A brain tumor segmentation system based on a multi-scale neural network

Abstract

The present invention relates to the technical field of medical image processing, and provides a brain tumor segmentation system based on a multi-scale neural network, including a data acquisition module, a data preprocessing module, a multi-scale neural network module, and a post-processing module; the system acquires brain MRI images through the data acquisition module, and performs normalization, resampling, and data augmentation processing through the preprocessing module; the multi-scale neural network module combines residual dense blocks, attention mechanisms, and dilated convolutions to extract multi-scale features, and gradually restores the spatial resolution through skip connections and a decoder to generate a brain tumor segmentation feature map; the post-processing module further performs denoising, region optimization, and boundary smoothing processing on the segmentation result to generate a complete and accurate tumor region, and superimposes it on the original MRI image for intuitive display, significantly improving the accuracy and efficiency of segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to a brain tumor segmentation system based on a multi-scale neural network. Background Art

[0002] With the development of technology, the combination of medical imaging technology and artificial intelligence provides important support for the accurate diagnosis of brain tumors; in recent years, more and more brain tumor segmentation systems have begun to adopt neural networks to improve the segmentation accuracy of brain MRI images and assist doctors in achieving more efficient and accurate diagnoses; however, traditional brain tumor segmentation systems still have the following deficiencies in application: First, when processing brain MRI images, traditional brain tumor segmentation systems have insufficient ability to recognize the morphological changes and complex features of tumors; since tumors usually have irregular shapes, diverse texture features, and complex boundary information, existing systems often lack a detailed capture of the tumor area, resulting in poor integrity and accuracy of the segmentation results, and prone to artifacts or missing some key tumor areas; Second, traditional brain tumor segmentation systems have insufficient utilization of multi-scale information and lack the ability to comprehensively extract and integrate local and global features; especially when dealing with small tumors or tumors with unclear contrast with surrounding normal tissues, traditional systems have low sensitivity to different spatial scale features and are difficult to effectively distinguish tumors from background tissues, directly affecting the reliability and performance of the segmentation; In addition, the existing systems have limited ability to express high-dimensional features, which easily leads to problems such as rough boundaries and discontinuous regions in the segmentation results; Therefore, there is an urgent need for a brain tumor segmentation system based on a multi-scale neural network that can more comprehensively extract and fuse multi-scale features and enhance the accurate recognition ability of tumor regions. Summary of the Invention

[0003] The present invention provides a brain tumor segmentation system based on a multi-scale neural network. By adopting a direction-sensitive convolution, a dilated convolution mechanism, a group attention mechanism, and the fusion of position attention and channel attention, high-precision segmentation of brain MRI image data is achieved. Specifically, the present invention extracts horizontal, vertical, and diagonal direction features of the image through direction-sensitive convolution, combines the dilated convolution mechanism to expand the receptive field, effectively captures multi-scale features, and enhances the recognition ability of the tumor region through multi-level feature extraction and integration. At the same time, a direction-aware convolution is used to optimize the group attention mechanism, and through grouped feature extraction, weight assignment, and weighted operations, the accuracy of feature expression is strengthened, and the spatial and channel feature weights are optimized by combining the position attention and channel attention mechanisms to further improve the segmentation accuracy. In addition, the present invention also uses a step-by-step upsampling method to restore the spatial resolution of the segmentation result, and in the post-processing stage, artifacts and noise regions are removed through denoising, region optimization, and boundary smoothing to generate an optimized segmentation result. This system can comprehensively extract and fuse multi-scale features, significantly improving the integrity and accuracy of tumor region segmentation.

[0004] The present invention provides a brain tumor segmentation system based on a multi-scale neural network. The system includes a data acquisition module, a data preprocessing module, a multi-scale neural network module, and a post-processing module.

[0005] The data acquisition module acquires brain MRI image data through a public dataset and clinical data.

[0006] The data preprocessing module normalizes, resamples, aligns the images, and performs data augmentation on the brain MRI image data to obtain preprocessed brain MRI image data.

