A method for ischemic stroke image segmentation based on wavelet pooling neural network

Abstract

The present invention belongs to the field of medical imaging technology, and specifically relates to an ischemic stroke image segmentation method based on a wavelet pooling neural network. The following steps are as follows: first, using the U-Net convolutional neural network as the basic network architecture, a convolution residual module and a convolution residual are added to each convolution block of the encoding path; then the preprocessed feature map is input into the convolution block of the encoding path; then the first preliminary effective feature layer is obtained, and then losslessly sampled by the wavelet pooling module; then the channel attention SE module is connected after the wavelet pooling layer, and channel feature redirection is performed through squeezing, excitation and recalibration; then after each downsampling operation, the feature map is input into the pyramid expansion coordinate attention module, and multi-scale feature fusion is achieved through void convolution and coordinate attention; then in the decoding path, the feature map is input into the spatial attention module after each upsampling operation, and finally the segmentation result is output. The segmentation result output by the present invention is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical imaging, and in particular relates to an ischemic stroke image segmentation method based on a wavelet pooling neural network. Background Art

[0002] Ischemic stroke is a highly dangerous brain disease that causes disability or death in adults. In the face of this disease, rapid and accurate diagnosis and treatment are particularly critical to reducing brain damage. During the diagnostic process, accurately segmenting ischemic lesions from stroke medical images is the key to timely diagnosis and treatment. Doctors need to assess the condition and formulate treatment plans based on the morphology of the segmented lesions. The current traditional method is for doctors to manually segment lesions, but this method is time-consuming and labor-intensive, and is prone to missed or misdiagnosis of lesions due to personal subjective factors, which can delay the best time for treatment. In view of this, the development of a neural network image segmentation method that can quickly and automatically complete accurate lesion segmentation is of great significance for improving the efficiency and accuracy of diagnosis and treatment, and assisting doctors in implementing timely treatment within the golden treatment time. Summary of the Invention

[0003] The purpose of the present invention is to provide an ischemic stroke image segmentation method based on a wavelet pooling neural network. In response to the problems of downsampling information loss and insufficient multi-scale capture capability in the current deep learning method for ischemic stroke segmentation, a ischemic stroke image segmentation method based on a wavelet pooling neural network is proposed. This method achieves lossless wavelet decomposition to retain image details by designing a new downsampling module and a multi-scale feature fusion module, and realizes fine edge segmentation of large lesions, correct regional positioning of small lesions, and correct detection and differentiation of multi-region lesions.

[0004] The technical solutions adopted by the present invention are as follows:

[0005] A method for ischemic stroke image segmentation based on wavelet pooling neural network includes the following steps:

[0006] Step 1: Using the U-Net convolutional neural network as the basic network architecture, add the convolutional residual module Identity Block and the convolutional residual block Conv Block to each convolution block in the encoding path;

[0007] Step 2: The preprocessed 256×256×4 feature map is input into the convolution block of the encoding path. After passing through the first double-layer convolution residual block (Conv Block), the feature map is converted into a 256×256×64 feature map with inconsistent dimensions before and after. After passing through the convolution residual module (Identity Block) once more, the preliminary effective feature map of the first layer is obtained, at which point the dimensions before and after the convolution are consistent.

[0008] Step 3: Increase the number of feature map channels to 64 and obtain the first preliminary effective feature layer, which is then losslessly sampled through the wavelet pooling module;

[0009] Step 4: Connect the channel attention SE module after the wavelet pooling layer to redirect channel features through squeezing, excitation and recalibration;

[0010] Step 5: After each downsampling operation, the feature map is input into the pyramid expansion coordinate attention module, which uses dilated convolution and coordinate attention to achieve multi-scale feature fusion to improve the network's receptive field and feature expression capabilities;

[0011] Step 6: In the decoding path, the feature map after each upsampling operation is input to the spatial attention module;

[0012] Step 7: Introduce a full-scale skip connection module between each encoder layer and decoder layer. The full-scale skip connection module improves the efficiency of feature transmission and reuse by fusing features of different scales, thereby capturing both fine-grained and coarse-grained semantics at all scales and obtaining a more accurate segmentation result.

