A brain tumor MRI image segmentation method, device and readable storage medium

Abstract

The application provides a brain tumor MRI image segmentation method, device and readable storage medium, and relates to the field of image detection. The brain tumor MRI image segmentation method comprises: acquiring a brain tumor MRI image to be segmented; constructing a DOU-Net neural network for brain tumor MRI image segmentation, training the DOU-Net neural network based on a preset brain tumor MRI sample data set, and obtaining a brain tumor MRI image segmentation model; and segmenting the brain tumor MRI image to be segmented by using the MRI image segmentation model to obtain a brain tumor MRI segmentation result. The DOU-Net neural network comprises a deep over-parameterized convolution block, a group normalization layer and a GELU activation function layer. The deep over-parameterized convolution layer is used to obtain a deep feature image, the group normalization layer is used to obtain a group normalized image, and the GELU activation function layer is used to obtain a nonlinear conversion image. The application can improve the accuracy of brain tumor image segmentation.

Description

Technical Field

[0001] This invention relates to the field of image detection, and more specifically, to a method, apparatus, and readable storage medium for segmenting brain tumor MRI images. Background Technology

[0002] Meningiomas are common primary intracranial tumors with high incidence and mortality rates. Their locational variability, size variation, and strong heterogeneity make diagnosis exceptionally difficult.

[0003] Brain tumor MRI (Magnetic Resonance Imaging) image segmentation refers to identifying and segmenting different regions of a brain tumor from MRI images. Segmentation of different regions allows for the extraction of detailed information about the patient's brain tumor, thus providing a basis for diagnosis. However, due to the multiple modalities present in brain tumor MRI images, achieving information complementarity inevitably increases the difficulty of segmenting brain tumor regions. Currently, deep neural network-based brain tumor segmentation methods are mainly used to segment brain tumor MRI images. However, this approach suffers from insufficient feature point extraction, low convergence speed, and poor stability, resulting in poor segmentation accuracy for brain tumor MRI images. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the segmentation accuracy of brain tumor MRI images.

[0005] To address the above problems, the present invention provides a brain tumor MRI image segmentation method, apparatus, and readable storage medium.

[0006] In a first aspect, the present invention provides a method for segmenting brain tumor MRI images, comprising:

[0007] Acquire MRI images of a brain tumor to be segmented, wherein the MRI images of the brain tumor to be segmented include the brain tumor region;

[0008] A DOU-Net neural network for brain tumor MRI image segmentation is constructed, and the DOU-Net neural network is trained based on a preset brain tumor MRI sample dataset to obtain a brain tumor MRI image segmentation model.

[0009] The brain tumor MRI image segmentation model is used to segment the brain tumor MRI image to be segmented to obtain the brain tumor MRI segmentation result.

[0010] The DOU-Net neural network includes a deep overparameterized convolutional block, a group normalization layer, and a GELU activation function layer. The deep overparameterized convolutional layer is used to perform deep convolution processing on the brain tumor MRI image to obtain a depth feature image. The group normalization layer is used to perform group normalization processing on the depth feature image to obtain a group normalized image. The GELU activation function layer is used to perform nonlinear transformation on the group normalized image to obtain a nonlinear transformed image.

[0011] Optionally, the DOU-Net neural network includes an encoder, a decoder, and a bottleneck layer connecting the encoder and the decoder. The depthwise overparameterized convolutional block includes a first depthwise overparameterized convolutional block in the encoder, a second depthwise overparameterized convolutional block in the decoder, and a third depthwise overparameterized convolutional block in the bottleneck layer. The group normalization layer includes a first group normalization layer in the encoder and a second group normalization layer in the decoder. The GELU activation function layer includes a first GELU activation function layer in the encoder and a second GELU activation function layer in the decoder.

[0012] Optionally, the depth-overparameterized convolutional block includes a 3D depth-overparameterized convolutional block.

