A neural network-based brain image segmentation method and device

Abstract

The application provides a brain image segmentation method and device based on a neural network, which comprises the following steps: cutting a first brain image into a plurality of sub-blocks; wherein the first brain image is a three-dimensional brain image in a standard brain space; inputting each sub-block of the first brain image into a corresponding convolutional neural network model to obtain a brain region segmentation result corresponding to the sub-block; and fusing the brain region segmentation results of the plurality of sub-blocks of the first brain image to obtain a brain region segmentation result of the first brain image. According to the application, the first brain image is divided into blocks, so that each convolutional neural network can precisely segment the corresponding sub-block, and the number of layers of the convolutional neural network required is small due to the small volume of the sub-block, and the segmentation time is short; and the segmentation results of the plurality of sub-blocks of the first brain image are fused to accurately obtain the overall segmentation result of the first brain image, thereby realizing efficient and accurate segmentation of the brain image.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a brain image segmentation method and apparatus based on neural networks. Background Technology

[0002] The cerebral cortex is the most developed and sophisticated organ in the human body, governing all bodily activities and regulating the balance between the body and its environment; it is the foundation of higher nervous activity. Different neural centers in the brain control different physiological activities. For example, damage to the motor language center can cause motor aphasia, where the muscles related to speech are not paralyzed, but the patient cannot speak. Damage to the auditory language center can cause sensory aphasia, where the patient can hear others speaking but cannot understand what is being said. Therefore, precise division of brain regions is crucial for clinicians to assess surgical risks.

[0003] Currently, due to the difficulty of brain region segmentation, there are few clinical practices using this method. The common approach is to use classic image processing, utilizing image preprocessing, registration, and morphological processing, with Freesurfer being a representative technique. However, Freesurfer requires redundant surface reconstruction and other operations, and its processing time is long due to limitations in graphics card memory. For example, segmenting a single brain image using Freesurfer on a single computer typically takes several hours.

[0004] To address the shortcomings of low efficiency in brain region segmentation in existing technologies, this invention provides a brain image segmentation method and apparatus. Summary of the Invention

[0005] This invention provides a brain image segmentation method and apparatus to address the shortcomings of low efficiency in brain image segmentation in existing technologies.

[0006] This invention provides a brain image segmentation method based on neural networks, comprising:

[0007] The first brain image is divided into multiple sub-blocks; wherein, the first brain image is a three-dimensional brain image in standard brain space;

[0008] Each sub-block of the first brain image is input into the corresponding convolutional neural network model of the sub-block to obtain the corresponding brain region segmentation result of the sub-block;

[0009] The brain region segmentation results of multiple sub-blocks of the first brain image are fused together to obtain the brain region segmentation result of the first brain image.

[0010] This invention also provides a brain image segmentation method based on neural networks, comprising:

[0011] The first brain image is blurred to obtain the second brain image; wherein, the first brain image is a three-dimensional brain image in standard brain space;

[0012] The first brain image and the second brain image are respectively divided into multiple sub-blocks;

[0013] Each sub-block of the second brain image is input into the corresponding low-resolution convolutional neural network model to obtain the low-resolution brain region segmentation results of each sub-block of the second brain image.

[0014] Each sub-block of the first brain image and its corresponding low-resolution brain region segmentation results are input into the corresponding high-resolution convolutional neural network model and fused to obtain the high-resolution brain region segmentation results of the first brain image.

[0015] This invention also provides a brain image segmentation method based on neural networks, comprising:

[0016] The original brain images were preprocessed and rigidly registered with standard brain images to obtain registered brain images.

[0017] The registered brain images are input into a convolutional neural network for brain region segmentation to obtain segmented brain images in standard brain space.

[0018] The segmented brain images in the standard brain space are inversely registered to the original brain space to obtain the segmented brain images corresponding to the original brain images.

[0019] This invention provides a neural network-based brain image segmentation device, comprising:

[0020] The segmentation module is used to segment the first brain image into multiple sub-blocks; wherein, the first brain image is a three-dimensional brain image in standard brain space;

[0021] The segmentation module is used to input each sub-block of the first brain image into the corresponding convolutional neural network model of the sub-block to obtain the brain region segmentation result of the sub-block;

[0022] The fusion module is used to fuse the brain region segmentation results of multiple sub-blocks of the first brain image to obtain the brain region segmentation result of the first brain image.

[0023] The present invention also provides a brain image segmentation device based on a neural network, comprising:

[0024] The fuzzing module blurs the first brain image to obtain a second brain image; wherein the first brain image is a three-dimensional brain image in standard brain space.

[0025] The segmentation module is used to segment the first brain image and the second brain image into multiple sub-blocks respectively;

[0026] The second segmentation module is used to input each sub-block of the second brain image into the corresponding low-resolution convolutional neural network model to obtain the low-resolution brain region segmentation results of each sub-block of the second brain image.

[0027] The first segmentation module inputs each sub-block of the first brain image and its corresponding low-resolution brain region segmentation results into the corresponding high-resolution convolutional neural network model, and fuses them to obtain the high-resolution brain region segmentation results of the first brain image.

[0028] This invention also provides a brain image segmentation device based on a neural network, comprising:

[0029] The registration module is used to preprocess the original brain images and rigidly register them with standard brain images to obtain registered brain images.

[0030] The third segmentation module is used to input the registered brain image into a convolutional neural network for brain region segmentation to obtain segmented brain images in standard brain space.

[0031] The inverse registration module is used to inversely register the segmented brain images in the standard brain space to the original brain space to obtain the segmented brain images corresponding to the original brain images.

[0032] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement all or part of the steps of any of the neural network-based brain image segmentation methods described above.

[0033] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements all or part of the steps of the neural network-based brain image segmentation method described above.

[0034] The present invention also provides a computer program product comprising computer-executable instructions, which, when executed, implement all or part of the steps of any of the neural network-based brain image segmentation methods described above.

[0035] The present invention provides a brain image segmentation method and apparatus based on neural networks, which achieves the following technical effects:

[0036] 1. Provides convolutional neural network models with various architectures for automatic segmentation of brain images to meet the diverse needs of users;

[0037] 2. In the architecture of multiple convolutional neural networks, the first brain image is divided into blocks and each sub-block is input into the corresponding convolutional neural network model for brain region segmentation. This allows each convolutional neural network to perform targeted and accurate segmentation of the corresponding sub-block. Furthermore, since the sub-blocks are small in size, fewer layers of the convolutional neural network are required, resulting in less segmentation time. By fusing the segmentation results of multiple sub-blocks of the first brain image, the overall segmentation result of the first brain image is accurately obtained, achieving efficient and accurate segmentation of brain images.