[0007] The multi-scale neural network module constructs a residual dense block through a dense connection mechanism and a residual connection mechanism. In the residual dense block, a direction-sensitive convolution and a dilated convolution mechanism are introduced to construct a direction-dilated residual block. A direction-aware convolution is used to optimize the group attention mechanism to construct an enhanced group attention mechanism, and then combined with position attention and channel attention to construct a DGAP module. The skip connection is optimized by stacking convolutions to construct a stacked skip connection. Combining the direction-dilated residual block, the DGAP module, and the stacked skip connection, a multi-scale direction-aware brain tumor segmentation network is constructed. The multi-scale direction-aware brain tumor segmentation network is used to process the preprocessed brain MRI image data to generate a brain tumor segmentation feature map. The multi-scale direction-aware brain tumor segmentation network includes a local texture edge feature extraction module, a direction-dilated residual block feature extraction module, a DGAP optimization module, a stacked skip connection module, a resolution restoration decoding module, and a decoder.

[0008] The post-processing module performs denoising, region optimization, boundary smoothing, and connectivity analysis on the brain tumor segmentation feature map, removes artifacts and small noise regions, generates an optimized segmentation result, and overlays the optimized segmentation result with the brain MRI image data to visually display the tumor location.

[0009] Further, the multi-scale neural network module generates the brain tumor segmentation feature map through the following specific steps:

[0010] The local texture and edge feature extraction module extracts the local texture and edge features of the brain MRI image data through convolution operations to generate preliminary feature data;

[0011] The directional dilated residual block feature extraction module performs multi-layer feature extraction on the preliminary feature data through directional dilated residual blocks to generate basic feature data;

[0012] The DGAP optimization module optimizes the basic feature data through the DGAP module and outputs channel attention feature data;

[0013] The stacked skip connection module processes the channel attention feature data through stacked skip connections to generate skip connection feature data;

[0014] The resolution restoration decoding module transfers the skip connection feature data to the decoder, gradually upsamples through transposed convolution, and restores the spatial resolution to be consistent with the spatial resolution of the preprocessed brain MRI image data to obtain the brain tumor segmentation feature map.

[0015] Further, the directional dilated residual block feature extraction module specifically includes the following sub-modules:

[0016] The layer-by-layer convolution feature extraction module: extracts the horizontal-direction convolution feature, vertical-direction convolution feature, and diagonal-direction convolution feature of the preliminary feature data through directional dilated residual blocks, and fuses the horizontal-direction convolution feature, vertical-direction convolution feature, and diagonal-direction convolution feature of the preliminary feature data to generate direction-enhanced multi-scale feature data;

[0017] ;

[0018] Among them, represents the direction-enhanced multi-scale feature data, , , represent dynamic weight parameters, representing the weights of the horizontal-direction convolution feature, vertical-direction convolution feature, and diagonal-direction convolution feature respectively; represents the preliminary feature data, represents the dilation rate of the horizontal-direction convolution, represents the horizontal-direction convolution feature, represents the dilation rate of the vertical-direction convolution, represents the vertical direction convolution feature, represents the dilation rate of the diagonal direction convolution, represents the diagonal direction convolution feature;

[0019] Feature splicing module: Splice the direction-enhanced multi-scale feature data of all direction dilation residual blocks in the channel dimension, and integrate the multi-level features in the direction-enhanced multi-scale feature data through the dense connection mechanism to generate multi-level fusion feature data;

[0020] ;

[0021] Among them, represents the multi-level fusion feature data, represents the splicing operation in the channel dimension, represents the direction-enhanced multi-scale feature data of each residual dense block;

[0022] Residual connection module: Add the multi-level fusion feature data and the preliminary feature data element by element, and retain the original information in the preliminary feature data through the residual connection mechanism to generate residual fusion feature data;

[0023] ;

[0024] Among them, represents the residual fusion feature data, the preliminary feature data;

[0025] Activation processing module: Perform activation processing on the residual fusion feature data to generate basic feature data.