[0013] Optionally, the group normalization layer includes a grouping unit, a normalization unit, and a merging unit. The grouping unit is used to divide the depth feature image into feature image groups in the channel dimension. The normalization unit is used to perform a normalization operation on the feature image groups. The merging unit is used to merge the normalized feature image groups to obtain the group normalized image.

[0014] Optionally, the expression for the GELU activation function in the GELU activation function layer is:

[0015]

[0016] Where x represents the group of normalized images.

[0017] Optionally, the brain tumor MRI images to be segmented include T1-weighted images, T2-weighted images, T1ce images, and free water-inhibited sequence images.

[0018] Optionally, the brain tumor MRI sample dataset includes brain tumor MRI training sample images and brain tumor MRI verification sample images, both of which are brain tumor MRI images with annotation information. The step of training the DOU-Net neural network based on the preset brain tumor MRI image sample dataset to obtain a brain tumor MRI image segmentation model includes:

[0019] Based on the brain tumor MRI training sample images, multiple DOU-Net neural networks were trained using five-fold cross-validation to obtain cross-validation results;

[0020] Based on the cross-validation results, select a DOU-Net neural network that meets the preset conditions;

[0021] The DOU-Net neural network that meets the preset conditions is validated based on the brain tumor MRI validation sample images, and the brain tumor MRI image segmentation model is constructed based on the DOU-Net neural network whose validation results meet the preset results.

[0022] Optionally, the DOU-Net neural network includes a 2D DOU-Net neural network, a 3D DOU-Net neural network, and a 3D DOU-Net cascaded neural network.

[0023] In a second aspect, the present invention provides a brain tumor MRI image segmentation device, comprising:

[0024] An image acquisition module is used to acquire an MRI image of a brain tumor to be segmented, wherein the MRI image of the brain tumor to be segmented includes the brain tumor region;

[0025] The image segmentation model construction module is used to construct a DOU-Net neural network for segmenting the brain tumor MRI images, and to train the DOU-Net neural network based on a preset brain tumor MRI sample dataset to obtain a brain tumor MRI image segmentation model.

[0026] The segmentation module is used to segment the brain tumor MRI image to be segmented using the brain tumor MRI image segmentation model to obtain the brain tumor MRI segmentation result;

[0027] The DOU-Net neural network includes a deep overparameterized convolutional block, a group normalization layer, and a GELU activation function layer. The deep overparameterized convolutional layer is used to perform deep convolution processing on the brain tumor MRI image to obtain a depth feature image. The group normalization layer is used to perform group normalization processing on the depth feature image to obtain a group normalized image. The GELU activation function layer is used to perform nonlinear transformation on the group normalized image to obtain a nonlinear transformed image.

[0028] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described brain tumor MRI image segmentation method.

[0029] Compared with existing technologies, the beneficial effects of this invention are as follows: After acquiring the MRI image of a brain tumor to be segmented, a brain tumor MRI image segmentation model is obtained by training a DOU-Net neural network. This model is then used to obtain the segmentation result of the brain tumor MRI image, thus achieving automatic segmentation of the brain tumor MRI image. Furthermore, the DOU-Net neural network employs deep overparameterized convolutional blocks, which, compared to traditional convolutional blocks, allows for the configuration of a separate convolutional kernel for each input channel. This not only reduces the number of parameters in the neural network but also increases the diversity of feature extraction, thereby improving the segmentation accuracy of the brain tumor MRI image. Moreover, the introduction of a group normalization layer, utilizing the sharing of normalization parameters within a group, reduces the computational load while enhancing the generalization ability of the segmentation model for features from different channels, improving the adaptability of the segmentation model to different scenarios. Finally, the introduction of a smooth and stable GELU activation function layer avoids the gradient vanishing problem compared to the traditional U-Net neural network, enhancing the robustness of the segmentation model. This invention improves the segmentation accuracy of brain tumor MRI images by enhancing the overall performance of the brain tumor MRI image segmentation model. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a brain tumor MRI image segmentation method according to an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the architecture of the DOU-Net neural network according to an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the structure of a brain tumor MRI image segmentation device according to an embodiment of the present invention.