[0038] 3. Standard brain maps and training samples can be flexibly selected, and the trained convolutional neural network is more suitable for the target population, and the brain region segmentation accuracy can be flexibly selected;

[0039] 4. Based on the basic training samples, simulated lesion samples were quickly obtained by increasing the coverage area, which improved the segmentation accuracy of the convolutional neural network for lesion regions in abnormal brain images. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating a neural network-based brain image segmentation method provided by the present invention.

[0042] Figure 2 This is a flowchart illustrating another brain image segmentation method based on neural networks provided by the present invention.

[0043] Figure 3 This is a flowchart illustrating another brain image segmentation method based on neural networks provided by the present invention.

[0044] Figure 4 This is a schematic diagram example of the structure of the convolutional neural network model in a neural network-based brain image segmentation method provided by the present invention;

[0045] Figure 5 This is one of the schematic diagrams of the brain region segmentation process in a neural network-based brain image segmentation method provided by the present invention;

[0046] Figure 6 This is one of the schematic diagrams of the brain region segmentation process in a neural network-based brain image segmentation method provided by the present invention;

[0047] Figure 7This is a schematic diagram of the structure of a neural network-based brain image segmentation device provided by the present invention;

[0048] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0049] Figure 9 This is the result of segmenting a real brain image using a brain image segmentation method according to an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0051] The following is combined with Figures 1-8 This invention describes a brain image segmentation method and apparatus based on a neural network. Figure 1 This is a flowchart illustrating a neural network-based brain image segmentation method provided by the present invention, as shown below. Figure 1 As shown, the method includes:

[0052] S11. Divide the first brain image into multiple sub-blocks; wherein, the first brain image is a three-dimensional brain image in standard brain space;

[0053] Specifically, this invention performs brain region segmentation on three-dimensional brain images to efficiently and accurately determine the brain region label of each pixel in the three-dimensional brain image.

[0054] The first brain image is a three-dimensional brain image in standard brain space, that is, a three-dimensional brain image registered to a standard brain image in a standard brain coordinate system. A standard brain image refers to a statistically obtained average brain image, for example, using multiple head images of the same resolution (usually 1*1*1mm). 3 The average brain regions are then cropped from the averaged brain regions to obtain a standard brain image (i.e., removing areas such as the neck from the head image and placing the brain region in the center). Standard brain images can be Bordmann atlases, Talairach brain images, MNI305 brain images, WholeBrain brain images, VOXEL-MAN brain images, BrainWeb brain images, etc. Specific standard brain images can be flexibly selected or statistically generated, for example, based on brain region precision (number of brain region labels) or target population type (skin color, nationality, gender, age group, etc.).

[0055] Brain images are configured in a standard brain space, with uniform pose and size. This ensures that each convolutional neural network (CNN) only needs to learn features of specific brain regions and process local brain images at those locations, improving brain image segmentation efficiency. The first brain image in the standard brain space is divided into multiple sub-blocks according to a preset segmentation strategy. This strategy could involve moving a cube along the axis of the standard brain space coordinate system with a preset step size to gradually segment into multiple sub-blocks, or moving a cube or other shaped cube to segment into multiple sub-blocks. The purpose of segmenting into sub-blocks is to reduce the number of pixels each CNN needs to process, thereby reducing the number of CNN model layers, accelerating convergence during model training, and effectively reducing the time spent on brain region segmentation during model application, thus lowering the requirements for GPU memory. It is understood that to ensure a complete brain region segmentation result for the first brain image, each pixel in the first brain image must exist in at least one sub-block, which can be achieved by limiting the preset step size. The brain image is then segmented into multiple sub-blocks, which are then input into the corresponding CNN models for processing.

[0056] S12. Input each sub-block of the first brain image into the corresponding convolutional neural network model of the sub-block to obtain the corresponding brain region segmentation result;

[0057] Specifically, in this invention, each sub-block of the brain image corresponds to a specific convolutional neural network (CNN) model. The correspondence between sub-blocks and CNNs is as follows: each sub-block corresponds to a separate CNN; or, sub-blocks with the same coordinate attribute are input into the same CNN model (for example, adjacent sub-blocks within a local area of ​​the brain image are input into the same CNN model; or, by dividing the sub-blocks into intervals containing the pixel horizontal coordinates, sub-blocks in the left brain are input into one CNN model, and sub-blocks in the right brain are input into another CNN model). Each sub-block of the first brain image is input into its corresponding CNN model, and brain region segmentation is performed on that sub-block to obtain the corresponding brain region segmentation result. In this invention, brain regions refer to areas divided according to brain function (e.g., primary motor cortex, somatosensory association cortex, premotor cortex, primary visual cortex, etc.). Accordingly, the brain region segmentation result includes a brain region classification label corresponding to each pixel in the sub-block. Considering that brain images may contain invalid regions (e.g., tumor pixels, skull pixels, hemorrhage pixels), invalid regions do not correspond to actual brain regions. Their corresponding brain region classification labels can be set to be the same as the brain region classification labels of the background pixels, for example, set to "0".

[0058] The convolutional neural network model is pre-trained. Specifically, it can be obtained as follows: First, brain image samples and their corresponding brain region classification reference labels are acquired. This can be done by directly obtaining existing brain image samples and brain region classification reference labels, or by having experts specifically annotate the brain image samples to provide brain region classification reference labels for each region in the brain image samples. The number of labels can be flexibly set according to the classification accuracy, for example, set to 125 categories. The brain image samples and their corresponding brain region classification reference labels are also segmented into multiple sub-block samples and the sub-block segmentation results (reference classification labels) for each sub-block sample, and then used to train the corresponding convolutional neural network for each sub-block sample.

[0059] S13. The segmentation results of multiple sub-blocks of the first brain image are fused to obtain the fused segmentation result of the first brain image.

[0060] Specifically, each sub-block of the first brain image contains only a portion of the pixels of the first brain image. Correspondingly, the segmentation result of the sub-block also only contains the local segmentation result of the first brain image. It is necessary to fuse the segmentation results of each sub-block. Since there may be overlapping areas between adjacent sub-blocks, it is necessary to determine the fusion label corresponding to the pixels in the overlapping area during the fusion process.

[0061] In this embodiment, the first brain image is a three-dimensional brain image in a standard brain space, which places the first brain image and the training samples of the convolutional neural network in a unified brain space, providing a basis for block segmentation and brain region segmentation. By inputting each sub-block of the first brain image into the corresponding convolutional neural network model for brain region segmentation, each convolutional neural network can perform targeted and accurate segmentation of the corresponding sub-block. Furthermore, due to the small size of the sub-blocks, fewer layers of the convolutional neural network are required, resulting in less segmentation time. By accurately fusing the segmentation results of multiple sub-blocks of the first brain image, the overall segmentation result of the first brain image is obtained, achieving efficient and rapid segmentation of the brain image.

