. The network consists of convolutional layers, pooling layers, and deconvolution layers, and uses jump connections to fuse features of different depths to make the segmentation results more accurate.

[0041] 5. In the registration process of the positioning fusion image and the detection fusion image, rigid registration and nonlinear registration are combined to match the global information and local information of the image respectively, so as to obtain a complete and accurate registration result.

[0042] 6. All brain anatomical structures in two-dimensional brain images can be located, and can be automatically located without human intervention, with higher positioning accuracy, and can be applied to various types of brain images. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 : is a flow chart of a method for localizing brain anatomical structures applicable to two-dimensional brain image data;

[0044] Figure 2 : It is a schematic diagram of the structure of a convolutional neural network used to build a feature extraction network;

[0045] Figure 3 : It is a structural diagram of a fully convolutional neural network used to build a fully convolutional feature region segmentation network. DETAILED DESCRIPTION

[0046] In order to make the purpose, advantages and features of the present invention more clearly and in detail, the following preferred embodiments are used for illustration and explanation. The embodiments are only typical examples of the application of the technical solution of the present invention. Any technical solution formed by equivalent replacement or equivalent transformation falls within the scope of protection claimed by the present invention.

[0047] The present invention discloses a method for locating brain anatomical structures applicable to two-dimensional brain image data, such as Figure 1 As shown, the following steps are included:

[0048] S1: Collect annotated three-dimensional brain images, enhance the three-dimensional brain image data and slice them to obtain a two-dimensional brain image database, extract the brain anatomical structure distribution features in the two-dimensional brain image database, and obtain a feature database;

[0049] In one embodiment, the step S1 includes:

[0050] S11: Collect annotated three-dimensional brain images to form a three-dimensional brain image database, where each three-dimensional brain image in the database corresponds to a three-dimensional annotated image;

[0051] Among them, each three-dimensional brain image corresponds to a three-dimensional annotated image marked with 672 brain anatomical structures. The pixels in the three-dimensional annotated image correspond to the grayscale values ​​in the brain anatomical structure grayscale mapping table, and the grayscale value of each pixel represents the brain anatomical structure to which the pixel belongs.

[0052] S12: performing data enhancement processing on each 3D brain image and its 3D annotated image by using affine transformation and grayscale stretching, and storing the 3D brain image and 3D annotated image data after the data enhancement processing into a 3D brain image database;

[0053] Among them, the affine transformation refers to: rotating in the x, y, and z directions respectively, with the rotation angle not exceeding 15° and the step size of the angle change being 1°. Bilinear interpolation is used when performing affine transformation on the three-dimensional brain image, and nearest neighbor interpolation is used when performing affine transformation on the three-dimensional annotated image, so as to obtain the three-dimensional brain image and the three-dimensional annotated image after affine transformation; the grayscale stretching refers to: using adaptive histogram equalization to perform grayscale stretching on the three-dimensional brain image, wherein the contrast threshold of the equalization is set to 1.2 and the grid size of the equalization is 4*4; through affine transformation and grayscale stretching, more three-dimensional brain images and three-dimensional annotated image data are obtained, thereby realizing the expansion of the three-dimensional brain image database.

[0054] S13: Slicing the three-dimensional brain image and the annotated image after data enhancement processing in the coronal direction to obtain a two-dimensional brain image database formed by the corresponding two-dimensional brain image and the two-dimensional annotated image;

[0055] Before each 2D brain image is stored in the 2D brain image database, bilinear interpolation sampling is used to obtain a 2D brain image DI with a size of 512 * 512 pixels. i , all 2D brain images DI i The two-dimensional brain image set DI is formed; before each two-dimensional annotated image is stored in the two-dimensional brain image database, the nearest neighbor interpolation sampling is used to obtain a two-dimensional annotated image DA with a picture size of 512 * 512 pixels. i , all 2D annotated images DA iForm a two-dimensional annotated image set DA, for each two-dimensional brain image DI in the two-dimensional brain image set DI i , can find the corresponding two-dimensional annotated image DA in the two-dimensional annotated image set DA i .

[0056] S14: Divide the number of pixels occupied by each brain anatomical structure in the two-dimensional annotated image corresponding to each two-dimensional brain image by the number of pixels occupied by the entire brain contour, calculate the area ratio occupied by each brain anatomical structure in the entire brain contour, and obtain a feature database of the two-dimensional brain image.