[0033] 21-Encoder; 211-First depthwise overparameterized convolutional block; 22-Decoder; 221-Second depthwise overparameterized convolutional block; 23-Bottleneck layer; 231-Third depthwise overparameterized convolutional block. Detailed Implementation

[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0036] The brain tumor MRI image segmentation method provided in this invention can be applied in a server-network-terminal device environment, where the terminal device communicates with the server via a network. The server performs a brain tumor MRI image segmentation model and acquires the brain tumor MRI image to be segmented transmitted from the terminal device. Then, the server inputs the brain tumor MRI image to be segmented into the brain tumor MRI image segmentation model to obtain the brain tumor MRI image segmentation result. The terminal device may include, but is not limited to, personal computers, laptops, smartphones, tablets, and portable wearable devices. The server may be a standalone server or a combination of multiple servers.

[0037] like Figure 1 and Figure 2 As shown, in order to solve the above-mentioned technical problems, this embodiment of the invention provides a brain tumor image segmentation method, including:

[0038] S11. Obtain the MRI image of the brain tumor to be segmented, wherein the MRI image of the brain tumor to be segmented includes the brain tumor region.

[0039] Specifically, MRI (Magnetic Resonance Imaging) is a non-invasive imaging technique that uses the magnetic resonance effect to capture electromagnetic signals generated by the human body and uses them to generate tomographic images of human structures. MRI can employ various imaging sequences, each producing MRI images with unique characteristics, showcasing the anatomical structure of the brain from multiple dimensions. This enhances the richness and diversity of brain information acquired, providing a data foundation for subsequent brain tumor MRI images. In this embodiment, the brain tumor MRI image to be segmented can be a pre-obtained image from MRI. This embodiment only utilizes this image for brain tumor segmentation and does not involve the image extraction process.

[0040] S12. Construct a DOU-Net neural network for brain tumor MRI image segmentation, and train the DOU-Net neural network based on a preset brain tumor MRI sample dataset to obtain a brain tumor MRI image segmentation model.

[0041] Specifically, the DOU-Net neural network is an improvement based on the U-Net neural network structure. By improving the convolutional layer, normalization layer and activation function layer in the U-Net neural network, it has stronger feature extraction ability and denoising performance, thus obtaining a brain tumor MRI image segmentation model with better segmentation performance. Moreover, in this embodiment, the brain tumor MRI sample dataset can be obtained from pre-stored images in the database or from a public image library, without any specific restrictions.

[0042] S13. The brain tumor MRI image to be segmented is segmented using the brain tumor MRI image segmentation model to obtain the brain tumor MRI segmentation result.

[0043] By utilizing MRI image segmentation models to segment MRI images of brain tumors, automatic and efficient segmentation of these images can be achieved, helping doctors accurately determine brain tumor data. Specifically, the brain tumor segmentation result refers to the region composed of brain tumors in the target brain MRI image, and the determination of the corresponding brain tumor type. Brain tumor types are divided into: edema (ED), non-enhancing tumor (NET), and enhancing tumor core (ET). These three types can form three segmented regions: enhancing tumor core region (ET), tumor core region (TC), and whole tumor region (WT). The tumor type in the tumor core region includes both enhancing tumor core and gangrene, while the brain type in the whole tumor region includes edema, enhancing tumor core, and gangrene.

[0044] The DOU-Net neural network includes a deep overparameterized convolutional block, a group normalization layer, and a GELU activation function layer. The deep overparameterized convolutional layer is used to perform deep convolution processing on the brain tumor MRI image to obtain a depth feature image. The group normalization layer is used to perform group normalization processing on the depth feature image to obtain a group normalized image. The GELU activation function layer is used to perform nonlinear transformation on the group normalized image to obtain a nonlinear transformed image.