[0062] Based on any of the above embodiments, in one embodiment, each sub-block of the first brain image corresponds to a convolutional neural network model.

[0063] Each specific sub-block corresponds to a convolutional neural network model, which can process brain imaging data at the corresponding location, thereby improving the speed and accuracy of brain region segmentation.

[0064] Based on any of the above embodiments, in one embodiment, the convolutional neural network is specifically a U-net network or a fully convolutional neural network.

[0065] Specifically, U-Net is a type of convolutional neural network that includes a contraction path for feature extraction and an expansion path using upsampling. During the upsampling stage, features extracted in the feature extraction (downsampling) stage are added, allowing more information about the original image texture to propagate in high-resolution layers (hidden layers). U-Net can achieve accurate segmentation results with very little training data. Fully convolutional neural networks replace the fully connected classification layers in ordinary convolutional neural networks with convolutional layers, upsampling deep features to restore the original image size and classifying each pixel at the pixel level.

[0066] In this embodiment, the segmentation accuracy of brain images can be improved by using U-net network or fully convolutional neural network.

[0067] Figure 2 This is a flowchart illustrating another neural network-based brain image segmentation method provided by the present invention, as shown below. Figure 2 As shown, the method includes:

[0068] S21. Blur the first brain image to obtain a second brain image; wherein, the first brain image is a three-dimensional brain image in standard brain space.

[0069] Specifically, blurring can reduce the resolution of brain images. In this invention, low-resolution brain images can be quickly segmented into low-resolution segmentation results, which can be used to improve the segmentation accuracy of high-resolution brain images. The blurring process does not change the size of the brain image; it only reduces the image quality, making the pixels of the second brain image reference the values ​​of surrounding pixels more. Specifically, the blurring process can use a Gaussian kernel to perform Gaussian distribution weighting on the high-resolution first brain image to obtain the corresponding low-resolution second brain image.

[0070] Furthermore, in order to maintain the size of the second brain image after blurring and preserve the quality of the edge pixels of the brain image, pixels can be added to the outer edge of the first brain image in a mirror-symmetric manner before blurring.

[0071] It should be noted that "high resolution" and "low resolution" in this invention do not mean that the resolution of brain images is within a specific resolution range. In this invention, "high resolution" and "low resolution" are relative concepts, that is, high-resolution brain images have higher image quality and contain more detailed features than low-resolution brain images.

[0072] S22. Divide the first brain image and the second brain image into multiple sub-blocks respectively;

[0073] Specifically, both the first and second brain images are segmented into blocks. The preset segmentation strategy can be that the first and second brain images adopt a uniform segmentation strategy. In this case, each sub-block in the first brain image corresponds one-to-one with each sub-block in the second brain image, and the corresponding blocks have the same position and size. Alternatively, the preset segmentation strategy can be that the first and second brain images adopt different specific segmentation strategies. In this case, the number of sub-blocks in the first and second brain images may be different, and correspondingly, the number of low-resolution convolutional neural network models required by the scheme may differ from the number of high-resolution convolutional neural network models required. Specific segmentation strategies can be referred to in the aforementioned embodiments, and will not be repeated here.

[0074] S23. Input each sub-block of the second brain image into the corresponding low-resolution convolutional neural network model to obtain the low-resolution brain region segmentation results of each sub-block of the second brain image.

[0075] Specifically, the correspondence between sub-blocks of the second brain image and low-resolution convolutional neural network models is as follows: each sub-block corresponds to a separate low-resolution convolutional neural network; or, sub-blocks with the same coordinate attribute are input into the same low-resolution convolutional neural network model (for example, adjacent sub-blocks within a local area of ​​the brain image are input into the same convolutional neural network; or, by dividing the sub-blocks into intervals where the pixel horizontal coordinates lie, sub-blocks in the left brain are input into one low-resolution convolutional neural network model, and sub-blocks in the right brain are input into another low-resolution convolutional neural network model). Each sub-block of the second brain image is input into the corresponding low-resolution convolutional neural network model for brain region segmentation, resulting in low-resolution brain region segmentation results for each sub-block. It should be noted that "low-resolution" in "low-resolution convolutional neural network" is not a limitation on the attributes of the convolutional neural network, but rather a conceptual distinction from "high-resolution convolutional neural network" mentioned below. The "low-resolution convolutional neural network" in this invention is trained using training samples with relatively low resolution. The training samples include low-resolution brain image sub-block samples corresponding to the location, and the low-resolution segmentation results corresponding to the sub-block samples. The sub-block sample contains a series of three-dimensional pixels, and the corresponding low-resolution segmentation result contains the brain region classification reference label for each pixel. Considering that brain images may contain invalid regions (such as tumor pixels, skull pixels, hemorrhage pixels), which do not correspond to actual brain regions, their corresponding brain region classification labels can be set to be the same as the brain region classification labels of the background pixels, for example, set to "0".

[0076] S24. Input each sub-block of the first brain image and its corresponding low-resolution brain region segmentation results into the corresponding high-resolution convolutional neural network model, and fuse them to obtain the high-resolution brain region segmentation results of the first brain image.

[0077] Specifically, the second brain image is a low-resolution brain image of the first brain image, and the low-resolution brain region segmentation results of the second brain image can be used in the brain region segmentation process of the first brain image.

[0078] The low-resolution brain region segmentation results of each sub-block in the second brain image can be directly used as the corresponding low-resolution brain region segmentation results of each sub-block in the first brain image. In this case, it is necessary to ensure that the first and second brain images are segmented using the same segmentation strategy, i.e., each sub-block corresponds one-to-one. Alternatively, the low-resolution brain region segmentation results of each sub-block in the second brain image can be fused, and then the corresponding low-resolution brain region segmentation results of each sub-block in the first brain image can be determined from the overall fused brain region segmentation result of the second brain image. In this case, the first and second brain images can be segmented using the same segmentation strategy or different segmentation strategies. The fusion process involves determining the comprehensive brain region classification label for each pixel. For example, for overlapping pixels, the fused brain region segmentation label is determined based on the brain region segmentation label of the pixel in each sub-block. By fusing the brain region segmentation results of sub-blocks, the segmentation accuracy of sub-block edges can be improved.

[0079] Similarly, the correspondence between sub-blocks of the first brain image and high-resolution convolutional neural network models is as follows: each sub-block corresponds to a high-resolution convolutional neural network; or, sub-blocks with the same coordinate attribute are input into the same high-resolution convolutional neural network model (for example, adjacent sub-blocks within a local area of ​​the brain image are input into the same convolutional neural network; or, by dividing the interval of the pixel horizontal coordinates in the sub-block, sub-blocks in the left brain are input into one high-resolution convolutional neural network model, and sub-blocks in the right brain are input into another high-resolution convolutional neural network model). Each sub-block and its corresponding low-resolution brain region segmentation result are input into the high-resolution convolutional neural network corresponding to that sub-block to obtain the high-resolution brain region segmentation result of that sub-block. Then, the brain region segmentation results of each sub-block are fused to obtain the high-resolution brain region segmentation result of the first brain image.