[0013] Preferably, in step 1, the residual module algorithm is described as follows: the residual module consists of two consecutive convolutional layers and a skip connection; the input feature passes through the first convolutional layer, then passes through the ReLU activation function, and the output result is passed to the second convolutional layer; the output of the second convolutional layer is added to the input feature to obtain the output of the residual module, which can avoid the gradient vanishing problem and improve the training efficiency and performance of the model.

[0014] Preferably, the step 3 specifically includes the following steps:

[0015] Step 301: In the wavelet pooling module, the discrete wavelet decomposition process is described as follows: the input image is first subjected to a one-dimensional wavelet transform in the horizontal direction, i.e., the input image is simultaneously subjected to high-pass and low-pass filters to extract the high and low frequency information in the horizontal direction; the high and low frequency components obtained in the first step are respectively subjected to a one-dimensional wavelet transform in the vertical direction, to obtain four image components after two-dimensional wavelet decomposition;

[0016] Step 302: The four image components with half the resolution obtained by the two-dimensional discrete wavelet transform are input to the pooling layer for processing. While the pooling layer expands the receptive field, it also compensates for the information loss caused by traditional downsampling, allowing the network to utilize high-frequency components in the image, such as edges, contours, and other areas with large changes. The resolution of the decomposed sub-image is halved, and the number of channels in the feature map is quadrupled. That is, by allocating the information encoding in the spatial dimension to the channel dimension, the preservation of all information is also guaranteed.

[0017] Step 303: The sub-images obtained by the two downsampling methods of maximum pooling and discrete wavelet decomposition are stacked in the channel dimension and used as the input of the next layer of network.

[0018] Preferably, in step 301, the four image components are: the component HH filtered by horizontal high frequency and vertical high frequency represents the details of the diagonal direction of the image, the component HL filtered by horizontal high frequency and vertical low frequency represents the horizontal edge information of the image, the component LH filtered by horizontal low frequency and vertical high frequency represents the vertical edge information of the image, and the component LL filtered by horizontal low frequency and vertical low frequency represents the approximate information of the image.

[0019] Preferably, the step 4 comprises the following steps:

[0020] Step 401: Given a feature map x with an input size of H×W×C, where H is the height of a single feature map, W is the width, and C is the number of channels; a squeezing operation compresses the input feature map along the spatial length and width dimensions, i.e., a global average pooling operation is performed, compressing each two-dimensional feature map into a real number and arranging them along the channel direction, thus obtaining a global angular feature vector 1×1×C with a spatial direction, which represents the description of each feature map on the channel. The formula for the channel is as follows:

[0021]

[0022] Among them, x c represents the cth channel of the input feature x; z c is the global average value of the c-th channel;

[0023] Step 402: In the excitation operation, the 1×1×C global feature vector obtained by the squeeze operation is first passed through a fully connected layer with a small number of neurons for inter-channel information interaction learning, while reducing the channel dimension to C / r to reduce computational complexity and parameter quantity; then it is activated by a ReLU layer; and then it is further passed through a fully connected layer of length C for further channel information fusion and weight learning, and activated by a sigmoid function to reduce the weight range to between 0 and 1. The size of the weight represents the importance of the corresponding channel feature map; the excitation formula is as follows:

[0024] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)),

[0025] Where W1 is a fully connected layer with dimension C / r, r is the drop ratio, W2 is a fully connected layer with dimension C, δ is the ReLU activation function, σ is the sigmoid activation function; s is the obtained attention weight;

[0026] Step 403: In the channel feature recalibration operation, the attention weight s obtained through the fully connected layer and nonlinear learning is multiplied by the feature map x before the squeeze operation. The formula is:

[0027]

[0028] Where, This is the final feature map with channel attention applied.

[0029] Preferably, the step 5 comprises the following steps:

[0030] Step 501: Perform multiple convolution operations with different dilation rates on the preliminary effective feature map. The dilation rates selected by the module are 1, 2, 3, and 4, respectively, to obtain a multi-scale feature map.

[0031] Step 502: The multi-scale feature maps obtained by the four dilated convolutions are concatenated according to the channel dimension, and then a 1×1 convolution is performed to interactively fuse the multi-scale information of each channel;

[0032] Step 503: The above feature map is subjected to coordinate attention again to redistribute the attention of channel information and spatial information, and finally an effective feature layer with multi-scale robustness and coordinate sensitivity is obtained.