[0045] Compared with existing technologies, the beneficial effects of this invention are as follows: A brain tumor MRI image segmentation model is obtained by training a DOU-Net neural network, and the brain tumor MRI image segmentation model is used to obtain the brain tumor MRI image segmentation result, thereby achieving automatic segmentation of brain tumor MRI images. Furthermore, the DOU-Net neural network uses deep overparameterized convolutional blocks, which, compared to traditional convolutional blocks, can configure a separate convolutional kernel for each input channel, not only reducing the number of parameters in the neural network but also increasing the diversity of feature extraction, thus improving the segmentation accuracy of brain tumor MRI images. Moreover, by introducing a group normalization layer, the sharing of normalization parameters within a group can reduce the amount of parameter computation while enhancing the generalization ability of the segmentation model for features from different channels, improving the adaptability of the segmentation model to different scenarios. Finally, by introducing a smooth and stable GELU activation function layer, the problem of gradient vanishing can be avoided, enhancing the stability of the segmentation model. This invention improves the segmentation accuracy of brain tumor MRI images by enhancing the overall performance of the brain tumor MRI image segmentation model.

[0046] Reference Figure 2 The DOU-Net neural network includes an encoder 21, a decoder 22, and a bottleneck layer 23 connecting the encoder 21 and the decoder 22. The depth-overparameterized convolutional blocks include a first depth-overparameterized convolutional block 211 in the encoder 21, a second depth-overparameterized convolutional block 221 in the decoder 22, and a third depth-overparameterized convolutional block 231 in the bottleneck layer 23. The group normalization layer includes a first group normalization layer in the encoder and a second group normalization layer in the decoder. The GELU activation function layer includes a first GELU activation function layer in the encoder and a second GELU activation function layer in the decoder.

[0047] Specifically, encoder 21 extracts features from the input data, progressively reducing the size and dimensionality of the feature maps through parameterized convolutional blocks and pooling layers. Decoder 22 upsamples and reconstructs the feature maps extracted by encoder 21, thereby generating an output of the same size as the input data. Bottleneck layer 23 refers to the layer containing the smallest feature map in the DOU-Net neural network; this layer typically has low spatial resolution and a high number of channels. In this embodiment, deep overparameterized convolutional blocks are used to replace the convolutional blocks in the U-Net neural network. This reduces the number of parameters in the brain tumor MRI image segmentation model without increasing computational cost, extracts richer and more accurate semantic features from the image, and reduces noise interference, thereby improving the recognition accuracy of brain tumors in brain tumor MRI images and thus improving the segmentation accuracy of the brain tumor MRI image segmentation model. Furthermore, the deep convolution operation in the deep overparameterized convolutional blocks enables rapid feature convergence, thereby improving the segmentation efficiency of the brain tumor MRI image segmentation model. The introduction of group normalization layers can enhance the generalization ability of the segmentation model to features from different channels while reducing the computational cost of parameters. The GELU activation function layer can avoid the gradient vanishing problem and enhance the stability of the brain tumor MRI image segmentation model.

[0048] In one embodiment, the depth-overparameterized convolutional block includes a 3D depth-overparameterized convolutional block.

[0049] Furthermore, by designing the depth overparameterized convolutional block as a 3D depth overparameterized convolutional block, it is possible to better capture the spatial features and contextual information of the data during the processing of three-dimensional data, thereby achieving accurate segmentation of the brain tumor MRI image to be segmented.

[0050] In one embodiment, the group normalization layer includes a grouping unit, a normalization unit, and a merging unit. The grouping unit is used to divide the depth feature image into feature image groups in the channel dimension. The normalization unit is used to perform a normalization operation on the feature image groups. The merging unit is used to merge the normalized feature image groups to obtain the group normalized image.

[0051] Specifically, the grouping unit is responsible for dividing the depth feature image into multiple feature image groups along the channel dimension; the normalization unit performs a normalization operation on each feature image group, and the normalization method used here can be batch normalization or group normalization, which can be selected according to actual needs. Normalization operation helps to accelerate the training convergence of the model, reduce internal covariate bias, and improve the generalization ability of the model; the merging unit merges the feature image groups that have undergone normalization operation to obtain the final group normalized image. The merging operation refers to connecting the feature images of each feature image group in sequence to form a complete normalized feature image, thereby improving the training speed and segmentation efficiency of the brain tumor MRI image segmentation model.