[0026] Furthermore, the DGAP optimization module specifically includes the following sub-modules:

[0027] Intra-group attention feature enhancement module: Process the basic feature data through the enhanced intra-group attention mechanism to generate intra-group attention feature data. The enhanced intra-group attention mechanism includes two layers of fully connected networks;

[0028] Position attention allocation module: Allocate spatial attention weights to the intra-group attention feature data through the position attention mechanism to obtain position attention feature data;

[0029] Channel weight optimization module: Generate the importance weights of each channel of the position attention feature data through the channel attention mechanism, and perform weighting on the position attention feature data with the importance weights of each channel to obtain channel attention feature data.

[0030] Furthermore, the intra-group attention feature enhancement module specifically includes the following sub-modules:

[0031] Group global feature extraction module: Group the basic feature data to generate grouped feature data. For each group of feature data in the grouped feature data, perform orientation-aware convolution to extract horizontal features, vertical features, and diagonal features, and then perform global average pooling to extract global information to obtain group global features;

[0032] ;

[0033] Among them, represents the index of feature grouping, represents the index of the orientation-aware convolution kernel, represents the perceived features of the th group of features in different directions, represents the grouped feature data of the th group, represents the th group of feature data for which orientation-aware convolution is performed, represents the th orientation-aware convolution kernel, including convolution kernels in horizontal, vertical, and diagonal directions, represents the convolution operation,

[0034] Attention weight generation module: Input the group global features into a two-layer fully connected network to generate attention weights;

[0035] Feature weighting module: Apply the attention weights to each group of feature data for weighting operation to generate weighted intra-group attention features;

[0036] Channel dimension restoration module: Concatenate the weighted intra-group attention features of all groups to restore the channel dimension and generate intra-group attention feature data.

[0037] Adopting the above solution, the beneficial effects obtained by the present invention are as follows:

[0038] By combining the orientation-sensitive convolution, dilated convolution mechanism, and group attention mechanism, the present invention effectively realizes the accurate extraction and optimization of multi-scale features in brain MRI images; specifically, through the orientation-sensitive convolution for feature extraction in horizontal, vertical, and diagonal directions, combined with the dilated convolution mechanism to expand the receptive field, the present invention enhances the ability to capture complex features in the tumor area, thereby improving the segmentation accuracy of tumor morphological features; this multi-scale feature extraction and optimization method solves the problem of insufficient capture of complex boundary information in traditional systems and effectively improves the integrity and accuracy of tumor segmentation results;

[0039] Meanwhile, the present invention adopts an optimized group attention mechanism. Through grouped feature extraction and weighting operations, it further enhances the attention ability of the segmentation system to the key information in the tumor area. Combining the feature weight allocation of the position attention and channel attention mechanisms, it improves the sensitivity to spatial and channel information, significantly reducing the errors and interferences in the recognition of the tumor area. This feature optimization technology enhances the recognition effect of the present invention on small tumors and regions with blurred boundaries in complex scenarios, effectively improving the accuracy and robustness of the segmentation results.

[0040] In addition, in the post-processing stage, the present invention further eliminates artifacts and small noise regions through denoising, region optimization, and boundary smoothing, generating a more complete and accurate segmentation result. By superimposing the segmentation result on the brain MRI image data, the tumor location is visually displayed, providing efficient technical support for clinical diagnosis and treatment planning. Overall, the present invention significantly improves the accuracy and integrity of brain tumor segmentation, solving the problem of unsatisfactory segmentation results in traditional systems. Brief Description of the Drawings

[0041] Figure 1 It is a schematic diagram of the modules of a brain tumor segmentation system based on a multi-scale neural network proposed by the present invention;

[0042] Figure 2 It is a schematic diagram of the structure of the multi-scale direction-aware brain tumor segmentation network proposed in Embodiment 2. Detailed Embodiments

[0043] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0044] Embodiment 1. According to Figure 1 , the present invention provides a brain tumor segmentation system based on a multi-scale neural network. The system includes a data acquisition module, a data preprocessing module, a multi-scale neural network module, and a post-processing module;