[0033] Preferably, in step 6: first, global feature extraction is performed on the input feature map of size H×W×C in the channel dimension, that is, the values ​​on all channels corresponding to each feature point are averaged and extreme values ​​are calculated to obtain two H×W×1 feature maps; the two feature maps representing global information are stacked in the channel direction, and then a 7×7 convolution operation is performed to reduce the number of channels to 1, and then a sigmoid function is used for activation to transform the value of the feature map to the range of 0 to 1, that is, to obtain an attention weight map in the spatial direction; the formula is:

[0034]

[0035] Among them, f 7×7 represents a convolution kernel of size 7×7, and σ represents the sigmoid activation function; Represents the two-dimensional feature map obtained by taking the average value of the channel dimension, Represents the two-dimensional feature map obtained by taking the maximum value of the channel dimension;

[0036] By performing matrix multiplication on the spatial attention weight map and the input feature map, we can obtain the output feature map recalibrated in the spatial direction; then a Conv Block is used to deepen the feature extraction, and the above process is repeated to obtain the final effective feature layer, and a layer of convolution and activation is performed to obtain the preliminary segmentation result.

[0037] Preferably, in step 7, the resolution of the feature maps of different scales in the coding path is first unified; the feature maps with more underlying information are and After 4 times non-overlapping maximum pooling downsampling and 2 times downsampling respectively, the feature maps after pooling are further extracted by 3×3 convolution; the feature maps of the same scale are After only 3×3 convolution, the decoding path feature map with small resolution but relatively high-level semantic information is transformed into and The five scale feature maps are upsampled by four and two times through bilinear interpolation, and each undergoes a 3×3 convolution. The resulting five scale feature maps with unified resolution are stacked in the channel dimension and uniformly passed through a 3×3 convolution layer for information fusion and channel interaction. This aggregates detailed information such as target boundaries contained in the underlying feature maps with target category and positioning information contained in the higher-level feature maps.

[0038] In the encoding stage, steps 1 to 5 are repeated until five preliminary valid feature layers with resolutions of 256×256, 128×128, 64×64, 32×32, and 16×16 are obtained; in the decoding stage, step 6 is repeated, and step 7 runs through all steps to obtain the final valid feature layer, and the final segmentation result is obtained after a layer of convolution and activation.

[0039] The technical effects achieved by the present invention are:

[0040] In the present invention, in order to solve the problem of information loss caused by downsampling, a method of lossless two-dimensional discrete wavelet pooling of feature maps is adopted to reduce the dimensionality of the feature maps in the spatial dimension, thereby achieving the effect of increasing the network receptive field of the traditional pooling layer while retaining all feature map information by increasing the number of feature map channels. In addition, since stroke images have individual differences according to different patients, resulting in different lesion locations, shapes, and sizes, the present invention effectively improves the adaptability of the network to inputs with lesions of different sizes, and can make full use of the contextual information between pixels from a global perspective to make more accurate judgments on segmentation edges, taking into account the segmentation of lesions of different sizes, and can make good lesion distinctions for inputs with multiple lesion areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is an overall flow chart of an ischemic stroke image segmentation method based on wavelet pooling neural network proposed in the present invention.

[0042] Figure 2The present invention proposes an image segmentation network model for an ischemic stroke image segmentation method based on a wavelet pooling neural network, wherein the input is a source image and the output is a segmentation result.

[0043] Figure 3 This is an example of the comparison between the stroke lesion area segmentation effect diagram of the ischemic stroke image segmentation method based on wavelet pooling neural network proposed in the present invention and the segmentation results of other segmentation algorithms. DETAILED DESCRIPTION

[0044] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0045] like Figure 1 As shown, a method for ischemic stroke image segmentation based on wavelet pooling neural network includes the following steps:

[0046] Step 1: Using the U-Net convolutional neural network as the basic network architecture, add the convolution residual module Identity Block and the convolution residual block Conv Block to each convolution block in the encoding path; the residual module algorithm is described as follows: the residual module consists of two consecutive convolution layers and a skip connection; the input features pass through the first convolution layer, then through the ReLU activation function, and the output result is passed to the second convolution layer; the output of the second convolution layer is added to the input features to obtain the output of the residual module, which can avoid the gradient vanishing problem and improve the training efficiency and performance of the model

[0047] Step 2: The preprocessed 256×256×4 feature map is input into the convolution block of the encoding path. After passing through the first double-layer convolution residual block Conv Block, the feature map is obtained as a 256×256×64 feature map with inconsistent front and back dimensions. It then passes through the convolution residual module Identity Block once more to obtain the preliminary effective feature map of the first layer. At this time, the dimensions before and after convolution are consistent. Identity Block is the identity mapping module used when the input and output dimensions are the same, and Conv Block is the convolution residual module used when the input and output dimensions are different.

[0048] In this embodiment, each convolutional block in the encoding path: In the UNet architecture, the encoder portion typically consists of multiple convolutional blocks. Each convolutional block typically contains two consecutive 3x3 convolutional layers, each followed by a ReLU activation function and a 2x2 max pooling layer for downsampling. The main purpose of the convolutional block is to extract features and reduce spatial dimensions so that the model can capture different levels of abstraction of the input image.

[0049] Residual Module: The residual module is designed to solve the gradient disappearance / explosion problem in deep networks.

[0050] Convolutional Residual Module Identity Block: Identity Block is a special residual module in which the size and depth of the input feature map are the same as the output feature map, so the input can be directly added to the output. This means that if the data dimension on the main path (i.e. the path through the convolutional layer) does not change, then the identity mapping can be used to simply add the input to the output, thus forming a residual connection.

[0051] Convolutional Residual Block (Conv Block): A Conv Block is also a type of residual module, but it involves a change in the size or depth of the input feature map. In this case, the input cannot be directly added to the output because the dimensions do not match. Therefore, the Conv Block includes an additional convolutional layer to adjust the input dimension to match the output on the main path, and then perform the addition operation.

[0052] In this embodiment, when the input and output dimensions of the residual block are the same, the input and the residual map are directly added at the output end, that is, the identity mapping module Identity Block used when the input and output dimensions are the same is used; when the input and output dimensions are different, the input needs to be dimensionally adjusted through a 1×1 convolution to make the two dimensions the same, and then added to the residual map at the output end, that is, the convolution residual module Conv Block used when the input and output dimensions are different is used; the above two residual modules are used in the network of the present invention to deepen the network depth and feature extraction level.

[0053] Step 3: Increase the number of feature map channels to 64 and obtain the first preliminary effective feature layer, which is then losslessly sampled through the wavelet pooling module. This includes the following steps:

[0054] Step 301: In the wavelet pooling module, the discrete wavelet decomposition process is described as follows: the input image is first subjected to a one-dimensional wavelet transform in the horizontal direction, that is, it is simultaneously subjected to high-pass and low-pass filters to extract the high- and low-frequency information in the horizontal direction; the high- and low-frequency components obtained in the first step are subjected to a one-dimensional wavelet transform in the vertical direction respectively, that is, four image components after two-dimensional wavelet decomposition are obtained; the four image components are: the component HH after horizontal high-frequency and vertical high-frequency filtering represents the details in the diagonal direction of the image, the component HL after horizontal high-frequency and vertical low-frequency filtering represents the horizontal edge information of the image, the component LH after horizontal low-frequency and vertical high-frequency filtering represents the vertical edge information of the image, and the component LL after horizontal low-frequency and vertical low-frequency filtering represents the approximate information of the image

[0055] Step 302: The four image components with half the resolution obtained by the two-dimensional discrete wavelet transform are input to the pooling layer for processing. While the pooling layer expands the receptive field, it also compensates for the information loss caused by traditional downsampling, allowing the network to utilize high-frequency components in the image, such as edges, contours, and other areas with large changes. The resolution of the decomposed sub-image is halved, and the number of channels in the feature map is quadrupled. That is, by allocating the information encoding in the spatial dimension to the channel dimension, the preservation of all information is also guaranteed.