[0080] The "high-resolution convolutional neural network" in this invention is trained using relatively high-resolution training samples. These training samples include high-resolution brain image sub-block samples corresponding to the given location, low-resolution segmentation results corresponding to the sub-block samples, and high-resolution segmentation results corresponding to the sub-block samples. Each sub-block sample contains a series of three-dimensional pixels, and the corresponding low / high-resolution segmentation results contain low / high-resolution brain region classification reference labels for each pixel.

[0081] In this embodiment, the first brain image is blurred to obtain the second brain image, resulting in two sets of brain images with different resolutions but related brain region segmentation results. By inputting each sub-block of the second brain image into the corresponding convolutional neural network for rapid segmentation, a low-resolution brain region segmentation result is obtained. Since each sub-block corresponds to a pre-trained convolutional neural network with a small number of network layers, the sub-blocks can be segmented quickly. By inputting the low-resolution brain region segmentation result into the high-resolution convolutional neural network of the corresponding sub-block in the first brain image, strong support is provided for the accurate segmentation of the high-resolution first brain image. Overall, this invention achieves rapid and accurate segmentation of the first brain image.

[0082] Based on any of the above embodiments, in one embodiment, each sub-block of the second brain image corresponds to a low-resolution convolutional neural network model, and each sub-block of the first brain image corresponds to a high-resolution convolutional neural network model.

[0083] Specifically, each sub-block corresponds to a convolutional neural network model, which can process brain imaging data at the corresponding location, thereby improving the speed and accuracy of brain region segmentation.

[0084] Based on the above embodiments, in one embodiment, the step of inputting each sub-block of the second brain image into a corresponding low-resolution convolutional neural network model to obtain low-resolution brain region segmentation results for each sub-block of the second brain image; and inputting each sub-block of the first brain image and its corresponding low-resolution brain region segmentation results into a corresponding high-resolution convolutional neural network model, and fusing them to obtain high-resolution brain region segmentation results for the first brain image, includes:

[0085] Each sub-block of the second brain image is input into the low-resolution convolutional neural network corresponding to the location of the sub-block to obtain the corresponding low-resolution brain region segmentation result of the sub-block.

[0086] The low-resolution brain region segmentation results of multiple sub-blocks of the second brain image are fused to obtain the brain region segmentation result of the second brain image.

[0087] Each sub-block of the first brain image, and the corresponding low-resolution brain region segmentation result of the sub-block in the brain region segmentation result of the second brain image, are input into the high-resolution convolutional neural network corresponding to the location of the sub-block to obtain the corresponding high-resolution brain region segmentation result of the sub-block.

[0088] The high-resolution brain region segmentation results of multiple sub-blocks of the first brain image are fused to obtain the brain region segmentation result of the first brain image.

[0089] Specifically, each sub-block of the low-resolution second brain image is segmented into brain regions using a corresponding convolutional neural network, and the segmentation results are then fused. From the overall segmentation result of the second brain image, the fused brain region segmentation result of the corresponding sub-block is extracted (equivalent to further segmenting the overall segmentation result of the second brain image according to the segmentation strategy of the first brain image, obtaining the corresponding low-resolution brain region segmentation result of the first brain image sub-block). This result, along with the corresponding sub-block from the first brain image, is then input into the high-resolution convolutional neural network corresponding to that sub-block for brain region segmentation, obtaining the high-resolution brain region segmentation result of that sub-block. Finally, the high-resolution brain region segmentation results of each sub-block are fused to obtain the high-resolution brain region segmentation result of the first brain image. Since each high / low-resolution convolutional neural network processes local sub-blocks of the brain image, the amount of data processing is relatively small, requiring no deep model layers, and its training and segmentation application processes can be performed efficiently. Simultaneously, by fusing the brain region segmentation results of sub-blocks, the segmentation accuracy of sub-block edges can be improved, and the segmentation accuracy of the high-resolution brain image can be improved through the low-resolution brain region segmentation results.

[0090] Based on any of the above embodiments, in one embodiment, the fusion process is implemented in the following manner:

[0091] Traverse the brain images to determine the current pixel to be fused, and the brain region segmentation label of the current pixel to be fused in each sub-block;

[0092] Based on the segmentation labels of the current pixel to be fused in each sub-block of the brain image, the segmentation label with the most occurrences is determined as the fusion label of the pixel to be fused by majority voting.

[0093] Repeat the above process until all pixels in the brain image are fused.

[0094] Specifically, the fusion label is determined pixel by pixel in the brain image. Since a pixel may appear in different sub-blocks, it has multiple segmentation labels. For the current pixel to be fused, its brain region segmentation label in each sub-block is determined (if it is not in a certain sub-block, that sub-block is ignored). The brain region segmentation labels of the current pixel to be fused are statistically analyzed, and its fusion label is determined by majority voting. For example, if the current pixel to be fused, 'a', appears in 6 sub-blocks, with brain region segmentation labels 31, 32, 32, 31, 33, and 31 respectively, label "31" receives the most votes, and "31" is determined as the fusion label for the current pixel to be fused, '31'. Furthermore, if the current pixel to be fused has more than one segmentation label with the highest number of segments, its fusion label is determined from the multiple segments with the highest number of segments among its surrounding pixels. Specifically, the pixels with already determined fusion labels can be statistically analyzed, and the pixel with the highest number of segments among its surrounding pixels is determined as the fusion label for that pixel. For example: The number of "32" and "33" labels of the current pixel to be merged is 3 (the maximum number), and it is determined that 10 pixels in the surrounding 26 pixels have already been identified as having merge labels. Among these 10 pixels, 3 have the merge label "32" and 1 has the merge label "33". Based on the merge labels of the surrounding pixels, it can be determined that the merge label of the current pixel to be merged is "32".

[0095] In this embodiment, the labels of pixels in the brain image are accurately determined by fusing the segmentation results.

[0096] Based on any of the above embodiments, in one embodiment, the sum of the overlapping areas of each sub-block and its adjacent sub-blocks accounts for a proportion greater than a first proportion threshold of each sub-block.

