Brain tumor segmentation system based on multi-scale neural network

Abstract

The invention relates to the technical field of medical image processing, and provides a brain tumor segmentation system based on a multi-scale neural network, which comprises a data acquisition module, a data preprocessing module, a multi-scale neural network module and a post-processing module. The system acquires a brain MRI image through the data acquisition module, and performs normalization, resampling and data enhancement processing through the preprocessing module; the multi-scale neural network module combines a residual dense block, an attention mechanism and cavity convolution to extract multi-scale features, gradually recovers spatial resolution through jump connection and a decoder, and generates a brain tumor segmentation feature map; and 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 superposes the tumor region to the original MRI image for visual display, so that the segmentation accuracy and efficiency are remarkably improved.

Description

Technical Field

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

[0002] With the development of science and technology, the combination of medical imaging technology and artificial intelligence has provided 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 diagnosis. However, the traditional brain tumor segmentation system still has the following shortcomings in application: First, when processing brain MRI images, the traditional brain tumor segmentation system has insufficient recognition ability for the morphological changes and complex features of tumors. Because tumors usually have irregular shapes, diverse texture features, and complex boundary information, the existing systems often lack detailed capture of the tumor area, resulting in poor integrity and accuracy of the segmentation results, which is easy to Artifacts may appear or some key tumor areas may be missed; secondly, traditional brain tumor segmentation systems do not make sufficient use 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 that do not contrast clearly with surrounding normal tissues, traditional systems have low sensitivity to features of different spatial scales, making it difficult to effectively distinguish tumors from background tissues, which directly affects the reliability and performance of segmentation; in addition, existing systems have limited ability to express high-dimensional features, which can easily lead to problems such as rough boundaries and discontinuous regions in 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 ability to accurately identify tumor areas. Summary of the invention

[0003] The present invention provides a brain tumor segmentation system based on a multi-scale neural network, which realizes high-precision segmentation of brain MRI image data by adopting direction-sensitive convolution, hole convolution mechanism, group attention mechanism, and fusion of position attention and channel attention. Specifically, the present invention extracts horizontal, vertical and diagonal direction features of the image through direction-sensitive convolution, and expands the receptive field by combining the hole convolution mechanism, effectively captures multi-scale features, and enhances the recognition ability of tumor areas by multi-level feature extraction and integration. At the same time, the direction-aware convolution is used to optimize the group attention mechanism, and the accuracy of feature expression is enhanced by group feature extraction, weight allocation and weighted operation, 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 eliminates artifacts and noise areas by denoising, region optimization and boundary smoothing in the post-processing stage to generate an optimized segmentation result. The system can comprehensively extract and fuse multi-scale features, and significantly improve the integrity and accuracy of tumor area segmentation.

[0004] The present invention provides a brain tumor segmentation system based on a multi-scale neural network, the system comprising a data acquisition module, a data pre-processing module, a multi-scale neural network module and a post-processing module;

[0005] The data acquisition module collects brain MRI image data through public datasets and clinical data;

[0006] A data preprocessing module normalizes, resamples, aligns and enhances 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 dense connection mechanism and residual connection mechanism. In the residual dense block, the direction-sensitive convolution and hole convolution mechanism are introduced to construct a direction hole residual block; the group attention mechanism is optimized through direction-aware convolution to construct an enhanced group attention mechanism, and then the position attention and channel attention are combined to construct a DGAP module; the jump connection is optimized through stacked convolution to construct a stacked jump connection; the multi-scale direction-aware brain tumor segmentation network is constructed by combining the direction hole residual block, DGAP module, and stacked jump connection. The multi-scale direction-aware brain tumor segmentation network is used to process the pre-processed 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 hole residual block feature extraction module, a DGAP optimization module, a stacked jump connection module, a resolution recovery decoding module, and a decoder;

[0008] The post-processing module performs denoising, regional optimization, boundary smoothing and connectivity analysis on the brain tumor segmentation feature map to remove artifacts and small noise areas, generate optimized segmentation results, and superimpose the optimized segmentation results with the brain MRI image data to intuitively display the tumor location.

[0009] Furthermore, the process of generating a brain tumor segmentation feature map by a multi-scale neural network module specifically includes the following contents:

[0010] 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;

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

[0012] The DGAP optimization module optimizes the basic feature data through the DGAP module and outputs the 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 recovery decoding module passes the skip connection feature data to the decoder, and gradually upsamples it through deconvolution to restore the spatial resolution to be consistent with the spatial resolution of the preprocessed brain MRI image data, thereby obtaining a brain tumor segmentation feature map.