[0052] Furthermore, the expression for the GELU activation function in the GELU activation function layer is:

[0053]

[0054] Where x represents the group of normalized images.

[0055] Compared to the ReLU activation function used in the existing U-Net neural network, the GELU activation function has a smoother curve, making the backpropagation process of the DOU-Net neural network more stable and helping to alleviate the gradient vanishing problem. In addition, the shape of the GELU activation function is close to the cumulative distribution function of the normal distribution, which helps to improve the modeling ability of the brain tumor MRI image segmentation model, making the model more stable and able to flexibly adapt to complex data scenarios.

[0056] In one embodiment, the brain tumor MRI images to be segmented include T1-weighted images, T2-weighted images, T1ce images, and free water-inhibited sequence images.

[0057] Furthermore, in MRI images of brain tumors to be segmented, T1-weighted imaging shows high signal intensity for fat and low signal intensity for water, thus appearing brighter and darker for fat in the image; T2-weighted imaging shows high signal intensity for water, thus appearing brighter and darker for other tissues in the image; T1ce imaging, performed after the use of contrast agents, enhances the contrast of certain tissues or lesions, thereby displaying these areas more clearly; Free Water Inhibition (FLAIR) imaging highlights abnormal signals within the brain parenchyma, such as gliomas or inflammation, by suppressing the signal of cerebrospinal fluid. Through the complementary information of multimodal images, the richness and accuracy of brain tumor information acquisition are improved, thus providing a data foundation for subsequent diagnosis and treatment.

[0058] In one embodiment, before segmenting the brain tumor MRI image to be segmented using the brain tumor MRI image segmentation model to obtain the brain tumor MRI segmentation result, the method further includes: performing data preprocessing on the brain tumor MRI image to be segmented, so that the acquired data is more suitable for analysis and processing by the brain tumor MRI image segmentation model. This data preprocessing includes, but is not limited to, image cropping, resampling, and standardization.

[0059] The specific steps of data preprocessing include:

[0060] The MRI image of the brain tumor to be segmented is cropped.

[0061] The brain tumor MRI image to be segmented after image cropping is resampled.

[0062] The resampled MRI images of the brain tumor to be segmented are standardized to complete the data preprocessing of the MRI images of the brain tumor to be segmented.

[0063] Specifically, image cropping involves cropping T1-weighted, T2-weighted, T1ce-weighted, and free water-suppressed sequence images from the brain tumor MRI images to be segmented to non-zero regions. First, a three-dimensional non-zero template is generated for each image (T1-weighted, T2-weighted, T1ce-weighted, and free water-suppressed sequence images) to identify which regions in the image are non-zero, i.e., regions containing valid information. Each modality has a corresponding non-zero template, and the non-zero template for the entire dataset is the union of the non-zero templates from all modalities. Then, based on the generated non-zero templates, the size and position of the bounding box used for cropping are determined. This bounding box will encompass all non-zero regions and surround them as closely as possible. Finally, based on the determined bounding box, the MRI images of each modality are cropped, removing the completely black background regions around the image and retaining the parts containing valid information. The cropped images from each modality are then reassembled for subsequent processing and analysis. Image cropping removes invalid background while reducing image size, thus improving computational efficiency. At the same time, the cropped image still retains important information related to the segmentation task, so it does not affect the final segmentation result.