[0097] Specifically, adjacent sub-blocks within multiple sub-blocks overlap to improve the brain region segmentation accuracy of pixels in the sub-block edge regions. A preset segmentation strategy is used to ensure that the sum of the overlapping regions of each sub-block with its multiple adjacent sub-blocks accounts for a proportion greater than a first proportion threshold. This first proportion threshold is set according to requirements, for example, to 50%. It is understood that the higher the first proportion threshold, the higher the brain region segmentation accuracy. The sum of the overlapping regions here refers to determining the overlapping regions between the sub-block and each of its adjacent sub-blocks, and then taking the union of these overlapping regions.

[0098] In this embodiment, the edge region segmentation accuracy of each sub-block is improved by setting a first proportional threshold.

[0099] Based on any of the above embodiments, in one embodiment, the high-resolution convolutional neural network and the low-resolution convolutional neural network are specifically U-net networks; or,

[0100] The high-resolution convolutional neural network and the low-resolution convolutional neural network are specifically fully convolutional neural networks.

[0101] Specifically, high-resolution and low-resolution convolutional neural networks can be U-net networks. U-net is a type of convolutional neural network that includes a contraction path for feature extraction and an expansion path using upsampling. In the upsampling stage, features extracted in the feature extraction (downsampling) stage are added, allowing more information about the original image texture to propagate in high-resolution layers. U-net can obtain accurate segmentation results with very few training images. High-resolution and low-resolution convolutional neural networks can also be fully convolutional neural networks. Fully convolutional neural networks replace the fully connected classification layers in ordinary convolutional neural networks with convolutional layers, upsampling deep features to restore the original image size, facilitating classification of each pixel at the pixel level.

[0102] In this embodiment, the accuracy of brain image segmentation can be improved by using U-net network or fully convolutional neural network.

[0103] Based on the above embodiments, in one embodiment, the method further includes: preprocessing the original brain image and registering it with a standard brain image to obtain the first brain image in a standard brain space.

[0104] Specifically, the raw brain image is a three-dimensional brain image to be processed, acquired from MRI medical images (e.g., T1, T2, FLAIR, T1ce images). Due to factors such as human error, equipment limitations, and environmental factors, the raw brain image may contain noise interference, requiring preprocessing such as image denoising and removal of redundant regions (e.g., the neck). The first brain image is a three-dimensional brain image in a standard brain space (coordinate system). Registering the raw brain image to be processed with the standard brain image in the standard brain space (coordinate system) yields the first brain image. This registration process is rigid and can include operations such as translation, rotation, and scaling. Similarly, the sub-blocks in the training samples of the aforementioned convolutional neural networks / low-resolution convolutional neural networks / high-resolution convolutional neural networks are also sub-blocks segmented from the brain image samples registered to the standard brain space. This ensures that the training sample sub-blocks of the convolutional neural network corresponding to each sub-block's location have similar structures to the sub-blocks to be segmented in the brain region, guaranteeing segmentation accuracy.

[0105] In this embodiment, the original brain images are preprocessed and registered to a standard brain space to obtain the first brain image. This ensures that the first brain image and the training samples of the convolutional neural network are in the same brain space, providing a basis for block segmentation and brain region segmentation, and guaranteeing the accuracy of brain image segmentation.

[0106] Based on the above embodiments, in one embodiment, the preprocessing includes at least one of the following:

[0107] Noise reduction, offset field correction, and grayscale normalization.

[0108] Specifically, images may experience abrupt changes in pixel grayscale during acquisition and transmission. To reduce the impact of these changes on subsequent processing, denoising can be performed on 3D brain images. For example, an anisotropic curvature smoothing filter can be used to denoise 3D brain images. This filter can smooth image noise while preserving as much of the original image's feature information as possible. Furthermore, the anisotropic curvature smoothing filter is faster than Gaussian smoothing. The denoising process requires referencing surrounding pixels to process the current pixel. Unlike the blurring process described earlier, the range of reference pixels in denoising is smaller, and the variance between the weights of the referenced surrounding pixels is smaller. Its effect is to minimize the influence of abnormal pixels on image features while preserving as much of the original image's feature information as possible.

[0109] Furthermore, during image acquisition, factors such as the patient's position within the scanner and uneven magnetic field strength within the scanner itself can cause brightness differences in medical images, which can severely impact segmentation accuracy. To address this, offset field correction can be applied to the brain images to avoid interfering with segmentation accuracy.

[0110] Brain images can also be normalized in grayscale, scaling the grayscale of images acquired by different devices to 0-1 to facilitate subsequent applications.

[0111] In this embodiment, noise reduction, offset field correction, and grayscale normalization are used to improve data processing efficiency and brain image segmentation accuracy.

[0112] Based on any of the above embodiments, in one embodiment, the step of preprocessing the original brain image and registering it with a standard brain image to obtain the first brain image in standard brain space includes:

[0113] Using the three-dimensional region of the standard brain image, the registered brain image after being registered to the standard brain space is cut to obtain the first brain image.

[0114] Specifically, the standard brain image includes a brain region mask, which shows the three-dimensional extent of the standard brain image. The registered brain image, after being registered to the standard brain space, is segmented using the brain region mask to obtain the first brain image. The first brain image, compared to the original brain image, removes regions outside the brain region mask. Depending on the specific brain region mask used, effects such as "removing the neck," "removing the skull," and "removing the mandible" can be achieved. Furthermore, during brain region segmentation, pixels outside the brain region mask are given a separate label, representing invalid regions, for example, a label of "0".

[0115] In this embodiment, non-interested regions in the original brain images were removed, improving data processing efficiency and enhancing the visual effect of brain region segmentation in brain images.

[0116] Based on any of the above embodiments, in one embodiment, the method further includes:

[0117] The brain region segmentation results of the first brain image are inversely registered to the original brain space to obtain the brain region segmentation results of the original brain image.

[0118] Specifically, the first brain image is a three-dimensional brain image registered to the standard brain space. After the first brain image is segmented into brain regions, the brain region segmentation results of the first brain image can be further inversely registered to the original brain space (coordinate system) based on the registration relationship in the registration process to obtain the original brain image after brain region segmentation.

[0119] This embodiment yields brain region segmentation results from the original brain images in the original brain space, facilitating intuitive observation by doctors and providing information support for the diagnosis and treatment process.

[0120] Figure 3 This is a flowchart illustrating another neural network-based brain image segmentation method provided by the present invention, as shown below. Figure 3 As shown, the method includes:

[0121] S31. The original brain images are preprocessed and rigidly registered with standard brain images to obtain registered brain images.

[0122] S32. Use a convolutional neural network to segment the registered brain images to obtain segmented brain images in standard brain space.

[0123] S33. Inversely register the segmented brain image in the standard brain space to the original brain space to obtain the segmented brain image corresponding to the original brain image.