[0057] Among them, since each three-dimensional brain image corresponds to an annotated image with 672 brain anatomical structures, each two-dimensional brain image obtained by slicing the three-dimensional brain image also corresponds to a two-dimensional annotated image with 672 brain anatomical structures. The pixels in the two-dimensional annotated image correspond to the grayscale values ​​in the brain anatomical structure grayscale mapping table. The grayscale value of each pixel represents the brain anatomical structure to which the pixel belongs. According to the brain anatomical structure grayscale mapping table, the DA of each two-dimensional annotated image is calculated. i The number of pixels occupied by each brain anatomical structure in the DA i The number of pixels occupied by each brain anatomical structure in the image is divided by the number of pixels occupied by the entire brain, thereby calculating the DA of each two-dimensional annotation image. i The area ratio of the 672 brain anatomical structures in the entire brain contour is obtained as a 672-dimensional feature vector DF i , the eigenvector DF i Represents a two-dimensional brain image DI i The brain anatomical structure distribution characteristics are calculated for all two-dimensional brain images DI in the two-dimensional brain image set DI i The brain anatomical structure distribution characteristics are used to obtain the feature database DF of the two-dimensional brain image.

[0058] S2: Using the two-dimensional brain image database and the feature database to obtain a feature extraction network, the most similar features between the image to be located and the feature database and their corresponding detected images are retrieved through the feature extraction network;

[0059] In one embodiment, the step S2 includes:

[0060] S21: Use convolutional neural network to train the two-dimensional brain image database and feature database to obtain feature extraction network;

[0061] Among them, Figure 2As shown, the convolutional neural network consists of 16 convolutional layers, pooling layers and fully connected layers. The front end of the network learns the features of the two-dimensional brain image through multiple convolutional layers, and then passes through an average pooling layer and a fully connected layer to convert the features learned by the convolutional layer into brain anatomical structure distribution features, and build a feature extraction network NF. After building the feature extraction network NF, the two-dimensional brain images in the two-dimensional brain image set DI are normalized, and the images without brain anatomical structures are removed. Then, the normalized images are input into the convolutional neural network, and the feature database DF of the two-dimensional brain image is used as the label, and the mean square error is used as the loss function to train the built feature extraction network NF. The trained feature extraction network NF can be used to extract the brain anatomical structure distribution features of the two-dimensional brain image.

[0062] S22: Extracting features of the image to be located using a feature extraction network;

[0063] Specifically, the image to be located TI is input into the feature extraction network NF to obtain the brain anatomical structure distribution feature TF of the image to be located TI.

[0064] S23: Compare the features of the image to be located with the features in the feature database to obtain the most similar features and their corresponding detected images.

[0065] Specifically, the “comparing the features of the image to be located with the features in the feature database” refers to calculating the Euclidean distance between the brain anatomical structure distribution feature TF and each brain anatomical structure distribution feature in the feature database DF of the two-dimensional brain image; the most similar feature refers to: in the two-dimensional brain image database DF, the brain anatomical structure distribution feature with the smallest Euclidean distance calculated with the features of the image to be located; let the most similar feature be DF m , through the most similar feature DF m Find the corresponding two-dimensional brain image DI in the two-dimensional brain image set DI m .

[0066] S3: Using the data in the two-dimensional brain image database to train a fully convolutional feature region segmentation network, the feature region of the image to be located and the feature region of the detected image are segmented through the fully convolutional feature region segmentation network;

[0067] In one embodiment, step S3 includes:

[0068] S31: marking the characteristic region of the two-dimensional brain image to obtain a marked image;

[0069] There are 12 characteristic regions of the two-dimensional brain image, and their grayscales are 1-12 respectively. The characteristic regions are marked by grayscale extraction results, and the marked results are integrated on the marked image.

[0070] In this embodiment, 12 regions or brain anatomical structures including Outline, act, CB, cc, CP, HIP, fr, fx, mtt, MH, PG, and VL are selected as feature regions of the two-dimensional brain image. In the two-dimensional brain image database, each two-dimensional brain image DI i In the corresponding two-dimensional annotation image DA i In the 2D labeled image DA, the grayscales of these 12 regions or brain anatomical structures range from 1 to 12. i The grayscale annotated in the image is extracted and their labeling results are integrated into an image, which is a two-dimensional brain image DI i Feature region annotation image DS i , the feature region annotated images of all the two-dimensional brain images in the two-dimensional brain image set form an annotated image set DS.

[0071] S32: normalizing each two-dimensional brain image and deleting images without brain anatomical structures, inputting the normalized two-dimensional brain images into a fully convolutional neural network, using the labeled images as labels, adopting multi-category cross entropy as a loss function, training the network, and constructing a fully convolutional feature region segmentation network;

[0072] Specifically, Figure 3 As shown, the fully convolutional neural network (Unet++) is improved from the Unet segmentation model, and consists of a convolutional layer, a pooling layer, and a deconvolution layer. It uses jump connections to fuse features of different depths, and builds a fully convolutional feature region segmentation network through a fully convolutional neural network.