[0045] The data acquisition module acquires brain MRI image data through public datasets and clinical data;

[0046] The data preprocessing module normalizes, resamples, aligns the images, and enhances the data of the brain MRI image data to obtain preprocessed brain MRI image data;

[0047] Multi-scale neural network module, through the dense connection mechanism and the residual connection mechanism, constructs a residual dense block. In the residual dense block, the direction-sensitive convolution and the dilated convolution mechanism are introduced to construct a directional dilated residual block; through the direction-aware convolution to optimize the group attention mechanism, constructs an enhanced version of the group attention mechanism, and then combines the position attention and the channel attention to construct a DGAP module; through stacking convolution to optimize the skip connection, constructs a stacked skip connection; combines the directional dilated residual block, the DGAP module, and the stacked skip connection to construct a multi-scale direction-aware brain tumor segmentation network, and uses the multi-scale direction-aware brain tumor segmentation network to process the preprocessed brain MRI image data to generate a brain tumor segmentation feature map; the multi-scale direction-aware brain tumor segmentation network includes a local texture edge feature extraction module, a directional dilated residual block feature extraction module, a DGAP optimization module, a stacked skip connection module, a resolution restoration decoding module, and a decoder;

[0048] The post-processing module, through denoising processing, region optimization, boundary smoothing, and connectivity analysis of the brain tumor segmentation feature map, eliminates the artifacts and small noise regions, generates an optimized segmentation result, and overlays the optimized segmentation result with the brain MRI image data to visually display the tumor location.

[0049] Example 2, according to Figure 2 , this example is based on Example 1. In this example, the process of the multi-scale neural network module generating the brain tumor segmentation feature map specifically includes the following contents:

[0050] The local texture edge feature extraction module extracts the local texture and edge features of the brain MRI image data through convolution operations to generate preliminary feature data;

[0051] The directional dilated residual block feature extraction module performs multi-layer feature extraction on the preliminary feature data through the directional dilated residual block to generate basic feature data;

[0052] The DGAP optimization module optimizes the basic feature data through the DGAP module and outputs channel attention feature data;

[0053] The stacked skip connection module processes the channel attention feature data through the stacked skip connection to generate skip connection feature data;

[0054] The resolution restoration decoding module passes the skip connection feature data to the decoder, and gradually upsamples through transposed convolution to restore the spatial resolution to be consistent with the spatial resolution of the preprocessed brain MRI image data to obtain the brain tumor segmentation feature map.

[0055] Example 3, this example is based on Example 2. In this example, the directional dilated residual block feature extraction module specifically includes the following sub-modules:

[0056] Layered Convolution Feature Extraction Module: Extract the horizontal convolution features, vertical convolution features, and diagonal convolution features of the preliminary feature data through the directional dilated residual blocks, and fuse the horizontal convolution features, vertical convolution features, and diagonal convolution features of the preliminary feature data to generate directional enhanced multi-scale feature data;

[0057] ;

[0058] Among them, represents the directional enhanced multi-scale feature data, , , represent the dynamic weight parameters, representing the weights of the horizontal, vertical, and diagonal convolution features respectively; represents the preliminary feature data, represents the dilation rate of the horizontal convolution, represents the horizontal convolution feature, represents the dilation rate of the vertical convolution, represents the vertical convolution feature, represents the dilation rate of the diagonal convolution, represents the diagonal convolution feature;

[0059] Feature Concatenation Module: Concatenate the directional enhanced multi-scale feature data of all directional dilated residual blocks in the channel dimension, and integrate the multi-level features in the directional enhanced multi-scale feature data through the dense connection mechanism to generate multi-level fusion feature data;

[0060] ;

[0061] Among them, represents the multi-level fusion feature data, represents the concatenation operation in the channel dimension, represents the directional enhanced multi-scale feature data of each residual dense block;

[0062] Residual Connection Module: Add the multi-level fusion feature data and the preliminary feature data element by element, and retain the original information in the preliminary feature data through the residual connection mechanism to generate residual fusion feature data;

[0063] ;

[0064] Among them, represents the residual fusion feature data, preliminary feature data;

[0065] Activation Processing Module: Perform activation processing on the residual fusion feature data to generate basic feature data.