[0124] Specifically, the raw brain images are three-dimensional brain images to be processed, obtained from MRI medical images (such as T1, T2, Flair, T1ce, etc.). Due to human factors, equipment, environment and other factors, there may be noise interference in the raw brain images, so preprocessing of the raw brain images is required, such as image denoising and offset field correction.

[0125] The preprocessed original brain image is registered with a standard brain image in standard brain space (coordinate system). The registration process aims to maximize the similarity between the original and standard brain images, facilitating brain region segmentation by the convolutional neural network. During registration, registration parameters are iteratively adjusted, using rigid registration methods such as translation, rotation, and scaling to ensure the similarity between the brain regions in the preprocessed brain image and the standard brain image reaches a similarity threshold. Then, a mask of brain regions from the standard brain image is used to extract the registered brain image, which is a 3D brain image in standard brain space with redundant regions removed. This registered brain image is then input into the convolutional neural network for automatic segmentation, resulting in a segmented brain image. The segmented brain image contains a brain region classification label for each pixel.

[0126] Convolutional neural networks are pre-trained using training samples and sample labels. It is understood that the training samples also need to be registered to the standard brain space. The pre-trained convolutional neural network can be used to segment brain regions from brain images to be segmented to obtain specific brain region segmentation results. Examples of convolutional neural networks include fully convolutional neural networks and U-net networks.

[0127] After obtaining the segmented brain images, the segmented brain images are inversely registered to the original brain space (coordinate system) according to the registration relationship in the aforementioned registration process, so as to obtain the segmented brain images corresponding to the original brain images, which are convenient for doctors to observe.

[0128] In this embodiment, noise interference is reduced and redundant data is removed through preprocessing; by registering the original brain images to a standard brain space, the original brain images and the training samples of the convolutional neural network model are on the same basis, enabling the convolutional neural network model to have a transfer basis and automatically segment brain images; by inversely registering the brain region segmentation results of the registered brain images to the original brain space, segmented brain images in the original brain space are obtained, which are convenient for doctors to observe intuitively and provide information support for the diagnosis and treatment process.

[0129] Based on any of the above embodiments, in one embodiment, the training samples of the convolutional neural network include amplified samples, which are simulated lesion samples obtained by adding coverage areas on the basis of the basic training samples.

[0130] Specifically, current automated brain segmentation systems on the market are all designed for segmenting normal human brains, and their accuracy is low when segmenting lesion regions in abnormal brains. To address this, lesion samples can be added to the training samples. Considering that 3D brain images of normal individuals are relatively easy to obtain, while abnormal brain image data is scarce and difficult to acquire, this embodiment adds a coverage area to the basic training samples. This added coverage area simulates lesion regions (e.g., tumor areas, hemorrhage areas), and sets corresponding reference labels for the coverage area, for example, using the same reference labels as invalid regions outside brain tissue (i.e., invalid regions).

[0131] In this embodiment, simulated lesion samples were quickly obtained by adding a coverage area on the basis of basic training samples, which improved the segmentation accuracy of the convolutional neural network for lesion regions in abnormal brain images.

[0132] The following describes the neural network-based brain image segmentation method provided by the present invention through a preferred embodiment.

[0133] STEP 1: Acquire raw brain images and preprocess them.

[0134] 1.1 Generate raw brain images based on the patient's head images (such as T1 images). Raw brain images usually include tissues such as the neck, with varying resolutions, such as 0.5*0.5*0.8mm3, and different image sizes, such as 512*512*256. Such raw images are not easy to process uniformly.

[0135] 1.2 Preprocess the raw brain images. Preprocessing can be selected from the following processing options.

[0136] (1) Denoise the original brain images. Specifically, an anisotropic curvature smoothing filter can be used, which can smooth image noise while preserving the feature information of the original image as much as possible.

[0137] Images may experience abrupt changes in pixel grayscale during acquisition and transmission. To mitigate the impact of these changes on subsequent processing, an anisotropic curvature smoothing filter is employed. Compared to Gaussian smoothing, this filter offers significantly faster processing speeds while better preserving image details.

[0138] (2) Correct the offset field of the denoised image. During image acquisition, factors such as the patient's position in the scanner and the scanner itself can cause brightness differences in the MRI image.

[0139] (3) Normalize the grayscale of the images obtained in the previous steps and scale the grayscale of all images acquired by different devices to 0 to 1 to facilitate the application of subsequent steps.

[0140] STEP 2: Register the preprocessed raw brain images to a standard brain space.

[0141] 2.1 Obtain standard brain images.

[0142] Standard Brain Imaging I s You can select according to your actual needs, for example, select Bordmann brain images with 125 partitions, with a size of 256*256*192 and a resolution of 1*1*1mm. 3 .

[0143] A standard brain refers to the average brain image obtained statistically, acquired as follows: using multiple brain images of the same resolution (usually 1*1*1mm). 3 The average is calculated, and the final result is the average brain region (i.e., the neck region is removed and the brain region is placed in the center).

[0144] 2.2 Register the preprocessed raw brain images to the standard brain space.

[0145] The registration process is rigid registration, which includes steps such as rotation, translation, and scaling. Here, grayscale-based feature extraction is selected, and the formula for the similarity evaluation metric is as follows:

[0146]

[0147] The optimization strategy chosen is gradient descent to quickly find the optimal solution. In the formula, x is a generic term referring to the grayscale value of the corresponding pixel in the standard brain image and the intermediate image during the registration process. Ω0 in the formula refers to the size range of the standard brain image. After the steps are completed, the registered brain image I1 and the image registered from the original brain image I0 to the standard brain image I1 are obtained. s The registration relationship (i.e., the transformation matrix) T.

[0148] 2.3 Use standard brain region masks to register brain images.

[0149] I1 was segmented using a standard brain region mask to obtain the first brain image, ensuring that the images processed in subsequent segmentation steps only contain brain regions, thereby improving segmentation accuracy.

[0150] STEP 3: Brain imaging brain region segmentation / model training process.

[0151] (1) Use a single convolutional neural network to segment brain regions in brain images.

[0152] Reference Figure 4 The brain image segmentation is illustrated using a single U-net network, such as... Figure 3As shown, U-Net contains a 4-layer structure (this is just an example; it can also have 5 or more layers). After registration, the image is input into the model and undergoes three downsampling operations, three upsampling operations, and finally output to the label field after a 1×1×1 convolution. The three downsampling operations consist of two 3×3×3 convolution operations and one 2×2×2 pooling operation; the three upsampling operations consist of one 2×2×2 deconvolution operation and two 3×3×3 convolution operations. In the upsampling stage, features extracted in the feature extraction (downsampling) stage are added through skip connections to improve the accuracy of brain region segmentation.