[0073] After constructing the fully convolutional feature region segmentation network, the two-dimensional brain images in the two-dimensional brain image set DI are normalized, and the images without brain anatomical structures are removed. The normalized images are input into the fully convolutional feature region segmentation network NS. The network is trained with the annotated image set DS as labels and multi-category cross entropy as the loss function, thereby obtaining a fully convolutional feature region segmentation network NS for extracting feature regions of two-dimensional brain images.

[0074] S33: inputting the image to be located into a fully convolutional feature region segmentation network to extract the feature region of the image to be located;

[0075] Specifically, the image to be located TI is sampled using bilinear interpolation to make its image size 512 * 512 pixels and then input into the fully convolutional feature region segmentation network NS to extract the feature region TS of the image to be located TI.

[0076] S34: Input the detected image into the fully convolutional feature region segmentation network to extract the feature region of the detected image.

[0077] Specifically, the detected image DI m After inputting the fully convolutional feature region segmentation network, the feature region DS of the detected image is extracted m Since the image size of each image in the two-dimensional brain image database is 512*512 pixel, the detected image does not need to be resized before being input into the fully convolutional feature region segmentation network.

[0078] S4: The image to be located and the detected image are respectively fused with their corresponding feature regions to obtain a fused image to be located and a detected fused image, and the fused image to be located and the detected fused image are registered to obtain a brain anatomical structure positioning result of the image to be located.

[0079] In one embodiment, step S4 includes:

[0080] S41: averaging the grayscale of each pixel on the image to be located and the feature area of ​​the image to be located, and making a one-to-one correspondence between the pixels on the two images to obtain a fused image to be located;

[0081] S42: averaging the grayscale of each pixel in the detected image and the feature area of ​​the detected image, and making a one-to-one correspondence between the pixels in the two images to obtain a detected fused image;

[0082] Specifically, averaging the grayscale of each pixel on the two images means multiplying the grayscale of all pixels on the two images by 0.5, and then matching all pixels of the two images one by one to obtain a fused image of the two images. Therefore, it is necessary to ensure that the sizes of the two images with the average grayscale are consistent. In this embodiment, since the image size of the image to be located TI and its feature area TS is the same, the detected image Di m With its characteristic area DS m The image size is the same as that of the detected image Di, and the grayscale of each pixel on the image to be located TI and its feature area TS is averaged to obtain the fused image TC corresponding to the image to be located TI. m With its characteristic area DS m Perform grayscale averaging to obtain the same image as the detected image Di m The corresponding fused image DC m .

[0083] S43: using mutual information as a similarity measure, rigidly registering the positioning fusion image and the detection fusion image to obtain a transformation matrix;

[0084] S44: Using mutual information as a similarity measure, the results after rigid registration are subjected to nonlinear registration of symmetric diffeomorphism to obtain the deformation field;

[0085] In this embodiment, after obtaining the fused image, firstly, the detected image Di m The corresponding fused image DCm The fused image TC corresponding to the image to be located TI uses mutual information as a similarity measure for rigid registration to match the global information of the image and obtain a transformation matrix M. Then, the result after rigid registration is subjected to symmetric differential homeomorphism nonlinear registration to match the local information of the image. The nonlinear registration also uses mutual information as a similarity measure to obtain a deformation field DDF.

[0086] S45: applying the transformation matrix and the deformation field to the annotated data corresponding to the detected fused image, and finally obtaining the brain anatomical structure positioning result of the fused image to be positioned.

[0087] In this embodiment, "applying the transformation matrix and the deformation field to the annotated data corresponding to the detected fused image" means: first using the transformation matrix to perform a linear transformation on the annotated data corresponding to the detected fused image, and then using the deformation field to perform a nonlinear transformation on the result of the linear transformation, and finally obtaining the brain anatomical structure positioning result of the fused image to be positioned.