[0066] Example 4. This example is based on Example 2. In this example, a residual dense block is used, which specifically includes the following sub-modules:

[0067] Layer-by-layer convolutional feature extraction module: Multilevel feature extraction and fusion are performed on the preliminary feature data through the residual dense block to generate residual dense feature data;

[0068] Feature concatenation module: Concatenation is performed on the residual dense feature data in the channel dimension. Through the dense connection mechanism, the multilevel features in the residual dense feature data are integrated, the low-level features and high-level features are reused, information loss is avoided, and multilevel fusion feature data is generated;

[0069] Residual connection module: The multilevel fusion feature data and the preliminary feature data are added element by element. Through the residual connection mechanism, the original information in the preliminary feature data is retained, the expression ability of the fusion features is enhanced, and at the same time, the problem of gradient disappearance that may occur in the deep network is alleviated, generating residual fusion feature data;

[0070] Activation processing module: Activation processing is performed on the residual fusion feature data to generate basic feature data.

[0071] Example 5. This example is based on Example 3. In this example, the DGAP optimization module specifically includes the following sub-modules:

[0072] Intra-group attention feature enhancement module: The basic feature data is processed through the enhanced group attention mechanism to generate intra-group attention feature data. The enhanced group attention mechanism includes two layers of fully connected networks;

[0073] Position attention allocation module: Spatial attention weights are allocated to the intra-group attention feature data through the position attention mechanism to obtain position attention feature data;

[0074] Channel weight optimization module: The importance weights of each channel of the position attention feature data are generated through the channel attention mechanism, and the position attention feature data is weighted by the importance weights of each channel to obtain channel attention feature data.

[0075] Example 6. This example is based on Example 5. In this example, the intra-group attention feature enhancement module specifically includes the following sub-modules:

[0076] Group global feature extraction module: The basic feature data is grouped to generate grouped feature data. For each group of feature data in the grouped feature data, horizontal features, vertical features, and diagonal features are extracted through orientation-aware convolution, and then global average pooling is performed to extract global information to obtain group global features;

[0077] ;

[0078] Among them, represents the index of the feature grouping, represents the index of the direction-aware convolution kernel, represents the perceived features of the features in the th group in different directions, represents the grouped feature data of the th group, represents performing direction-aware convolution on the th group of feature data, represents the

[0079] Attention weight generation module: Input the group global features into a two-layer fully connected network to generate attention weights;

[0080] Feature weighting module: Apply the attention weights to each group of feature data for weighting operation to generate weighted intra-group attention features;

[0081] Channel dimension restoration module: Concatenate the weighted intra-group attention features of all groups to restore the channel dimension and generate intra-group attention feature data.

[0082] Embodiment 7, this embodiment is based on Embodiment 5. In this embodiment, the intra-group attention feature enhancement module specifically includes the following sub-modules:

[0083] Group global feature extraction module: Group the basic feature data to generate grouped feature data, perform global average pooling on each group of feature data in the grouped feature data to extract global information, and obtain group global features;

[0084] Attention weight generation module: Input the group global features into a two-layer fully connected network to generate attention weights;

[0085] Feature weighting module: Apply the attention weights to each group of feature data for weighting operation to generate weighted intra-group attention features;

[0086] Channel dimension restoration module: Concatenate the weighted intra-group attention features of all groups to restore the channel dimension and generate intra-group attention feature data.