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

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

[0017] ;

[0018] in, represents direction-enhanced multi-scale feature data, , , Represents the dynamic weight parameter, which represents the weight of the convolution feature in the horizontal, vertical and diagonal directions respectively; represents 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 diagonal convolution dilation rate, Represents diagonal convolution features;

[0019] Feature splicing module: splices the directional enhanced multi-scale feature data of all directional hole residual blocks in the channel dimension, integrates the multi-level features in the directional enhanced multi-scale feature data through a dense connection mechanism, and generates multi-level fused feature data;

[0020] ;

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

[0022] Residual connection module: adds the multi-level fusion feature data and the preliminary feature data element by element, retains the original information in the preliminary feature data through the residual connection mechanism, and generates residual fusion feature data;

[0023] ;

[0024] in, Represents residual fusion feature data, preliminary characterization data;

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

[0026] Furthermore, the DGAP optimization module specifically includes the following submodules:

[0027] 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.

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

[0029] Channel weight optimization module: Generate the importance weight of each channel of the position attention feature data through the channel attention mechanism, and weight the importance weight of each channel and the position attention feature data to obtain the 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 group feature data, perform direction-aware convolution on each group of feature data to extract horizontal features, vertical features, and diagonal features, and then perform global average pooling to extract global information and obtain group global features;

[0032] ;

[0033] in, represents the index of the feature grouping, represents the index of the direction-aware convolution kernel, Indicates The perceptual characteristics of group features in different directions, Indicates The grouping characteristic data of the group, Indicates Group characteristic data Perform direction-aware convolution, Indicates direction-aware convolution kernels, including horizontal, vertical, and diagonal convolution kernels. represents the convolution operation, represents the sum of all direction-aware convolution kernels;

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

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

[0036] Channel dimension restoration module: concatenates the weighted intra-group attention features of all groups, restores the channel dimension, and generates intra-group attention feature data.

[0037] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0038] The present invention effectively realizes the accurate extraction and optimization of multi-scale features in brain MRI images by adopting a combination of direction-sensitive convolution, dilated convolution mechanism and group attention mechanism; specifically, the present invention extracts features in the horizontal, vertical and diagonal directions through direction-sensitive convolution, and expands the receptive field by combining the dilated convolution mechanism, thereby enhancing the ability to capture complex features of the tumor area and 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] At the same time, the present invention adopts an optimized group attention mechanism, and further enhances the segmentation system's ability to focus on key information of the tumor region through group feature extraction and weighted operations; combines the feature weight allocation of the position attention and channel attention mechanisms to improve the sensitivity to spatial and channel information, and significantly reduces the error and interference in tumor region identification; this feature optimization technology enhances the recognition effect of the present invention on small tumors and areas with blurred boundaries in complex scenarios, and effectively improves 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 areas through denoising, regional optimization and boundary smoothing, generating more complete and accurate segmentation results; by superimposing the segmentation results with brain MRI image data, the tumor location is intuitively displayed, providing efficient technical support for clinical diagnosis and treatment planning; overall, the present invention significantly improves the accuracy and completeness of brain tumor segmentation, and solves the problem of unsatisfactory segmentation effect in traditional systems. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0042] Figure 2 Schematic diagram of the structure of the multi-scale direction-aware brain tumor segmentation network proposed in Example 2. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection 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 comprising a data acquisition module, a data pre-processing module, a multi-scale neural network module and a post-processing module;

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

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

[0047] The multi-scale neural network module constructs a residual dense block through dense connection mechanism and residual connection mechanism. In the residual dense block, the direction-sensitive convolution and hole convolution mechanism are introduced to construct a direction hole residual block; the group attention mechanism is optimized through direction-aware convolution to construct an enhanced group attention mechanism, and then the position attention and channel attention are combined to construct a DGAP module; the jump connection is optimized through stacked convolution to construct a stacked jump connection; the multi-scale direction-aware brain tumor segmentation network is constructed by combining the direction hole residual block, DGAP module, and stacked jump connection. The multi-scale direction-aware brain tumor segmentation network is used to process the pre-processed 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 hole residual block feature extraction module, a DGAP optimization module, a stacked jump connection module, a resolution recovery decoding module, and a decoder;

[0048] The post-processing module performs denoising, regional optimization, boundary smoothing and connectivity analysis on the brain tumor segmentation feature map to remove artifacts and small noise areas, generate optimized segmentation results, and superimpose the optimized segmentation results with the brain MRI image data to intuitively display the tumor location.