[0088] There are many implementation methods of the present invention, and all technical solutions formed by equivalent transformation or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A method for localizing brain anatomical structures applicable to two-dimensional brain image data, characterized in that: The following steps are involved: S1: Collect annotated three-dimensional brain images, enhance the three-dimensional brain image data and slice them to obtain a two-dimensional brain image database, extract the brain anatomical structure distribution features in the two-dimensional brain image database, and obtain a feature database; S2: Using the two-dimensional brain image database and the feature database to obtain a feature extraction network, the most similar features between the image to be located and the feature database and their corresponding detected images are retrieved through the feature extraction network; S3: Using the data in the two-dimensional brain image database to train a fully convolutional feature region segmentation network, the feature region of the image to be located and the feature region of the detected image are segmented through the fully convolutional feature region segmentation network; S4: fusing the image to be located and the detected image with their corresponding feature regions respectively to obtain a fused image to be located and a detected fused image, and registering the fused image to be located and the detected fused image to obtain a brain anatomical structure localization result of the image to be located; The step S3 comprises: S31: marking the characteristic region of the two-dimensional brain image to obtain a marked image; S32: normalizing each two-dimensional brain image and deleting images without brain anatomical structures, inputting the normalized two-dimensional brain images into a fully convolutional neural network, using the labeled images as labels, adopting multi-category cross entropy as a loss function, training the network, and constructing a fully convolutional feature region segmentation network; S33: inputting the image to be located into a fully convolutional feature region segmentation network to extract the feature region of the image to be located; S34: inputting the detected image into a fully convolutional feature region segmentation network to extract the feature region of the detected image; The step S4 comprises: S41: averaging the grayscale of each pixel on the image to be located and the feature area of ​​the image to be located, and making a one-to-one correspondence between the pixels on the two images to obtain a fused image to be located; S42: averaging the grayscale of each pixel in the detected image and the feature area of ​​the detected image, and making a one-to-one correspondence between the pixels in the two images to obtain a detected fused image; S43: using mutual information as a similarity measure, rigidly registering the positioning fusion image and the detection fusion image to obtain a transformation matrix; S44: Using mutual information as a similarity measure, the results after rigid registration are subjected to nonlinear registration of symmetric diffeomorphism to obtain the deformation field; S45: applying the transformation matrix and the deformation field to the annotated data corresponding to the detected fused image, and finally obtaining the brain anatomical structure positioning result of the fused image to be positioned.

2. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 1, characterized in that: The step S1 comprises: S11: Collect annotated three-dimensional brain images to form a three-dimensional brain image database, where each three-dimensional brain image in the database corresponds to a three-dimensional annotated image; S12: performing data enhancement processing on each 3D brain image and its 3D annotated image by using affine transformation and grayscale stretching, and storing the 3D brain image and 3D annotated image data after the data enhancement processing into a 3D brain image database; S13: Slicing the three-dimensional brain image and the annotated image after data enhancement processing in the coronal direction to obtain a two-dimensional brain image database formed by the corresponding two-dimensional brain image and the two-dimensional annotated image; S14: Divide the number of pixels occupied by each brain anatomical structure in the two-dimensional annotated image corresponding to each two-dimensional brain image by the number of pixels occupied by the entire brain contour, calculate the area ratio occupied by each brain anatomical structure in the entire brain contour, and obtain a feature database of the two-dimensional brain image.

3. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 2, characterized in that: Each 3D brain image corresponds to a 3D annotated image with 672 brain anatomical structures marked. The pixels in the 3D annotated image correspond to the grayscale values ​​in the brain anatomical structure grayscale mapping table, and the grayscale value of each pixel represents the brain anatomical structure to which the pixel belongs.

4. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 2, characterized in that: The affine transformation refers to: rotating in the x, y, and z directions respectively, with the rotation angle not exceeding 15° and the step size of the angle change being 1°. Bilinear interpolation is used when performing affine transformation on the three-dimensional brain image, and nearest neighbor interpolation is used when performing affine transformation on the three-dimensional annotated image, so as to obtain the three-dimensional brain image and the three-dimensional annotated image after affine transformation; the grayscale stretching refers to: using adaptive histogram equalization to perform grayscale stretching on the three-dimensional brain image, wherein the contrast threshold of the equalization is set to 1.2 and the grid size of the equalization is 4*4.

5. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 2, characterized in that: The step S2 comprises: S21: Use convolutional neural network to train the two-dimensional brain image database and feature database to obtain feature extraction network; S22: Extracting features of the image to be located using a feature extraction network; S23: Compare the features of the image to be located with the distribution features of the brain anatomical structure in the feature database to obtain the most similar features and their corresponding detected images.

6. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 5, characterized in that: The most similar feature refers to: a brain anatomical structure distribution feature in the two-dimensional brain image database that has the smallest Euclidean distance to the feature of the image to be located.

7. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 2, characterized in that: In the step S31, there are 12 characteristic regions of the two-dimensional brain image, and their grayscales are 1-12 respectively. The characteristic regions are marked by grayscale extraction results, and the marked results are integrated on the marked image.

8. The method for locating brain anatomical structures applicable to two-dimensional brain image data according to claim 2, characterized in that: The two-dimensional brain image is sampled using bilinear interpolation before being stored in the two-dimensional brain image database, the two-dimensional annotated image is sampled using nearest neighbor interpolation before being stored in the two-dimensional brain image database, and the image to be located is sampled using bilinear interpolation before being input into the full convolution feature region segmentation network. The size of the sampled images is 512*512 pixel.

Images