[0064] The resampling process includes: determining the target spatial size, determining the target size for each image, adjusting the image size, and processing the segmented and labeled images. The purpose of resampling is to address the issue of inconsistent actual spatial sizes (spacing) represented by individual voxels in different images within a dataset of brain MRI images. Resampling first requires determining the size of the target space. In most cases, the median of the different intervals between images in a brain MRI dataset is used. However, in anisotropic datasets, the interval value at the 10th percentile is taken as the target space size for the coordinates of the largest interval. For anisotropy, the criterion is whether, among the three dimensions of the median interval, the interval in one dimension is more than three times the interval in another dimension, and the median size of that dimension is less than one-third of the median size of the other dimension. Then, since the product of the interval and shape in each image is a constant representing the size of the entire image in actual space, the target size of each image can be determined based on the size of the target space. Finally, the `resize` function from the `skimage` library is called to adjust the size of each image. Different resize strategies are applied depending on whether anisotropy exists. If there is no anisotropy, third-order spline interpolation is performed on the 3D image. If anisotropy exists, assuming the dimension with the largest interval is the z-axis, third-order spline interpolation is performed only in the xy-plane of the image, while nearest-neighbor interpolation is used on the z-axis. For segmented labeled images, nearest neighbor interpolation is used in all three dimensions, regardless of anisotropy. Convolutional neural networks operate only in voxel space, ignoring size information in the actual physical space. Resampling during data preprocessing can effectively avoid this discrepancy.

[0065] It should be explained that third-order spline interpolation is an interpolation method typically used to estimate the curve or surface between a given set of data points. In third-order spline interpolation, the curve between data points is modeled as a piecewise function consisting of a set of cubic polynomials. These cubic polynomials satisfy continuity at connection points, continuity of the first derivative, and continuity of the second derivative, thus ensuring the smoothness and continuity of the curve.

[0066] Standardization refers to z-scoring normalization of each 3D image, which uses the image's mean and standard deviation to standardize the image. This process involves subtracting the average intensity of all pixels in the image from the intensity of each pixel, and then dividing by the standard deviation of all pixels in the image. If the image size is reduced by more than 1 / 4 during the data cropping stage, standardization is only performed on the non-zero regions after cropping. The advantage of standardization is that it ensures that the gray values ​​of each image have the same distribution, thus avoiding the possibility of outliers negatively impacting the training model.

[0067] In one embodiment, the brain tumor MRI sample dataset includes brain tumor MRI training sample images and brain tumor MRI verification sample images, both of which are brain tumor MRI images with labeled information. The step of training the DOU-Net neural network based on the preset brain tumor MRI image sample dataset to obtain a brain tumor MRI image segmentation model includes:

[0068] Based on the brain tumor MRI training sample images, multiple DOU-Net neural networks were trained using five-fold cross-validation to obtain cross-validation results;

[0069] Based on the cross-validation results, select a DOU-Net neural network that meets the preset conditions;

[0070] The DOU-Net neural network that meets the preset conditions is validated based on the brain tumor MRI validation sample images, and the brain tumor MRI image segmentation model is constructed based on the DOU-Net neural network whose validation results meet the preset results.

[0071] In brain tumor MRI sample datasets with noise or interference, the results of a single DOU-Net neural network may fluctuate significantly. This embodiment reduces the volatility of DOU-Net neural networks by integrating the results of multiple DOU-Net neural networks, thereby enhancing the robustness of the brain tumor MRI image segmentation model. Simultaneously, it reduces the bias and variance of each DOU-Net neural network, thus improving the segmentation accuracy of the brain tumor MRI image segmentation model.

[0072] Furthermore, the DOU-Net neural network includes a 2D DOU-Net neural network, a 3D DOU-Net neural network, and a 3D DOU-Net cascaded neural network.

[0073] The DOU-Net neural network is used to process two-dimensional images; the 3D DOU-Net neural network is used to manipulate pixels in three-dimensional images, suitable for processing three-dimensional MRI images; the 3D DOU-Net cascaded neural network requires independent configuration, design, and training. Furthermore, the DOU-Net neural network comprises two 3D DOU-Net cascaded neural networks. The first 3D DOU-Net cascaded neural network operates on the downsampled image to obtain low-resolution image segmentation results; the second 3D DOU-Net cascaded neural network further refines the results obtained from the first 3D DOU-Net cascaded neural network to obtain more refined and complementary segmentation results. During the training of the two 3D DOU-Net cascaded neural networks, depthwise overparameterized convolution is used to ensure that the output and input shapes are identical, and GELU is used as the activation function.