[0049] Embodiment 2, according to Figure 2 This embodiment is based on the first embodiment. In this embodiment, the process of generating a brain tumor segmentation feature map by a multi-scale neural network module 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 hole residual block feature extraction module performs multi-layer feature extraction on the preliminary feature data through the directional hole residual block to generate basic feature data;

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

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

[0054] The resolution recovery decoding module passes the skip connection feature data to the decoder, and gradually upsamples it through deconvolution to restore the spatial resolution to be consistent with the spatial resolution of the preprocessed brain MRI image data, thereby obtaining a brain tumor segmentation feature map.

[0055] Embodiment 3: This embodiment is based on embodiment 2. In this embodiment, the directional hole residual block feature extraction module specifically includes the following submodules:

[0056] Layer-by-layer convolution feature extraction module: extracts the horizontal convolution features, vertical convolution features and diagonal convolution features of the preliminary feature data through the directional hole residual block, and fuses the horizontal convolution features, vertical convolution features and diagonal convolution features of the preliminary feature data to generate direction-enhanced multi-scale feature data;

[0057] ;

[0058] in, represents direction-enhanced multi-scale feature data, , , Represents the dynamic weight parameter, which represents the weight of the convolution feature in the horizontal, vertical and diagonal directions respectively; represents 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 diagonal convolution dilation rate, Represents diagonal convolution features;

[0059] Feature splicing module: splices the directional enhanced multi-scale feature data of all directional hole residual blocks in the channel dimension, integrates the multi-level features in the directional enhanced multi-scale feature data through a dense connection mechanism, and generates multi-level fused feature data;

[0060] ;

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

[0062] Residual connection module: adds the multi-level fusion feature data and the preliminary feature data element by element, retains the original information in the preliminary feature data through the residual connection mechanism, and generates residual fusion feature data;

[0063] ;

[0064] in, Represents residual fusion feature data, preliminary characterization data;

[0065] Activation processing module: performs activation processing on the residual fusion feature data to generate basic feature data.

[0066] Embodiment 4: This embodiment is based on embodiment 2. In this embodiment, a residual dense block is used, which specifically includes the following submodules:

[0067] Layer-by-layer convolution feature extraction module: Perform multi-level feature extraction and fusion on the preliminary feature data through residual dense blocks to generate residual dense feature data;

[0068] Feature splicing module: splices the residual dense feature data in the channel dimension, integrates the multi-level features in the residual dense feature data through the dense connection mechanism, reuses low-level features and high-level features, avoids information loss, and generates multi-level fusion feature data;

[0069] Residual connection module: adds multi-level fusion feature data and preliminary feature data element by element, retains the original information in the preliminary feature data through the residual connection mechanism, enhances the expression ability of fusion features, alleviates the gradient vanishing problem that may occur in deep networks, and generates residual fusion feature data;

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

[0071] Embodiment 5: This embodiment is based on embodiment 3. In this embodiment, the DGAP optimization module specifically includes the following submodules:

[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: allocates spatial attention weights to the attention feature data within the group through the position attention mechanism to obtain position attention feature data;

[0074] Channel weight optimization module: Generate the importance weight of each channel of the position attention feature data through the channel attention mechanism, and weight the importance weight of each channel and the position attention feature data to obtain the channel attention feature data.

[0075] Embodiment 6: This embodiment is based on embodiment 5. In this embodiment, the intra-group attention feature enhancement module specifically includes the following submodules:

[0076] Group global feature extraction module: Group the basic feature data to generate group feature data, perform direction-aware convolution on each group of feature data to extract horizontal features, vertical features, and diagonal features, and then perform global average pooling to extract global information and obtain group global features;

[0077] ;

[0078] in, represents the index of the feature grouping, represents the index of the direction-aware convolution kernel, Indicates The perceptual characteristics of group features in different directions, Indicates The grouping characteristic data of the group, Indicates Group characteristic data Perform direction-aware convolution, Indicates direction-aware convolution kernels, including horizontal, vertical, and diagonal convolution kernels. represents the convolution operation, represents the sum of all direction-aware convolution kernels;

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

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

[0081] Channel dimension restoration module: concatenates the weighted intra-group attention features of all groups, restores the channel dimension, and generates 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 submodules:

[0083] Group global feature extraction module: groups basic feature data to generate group feature data, performs global average pooling on each group of feature data in the group feature data, extracts global information, and obtains group global features;

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

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

[0086] Channel dimension restoration module: concatenates the weighted intra-group attention features of all groups, restores the channel dimension, and generates intra-group attention feature data.