[0153] It should be noted that the above convolution process does not change the image size. Specifically, pixels can be added at the edges of sub-blocks using a mirror-symmetric method before performing the convolution operation. The purpose is to maintain the image size, facilitating brain region classification for each pixel at the pixel level. Furthermore, the number of output categories in the final label field depends on the chosen atlas. For example, the Brodmann atlas outputs 125 categories of labels.

[0154] The training process for convolutional neural network models requires inputting reference brain region classification labels along with brain images during the model training phase. Those skilled in the art can understand the specific training process based on the above content, so it will not be elaborated further here.

[0155] (2) Use multiple convolutional neural networks to segment brain regions in brain images.

[0156] Reference Figure 5 This is illustrated by using multiple U-net networks for brain image segmentation, such as... Figure 5 As shown, 27 U-Nets (3×3×3, or other cubic numbers, the specific number is not shown in the figure) are required. First, the original image is divided into 27 overlapping sub-blocks. Different U-Net networks are used to segment the brain regions of each sub-block to obtain the brain region segmentation results of each sub-block. Then, the brain region segmentation results of each sub-block are fused. For overlapping regions, the fusion label is determined by majority voting.

[0157] For the model training process of each convolutional neural network, it is necessary to input the brain region classification reference labels of the sub-blocks along with the sub-block brain images during the model training stage for model training. Those skilled in the art can understand the specific training process based on the above content, so it will not be elaborated here.

[0158] (3) Use a dual-resolution cascaded architecture to segment brain regions from brain images.

[0159] The first brain image to be segmented is blurred to obtain a low-resolution second brain image. The second brain image is processed by multiple convolutional neural networks in the above architecture (2) to obtain the brain region segmentation result of the second brain image.

[0160] For high-resolution first brain imaging, such as Figure 6 As shown, the same segmentation strategy is used to divide the data into blocks. For each sub-block, the low-resolution brain region segmentation result of the sub-block and the corresponding second brain image is input into the corresponding (high-resolution) convolutional neural network model to obtain the (high-resolution) brain region segmentation result of the sub-block. Then, the brain region segmentation results of each sub-block are fused to obtain the final brain region segmentation result of the first atlas. The fusion process is the same as above.

[0161] For the training process of high-resolution convolutional neural network models, it is necessary to input the low-resolution brain region segmentation results corresponding to the sub-block and the (high-resolution) brain region classification reference labels of the sub-block along with the sub-block brain images during the model training stage, in order to train the model. Those skilled in the art can understand the specific training process based on the above content, and it will not be elaborated here.

[0162] STEP4, Inverse Registration.

[0163] Based on the registration relationship T in the registration process, the brain region segmentation result of the first brain image is inversely registered to the original brain space to obtain the brain region segmentation result corresponding to the original brain image.

[0164] This embodiment uses a single convolutional neural network to automatically segment brain regions from brain images. Multiple convolutional neural networks are used to efficiently segment the brain map through block cutting and fusion, while reducing training overhead. A cascaded dual-resolution convolutional neural network further improves the accuracy of brain image segmentation.

[0165] The results of segmenting a real brain image using the brain image segmentation method of this embodiment are shown in the appendix. Figure 9 .

[0166] The neural network-based brain image segmentation device provided by the present invention will be described below. The neural network-based brain image segmentation device described below and the neural network-based brain image segmentation method described above can be referred to and correspond to each other.

[0167] Figure 7 This is a schematic diagram of the structure of a neural network-based brain image segmentation device provided by the present invention, as shown below. Figure 7 As shown, the device includes:

[0168] The segmentation module 71 is used to segment the first brain image into multiple sub-blocks; wherein, the first brain image is a three-dimensional brain image in standard brain space;

[0169] The segmentation module 72 is used to input each sub-block of the first brain image into the convolutional neural network model corresponding to the location of the sub-block to obtain the corresponding brain region segmentation result of the sub-block;

[0170] The fusion module 73 is used to fuse the brain region segmentation results of multiple sub-blocks of the first brain image to obtain the brain region segmentation result of the first brain image.

[0171] Based on the above embodiments, in one embodiment, the convolutional neural network is specifically a U-net network or a fully convolutional neural network.

[0172] This invention also provides another neural network-based brain image segmentation device, the device comprising:

[0173] The fuzzing module blurs the first brain image to obtain a second brain image; wherein the first brain image is a three-dimensional brain image in standard brain space.

[0174] The segmentation module is used to segment the first brain image and the second brain image into multiple sub-blocks respectively;

[0175] The second segmentation module is used to input each sub-block of the second brain image into the corresponding low-resolution convolutional neural network model to obtain the low-resolution brain region segmentation results of each sub-block of the second brain image.

[0176] The first segmentation module inputs each sub-block of the first brain image and its corresponding low-resolution brain region segmentation results into the corresponding high-resolution convolutional neural network model, and fuses them to obtain the high-resolution brain region segmentation results of the first brain image.

[0177] Based on the above embodiments, in one embodiment, the second segmentation module is used for:

[0178] Each sub-block of the second brain image is input into the low-resolution convolutional neural network corresponding to the location of the sub-block to obtain the corresponding low-resolution brain region segmentation result of the sub-block.

[0179] The low-resolution brain region segmentation results of multiple sub-blocks of the second brain image are fused to obtain the brain region segmentation result of the second brain image.

[0180] The first segmentation module is used for:

[0181] Each sub-block of the first brain image, and the corresponding low-resolution brain region segmentation result of the sub-block in the brain region segmentation result of the second brain image, are input into the high-resolution convolutional neural network corresponding to the location of the sub-block to obtain the corresponding high-resolution brain region segmentation result of the sub-block.

[0182] The high-resolution brain region segmentation results of multiple sub-blocks of the first brain image are fused to obtain the brain region segmentation result of the first brain image.

[0183] Based on any of the above embodiments, in one embodiment, the fusion process is implemented in the following manner:

[0184] Traverse the brain images to determine the current pixel to be fused, and the brain region segmentation label of the current pixel to be fused in each sub-block;

[0185] Based on the segmentation labels of the current pixel to be fused in each sub-block of the brain image, the segmentation label with the most occurrences is determined as the fusion label of the pixel to be fused by majority voting.

[0186] Repeat the above process until all pixels in the brain image are fused.

[0187] Based on any of the above embodiments, in one embodiment, the sum of the overlapping areas of each sub-block and its adjacent sub-blocks accounts for a proportion greater than a first proportion threshold of each sub-block.

[0188] Based on any of the above embodiments, in one embodiment, the high-resolution convolutional neural network and the low-resolution convolutional neural network are specifically U-net networks; or,

[0189] The high-resolution convolutional neural network and the low-resolution convolutional neural network are specifically fully convolutional neural networks.

[0190] Based on any of the above embodiments, in one embodiment, the device further includes a pre-processing module for preprocessing the original brain image and registering it with a standard brain image to obtain the first brain image in standard brain space.