[0074] It should be explained that the loss function of the brain tumor MRI segmentation model includes a binary cross-entropy loss function and a dice loss function, and the expression of the loss function is as follows:

[0075] Loss = L BCE +L Dice ;

[0076] Where Loss represents the loss function, L BCE L represents the binary cross-entropy loss function. Dice Represents the dice loss function;

[0077] The evaluation metric for the loss function includes the dice similarity coefficient, which is expressed as follows:

[0078]

[0079] Where X represents the real mask image, Y represents the algorithm segmentation image, |X∩Y| represents the number of pixels at the intersection of the real mask image and the algorithm segmentation image, and |X|+|Y| represents the total number of pixels in the real mask image and the algorithm segmentation image.

[0080] For MRI images of small brain tumors to be segmented, when mapping tumor tissue to normal tissue areas, the tumor tissue in small lesions is easily treated as background, making it difficult for neural networks to extract its features. This embodiment uses the DSC method to effectively evaluate the segmentation results and improve the accuracy of brain tumor identification in small lesion areas.

[0081] Reference Figure 3 The present invention provides a brain tumor MRI image segmentation device, comprising:

[0082] An image acquisition module is used to acquire an MRI image of a brain tumor to be segmented, wherein the MRI image of the brain tumor to be segmented includes the brain tumor region;

[0083] The image segmentation model construction module is used to construct a DOU-Net neural network for segmenting the brain tumor MRI images, and to train the DOU-Net neural network based on a preset brain tumor MRI sample dataset to obtain a brain tumor MRI image segmentation model.

[0084] The segmentation module is used to segment the brain tumor MRI image to be segmented using the brain tumor MRI image segmentation model to obtain the brain tumor MRI segmentation result;

[0085] The DOU-Net neural network includes a deep overparameterized convolutional block, a group normalization layer, and a GELU activation function layer. The deep overparameterized convolutional layer is used to perform deep convolution processing on the brain tumor MRI image to obtain a depth feature image. The group normalization layer is used to perform group normalization processing on the depth feature image to obtain a group normalized image. The GELU activation function layer is used to perform nonlinear transformation on the group normalized image to obtain a nonlinear transformed image.

[0086] This invention provides an electronic device including a memory and a processor; the memory is used to store a computer program; the processor is used to implement the brain tumor MRI image segmentation method described above when the computer program is executed. This electronic device may be a computer or a server, etc.

[0087] In another embodiment of the present invention, a computer-readable storage medium is provided on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned brain tumor MRI image segmentation method is implemented.

[0088] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, 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 units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0089] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for segmenting brain tumor MRI images, characterized in that, include: Acquire MRI images of a brain tumor to be segmented, wherein the MRI images of the brain tumor to be segmented include the brain tumor region; A DOU-Net neural network for brain tumor MRI image segmentation is constructed, and the DOU-Net neural network is trained based on a preset brain tumor MRI sample dataset to obtain a brain tumor MRI image segmentation model. The brain tumor MRI image segmentation model is used to segment the brain tumor MRI image to be segmented to obtain the brain tumor MRI segmentation result. The DOU-Net neural network includes a deep overparameterized convolutional block, a group normalization layer, and a GELU activation function layer. It also includes an encoder, a decoder, and a bottleneck layer connecting the encoder and decoder. The deep overparameterized convolutional block includes a first deep overparameterized convolutional block in the encoder, a second deep overparameterized convolutional block in the decoder, and a third deep overparameterized convolutional block in the bottleneck layer. The group normalization layer includes a first group normalization layer in the encoder and a second group normalization layer in the decoder. The GELU activation function layer includes a first GELU activation function layer in the encoder and a second GELU activation function layer in the decoder. The overparameterized convolutional block also includes a 3D depth overparameterized convolutional block, which is used to perform depth convolution processing on the brain tumor MRI image to obtain a depth feature image. The group normalization layer is used to perform group normalization processing on the depth feature image to obtain a group normalized image. The GELU activation function layer is used to perform nonlinear transformation on the group normalized image to obtain a nonlinear transformed image. The group normalization layer also includes a grouping unit, a normalization unit, and a merging unit. The grouping unit is used to divide the depth feature image into feature image groups in the channel dimension. The normalization unit is used to perform a normalization operation on the feature image groups. The merging unit is used to merge the normalized feature image groups to obtain the group normalized image. The expression for the GELU activation function in the GELU activation function layer is: ； in, This represents the group of normalized images.