[0087] The present invention and its embodiments are described above, which is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited to this. In short, if ordinary technicians in this field are inspired by it and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection 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, wherein 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; characterized in that: The system also includes a multi-scale neural network module; The multi-scale neural network module constructs a residual dense block through a dense connection mechanism and a residual connection mechanism, and introduces a direction-sensitive convolution and a hole convolution mechanism in the residual dense block to construct a direction hole residual block; optimizes the group attention mechanism through direction-aware convolution to construct an enhanced version of the group attention mechanism, and then combines position attention and channel attention to construct a DGAP module; optimizes jump connections through stacked convolutions to construct stacked jump connections; combines direction hole residual blocks, DGAP modules, and stacked jump connections to construct a multi-scale direction-aware brain tumor segmentation network, and uses the multi-scale direction-aware brain tumor segmentation network to process pre-processed 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 hole residual block feature extraction module, a DGAP optimization module, a stacked jump connection module, a resolution recovery decoding module, and a decoder.

2. The brain tumor segmentation system based on multi-scale neural network according to claim 1, characterized in that: The process of generating a brain tumor segmentation feature map by the multi-scale neural network module specifically includes the following contents: 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 directional hole residual block feature extraction module performs multi-layer feature extraction on the preliminary feature data through the directional hole residual block to generate basic feature data; The DGAP optimization module optimizes the basic feature data through the DGAP module and outputs the channel attention feature data; The stacked skip connection module processes the channel attention feature data through stacked skip connections to generate skip connection feature data; The resolution recovery decoding module passes the skip connection feature data to the decoder, and gradually upsamples it through deconvolution to restore the spatial resolution to be consistent with the spatial resolution of the preprocessed brain MRI image data, thereby obtaining a brain tumor segmentation feature map.

3. The brain tumor segmentation system based on multi-scale neural network according to claim 2, characterized in that: The directional hole residual block feature extraction module specifically includes the following submodules: Layer-by-layer convolution feature extraction module: extracts the horizontal convolution features, vertical convolution features and diagonal convolution features of the preliminary feature data through the directional hole residual block, and fuses the horizontal convolution features, vertical convolution features and diagonal convolution features of the preliminary feature data to generate direction-enhanced multi-scale feature data; Feature splicing module: splices the directional enhanced multi-scale feature data of all directional hole residual blocks in the channel dimension, integrates the multi-level features in the directional enhanced multi-scale feature data through a dense connection mechanism, and generates multi-level fused feature data; Residual connection module: adds the multi-level fusion feature data and the preliminary feature data element by element, retains the original information in the preliminary feature data through the residual connection mechanism, and generates 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 multi-scale neural network according to claim 2, characterized in that: The DGAP optimization module specifically Includes the following submodules: 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. Position attention allocation module: allocates spatial attention weights to the attention feature data within the group through the position attention mechanism to obtain position attention feature data; Channel weight optimization module: Generate the importance weight of each channel of the position attention feature data through the channel attention mechanism, and weight the importance weight of each channel and the position attention feature data to obtain the channel attention feature data.

5. The brain tumor segmentation system based on multi-scale neural network according to claim 4, characterized in that: The intra-group attention feature enhancement module specifically includes the following sub-modules: Group global feature extraction module: Group the basic feature data to generate group feature data, perform direction-aware convolution on each group of feature data to extract horizontal features, vertical features, and diagonal features, and then perform global average pooling to extract global information and obtain group global features; Attention weight generation module: the group global features are passed into the two-layer fully connected network to generate attention weights; Feature weighting module: applies attention weights to each group of feature data for weighted operations to generate weighted intra-group attention features; Channel dimension restoration module: concatenates the weighted intra-group attention features of all groups, restores the channel dimension, and generates intra-group attention feature data.

Images