[0087] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. In general, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural manners and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. A brain tumor segmentation system based on a multi-scale neural network, comprising a data acquisition module and a data preprocessing module. The data acquisition module acquires brain MRI image data; the data preprocessing module preprocesses the brain MRI image data to obtain preprocessed brain MRI image data. It is characterized in that: The system further includes a multi-scale neural network module; The multi-scale neural network module constructs residual dense blocks through a dense connection mechanism and a residual connection mechanism. In the residual dense blocks, a directional sensitive convolution and a dilated convolution mechanism are introduced to construct directional dilated residual blocks; an enhanced group attention mechanism is constructed by optimizing the group attention mechanism through directional perception convolution, and then combined with position attention and channel attention to construct a DGAP module; stacked convolutions are used to optimize skip connections to construct stacked skip connections; combining the directional dilated residual blocks, the DGAP module, and the stacked skip connections to construct a multi-scale directional perception brain tumor segmentation network, and using the multi-scale directional perception brain tumor segmentation network to process the preprocessed brain MRI image data to generate brain tumor segmentation feature maps; the multi-scale directional perception brain tumor segmentation network includes a local texture edge feature extraction module, residual dense blocks, a DGAP module, a stacked skip connection module, a resolution restoration decoding module, and a decoder.

2. The brain tumor segmentation system based on a multi-scale neural network according to claim 1, characterized in that: The process of the multi-scale neural network module generating brain tumor segmentation feature maps specifically includes the following content: The local texture edge feature extraction module extracts the local texture and edge features of the brain MRI image data through convolution operations to generate preliminary feature data; The preliminary feature data is subjected to multi-layer feature extraction through the residual dense blocks to generate basic feature data; The basic feature data is optimized through the DGAP module to output channel attention feature data; The stacked skip connection module processes the channel attention feature data through the stacked skip connections to generate skip connection feature data; The resolution restoration decoding module passes the skip connection feature data to the decoder, and gradually upsamples through deconvolution to restore the spatial resolution to be the same as that of the preprocessed brain MRI image data, obtaining brain tumor segmentation feature maps.

3. A brain tumor segmentation system based on a multi-scale neural network according to claim 2, characterized in that: The specific of the residual dense block includes the following sub-modules: Layer-by-layer convolution feature extraction module: extracts the horizontal direction convolution feature, vertical direction convolution feature, and diagonal direction convolution feature of the preliminary feature data through the directional dilated residual blocks, and fuses the horizontal direction convolution feature, vertical direction convolution feature, and diagonal direction convolution feature of the preliminary feature data to generate direction-enhanced multi-scale feature data; Feature splicing module: splices the direction-enhanced multi-scale feature data of all the directional dilated residual blocks in the channel dimension, and integrates the multi-level features in the direction-enhanced multi-scale feature data through the dense connection mechanism to generate multi-level fusion feature data; Residual connection module: adds the multi-level fusion feature data and the preliminary feature data element by element, and retains the original information in the preliminary feature data through the residual connection mechanism to generate residual fusion feature data; Activation processing module: performs activation processing on the residual fusion feature data to generate basic feature data.

4. The brain tumor segmentation system based on a multi-scale neural network according to claim 2, characterized in that: The specific of the DGAP module includes the following sub-modules: Intra-group attention feature enhancement module: processes the basic feature data through the enhanced group attention mechanism to generate intra-group attention feature data, and the enhanced group attention mechanism includes two layers of fully connected networks; Position Attention Allocation Module: Allocate spatial attention weights to the in-group attention feature data through the position attention mechanism to obtain position attention feature data; Channel Weight Optimization Module: Generate the importance weights of each channel of the position attention feature data through the channel attention mechanism, and perform weighting through the importance weights of each channel and the position attention feature data to obtain channel attention feature data.

5. The brain tumor segmentation system based on a multi-scale neural network according to claim 4, characterized in that: The in-group attention feature enhancement module specifically includes the following sub-modules: Group Global Feature Extraction Module: Group the basic feature data to generate grouped feature data, perform orientation-aware convolution on each group of feature data in the grouped feature data to extract horizontal features, vertical features, and diagonal features, and then perform global average pooling to extract global information to obtain group global features; Attention Weight Generation Module: Input the group global features into a two-layer fully connected network to generate attention weights; Feature Weighting Module: Apply the attention weights to each group of feature data for weighting operations to generate weighted in-group attention features; Channel Dimension Restoration Module: Concatenate the weighted in-group attention features of all groups to restore the channel dimension and generate in-group attention feature data.

Images