2. The brain tumor MRI image segmentation method according to claim 1, characterized in that, The MRI images of the brain tumor to be segmented include T1-weighted images, T2-weighted images, T1ce images, and free water-inhibited sequence images.

3. The brain tumor MRI image segmentation method according to claim 1, characterized in that, The brain tumor MRI sample dataset includes brain tumor MRI training sample images and brain tumor MRI verification sample images. Both the brain tumor MRI training sample images and the brain tumor MRI verification sample images are brain tumor MRI images with labeled information. The step of training the DOU-Net neural network based on the preset brain tumor MRI image sample dataset to obtain a brain tumor MRI image segmentation model includes: Based on the brain tumor MRI training sample images, multiple DOU-Net neural networks were trained using five-fold cross-validation to obtain cross-validation results. Based on the cross-validation results, select a DOU-Net neural network that meets the preset conditions; The DOU-Net neural network that meets the preset conditions is validated based on the brain tumor MRI validation sample images, and the brain tumor MRI image segmentation model is constructed based on the DOU-Net neural network whose validation results meet the preset results.

4. The brain tumor MRI image segmentation method according to claim 3, characterized in that, The DOU-Net neural network includes a 2D DOU-Net neural network, a 3D DOU-Net neural network, and a 3D DOU-Net cascaded neural network.

5. A brain tumor MRI image segmentation device, characterized in that, include: An image acquisition module is used to acquire an MRI image of a brain tumor to be segmented, wherein the MRI image of the brain tumor to be segmented includes the brain tumor region; The image segmentation model construction module is used to construct a DOU-Net neural network for segmenting the brain tumor MRI images, and to train the DOU-Net neural network based on a preset brain tumor MRI sample dataset to obtain a brain tumor MRI image segmentation model. The segmentation module is used to segment the brain tumor MRI image to be segmented using the brain tumor MRI image segmentation model to obtain the brain tumor MRI segmentation result; The DOU-Net neural network includes a deep overparameterized convolutional block, a group normalization layer, and a GELU activation function layer. It also includes an encoder, a decoder, and a bottleneck layer connecting the encoder and decoder. The deep overparameterized convolutional block includes a first deep overparameterized convolutional block in the encoder, a second deep overparameterized convolutional block in the decoder, and a third deep overparameterized convolutional block in the bottleneck layer. The group normalization layer includes a first group normalization layer in the encoder and a second group normalization layer in the decoder. The GELU activation function layer includes a first GELU activation function layer in the encoder and a second GELU activation function layer in the decoder. The overparameterized convolutional block also includes a 3D depth overparameterized convolutional block, which is used to perform depth convolution processing on the brain tumor MRI image to obtain a depth feature image. The group normalization layer is used to perform group normalization processing on the depth feature image to obtain a group normalized image. The GELU activation function layer is used to perform nonlinear transformation on the group normalized image to obtain a nonlinear transformed image. The group normalization layer also includes a grouping unit, a normalization unit, and a merging unit. The grouping unit is used to divide the depth feature image into feature image groups in the channel dimension. The normalization unit is used to perform a normalization operation on the feature image groups. The merging unit is used to merge the normalized feature image groups to obtain the group normalized image. The expression for the GELU activation function in the GELU activation function layer is: ； in, This represents the group of normalized images.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the brain tumor MRI image segmentation method as described in any one of claims 1 to 4.

Images