[0191] Based on any of the above embodiments, in one embodiment, the preprocessing includes at least one of the following:

[0192] Noise reduction, offset field correction, and grayscale normalization.

[0193] Based on any of the above embodiments, in one embodiment, the preceding module includes a cropping unit, used to cut the registered brain image after registration to the standard brain space using the three-dimensional region range of the standard brain image, to obtain the first brain image.

[0194] Based on any of the above embodiments, in one embodiment, the device further includes an inverse registration module, used to inversely register the brain region segmentation result of the first brain image to the original brain space to obtain the brain region segmentation result of the original brain image.

[0195] This invention also provides another brain image segmentation device based on a neural network, the device comprising:

[0196] The registration module is used to preprocess the original brain images and rigidly register them with standard brain images to obtain registered brain images.

[0197] The third segmentation module is used to input the registered brain image into a convolutional neural network for brain region segmentation to obtain segmented brain images in standard brain space.

[0198] The inverse registration module is used to inversely register the segmented brain images in the standard brain space to the original brain space to obtain the segmented brain images corresponding to the original brain images.

[0199] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute all or part of the steps of the neural network-based brain image segmentation methods provided above.

[0200] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0201] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, and when the program instructions are executed by a computer, the computer is able to perform all or part of the steps of the neural network-based brain image segmentation methods provided above.

[0202] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform all or part of the steps of the neural network-based brain image segmentation methods provided above.

[0203] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0204] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0205] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A neural network-based brain image segmentation method, characterized by, The method comprises: blurring a first brain image to obtain a second brain image, wherein the first brain image is a three-dimensional brain image in standard brain space; dividing the first brain image and the second brain image into a plurality of sub-blocks respectively, wherein the division strategy of the first brain image is different from the division strategy of the second brain image; inputting each sub-block of the second brain image into a corresponding low-resolution convolutional neural network model to obtain a low-resolution brain region segmentation result of each sub-block of the second brain image; inputting each sub-block of the first brain image and the corresponding low-resolution brain region segmentation result into a corresponding high-resolution convolutional neural network model, and fusing to obtain a high-resolution brain region segmentation result of the first brain image; the inputting each sub-block of the second brain image into a corresponding low-resolution convolutional neural network model to obtain a low-resolution brain region segmentation result of each sub-block of the second brain image, and the inputting each sub-block of the first brain image and the corresponding low-resolution brain region segmentation result into a corresponding high-resolution convolutional neural network model, and fusing to obtain a high-resolution brain region segmentation result of the first brain image, comprising: inputting each sub-block of the second brain image into a corresponding low-resolution convolutional neural network model to obtain a low-resolution brain region segmentation result of each sub-block of the second brain image; fusing the low-resolution brain region segmentation results of the plurality of sub-blocks of the second brain image to obtain a brain region segmentation result of the second brain image; inputting each sub-block of the first brain image and the corresponding low-resolution brain region segmentation result in the brain region segmentation result of the second brain image into a corresponding high-resolution convolutional neural network to obtain a corresponding high-resolution brain region segmentation result of the sub-block; fusing the high-resolution brain region segmentation results of the plurality of sub-blocks of the first brain image to obtain a brain region segmentation result of the first brain image.

2. The brain image segmentation method of claim 1, wherein, The fusion process is realized by the following way: traversing the brain image to determine a current pixel point to be fused and a brain region segmentation label of the current pixel point to be fused in each sub-block; determining the most numerous segmentation label in the segmentation labels as the fusion label of the pixel point to be fused by majority voting according to the segmentation labels of the current pixel point to be fused in each sub-block of the brain image; repeating the above process until all pixel points in the brain image are fused.

3. The brain image segmentation method of claim 1, wherein, The sum of the overlapping areas of each sub-block and adjacent sub-blocks in the plurality of sub-blocks accounts for more than a first proportion threshold of each sub-block.

4. The brain image segmentation method of claim 1, wherein, The method further comprises: preprocessing an original brain image and registering the original brain image with a standard brain image to obtain the first brain image in standard brain space.

5. The brain image segmentation method of claim 4, wherein, The preprocessing comprises one or more of the following: noise reduction processing, bias field correction, and gray scale normalization.

6. The brain image segmentation method of claim 4, wherein, The preprocessing of the original brain image and the registration of the original brain image with the standard brain image to obtain the first brain image in standard brain space comprises: cutting the registered brain image registered to the standard brain space using the three-dimensional region range of the standard brain image to obtain the first brain image.

7. The brain image segmentation method of claim 1, wherein, The method further comprises: The brain region segmentation result of the first brain image is inversely registered to the original brain space to obtain a brain region segmentation result of the original brain image.

8. A neural network-based brain image segmentation device, characterized by, The method comprises the following steps: The method comprises the following steps: A blurring module is configured to blur the first brain image to obtain a second brain image, wherein the first brain image is a three-dimensional brain image in a standard brain space; A segmentation module is configured to segment the first brain image and the second brain image into a plurality of sub-blocks, respectively, wherein the segmentation strategy of the first brain image is different from that of the second brain image; A second segmentation module is configured to input each sub-block of the second brain image into a corresponding low-resolution convolutional neural network model to obtain a low-resolution brain region segmentation result of each sub-block of the second brain image; A first segmentation module is configured to input each sub-block of the first brain image and the corresponding low-resolution brain region segmentation result into a corresponding high-resolution convolutional neural network model and fuse the results to obtain a high-resolution brain region segmentation result of the first brain image; The second segmentation module and the first segmentation module are specifically configured to: input each sub-block of the second brain image into a low-resolution convolutional neural network corresponding to the position of the sub-block to obtain a corresponding low-resolution brain region segmentation result of the sub-block; fuse the low-resolution brain region segmentation results of the plurality of sub-blocks of the second brain image to obtain a brain region segmentation result of the second brain image; input each sub-block of the first brain image and the corresponding low-resolution brain region segmentation result of the sub-block in the brain region segmentation result of the second brain image into a high-resolution convolutional neural network corresponding to the position of the sub-block to obtain a corresponding high-resolution brain region segmentation result of the sub-block; 9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, fuse the high-resolution brain region segmentation results of the plurality of sub-blocks of the first brain image to obtain a brain region segmentation result of the first brain image. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The processor executes the program to implement all or part of the steps of the brain image segmentation method according to any one of claims 1 to 7.

11. A computer program product, the computer program product comprising computer executable instructions, characterised in that, The computer program is executed by the processor to implement all or part of the steps of the brain image segmentation method according to any one of claims 1 to 7. The instructions, when executed, are configured to implement all or part of the steps of the brain image segmentation method according to any one of claims 1 to 7.

Images