[0056] Step 303: The sub-images obtained by the two downsampling methods of maximum pooling and discrete wavelet decomposition are stacked in the channel dimension and used as the input of the next layer of network;

[0057] Step 4: Connect the channel attention SE module after the wavelet pooling layer to redirect channel features through squeezing, excitation and recalibration;

[0058] Step 401: Given a feature map x with an input size of H×W×C, where H is the height of a single feature map, W is the width, and C is the number of channels; a squeezing operation compresses the input feature map along the spatial length and width dimensions, i.e., a global average pooling operation is performed, compressing each two-dimensional feature map into a real number and arranging them along the channel direction, thus obtaining a global angular feature vector 1×1×C with a spatial direction, which represents the description of each feature map on the channel. The formula for the channel is as follows:

[0059]

[0060] Among them, x c represents the cth channel of the input feature x; z c is the global average value of the c-th channel;

[0061] Step 402: In the excitation operation, the 1×1×C global feature vector obtained by the squeeze operation is first passed through a fully connected layer with a small number of neurons for inter-channel information interaction learning, while reducing the channel dimension to C / r to reduce computational complexity and parameter quantity; then it is activated by a ReLU layer; and then it is further passed through a fully connected layer of length C for further channel information fusion and weight learning, and activated by a sigmoid function to reduce the weight range to between 0 and 1. The size of the weight represents the importance of the corresponding channel feature map; the excitation formula is as follows:

[0062] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)),

[0063] Where W1 is a fully connected layer with dimension C / r, r is the drop ratio, W2 is a fully connected layer with dimension C, δ is the ReLU activation function, σ is the sigmoid activation function; s is the obtained attention weight;

[0064] Step 403: In the channel feature recalibration operation, the attention weight s obtained through the fully connected layer and nonlinear learning is multiplied by the feature map x before the squeeze operation. The formula is:

[0065]

[0066] Where, This is the final feature map with channel attention applied.

[0067] Step 5: After each downsampling operation, the feature map is input into the pyramid expansion coordinate attention module, which uses dilated convolution and coordinate attention to achieve multi-scale feature fusion to improve the network's receptive field and feature expression capabilities;

[0068] Step 501: Perform multiple convolution operations with different dilation rates on the preliminary effective feature map. The dilation rates selected by the module are 1, 2, 3, and 4, respectively, to obtain a multi-scale feature map.

[0069] Step 502: The multi-scale feature maps obtained by the four dilated convolutions are concatenated according to the channel dimension, and then a 1×1 convolution is performed to interactively fuse the multi-scale information of each channel;

[0070] Step 503: The above feature map is subjected to coordinate attention to redistribute the attention of channel information and spatial information, and finally an effective feature layer with multi-scale robustness and coordinate sensitivity is obtained;

[0071] Step 6: In the decoding path, the feature map after each upsampling operation is input to the spatial attention module;

[0072] In step 6: First, global feature extraction is performed on the input feature map of size H×W×C in the channel dimension, that is, the values ​​on all channels corresponding to each feature point are averaged and extreme values ​​are calculated to obtain two H×W×1 feature maps; the two feature maps representing global information are stacked in the channel direction, and then a 7×7 convolution operation is performed to reduce the number of channels to 1. Subsequently, a sigmoid function is used for activation to transform the value of the feature map to the range of 0 to 1, that is, the attention weight map in the spatial direction is obtained; its formula is:

[0073]

[0074] Among them, f 7×7 represents a convolution kernel of size 7×7, and σ represents the sigmoid activation function; Represents the two-dimensional feature map obtained by taking the average value of the channel dimension, Represents the two-dimensional feature map obtained by taking the maximum value of the channel dimension;

[0075] Performing matrix multiplication on the spatial attention weight map and the input feature map will yield a recalibrated output feature map in the spatial direction. This is then followed by a Conv Block to deepen feature extraction, and the above process is repeated to obtain the final effective feature layer. A further layer of convolution and activation is then performed to obtain the preliminary segmentation result.

[0076] Step 7: A full-scale skip connection module is introduced between each encoder layer and decoder layer. The full-scale skip connection module improves the efficiency of feature transmission and reuse by fusing features of different scales, thereby capturing both fine-grained and coarse-grained semantics at all scales and obtaining a more accurate segmentation result.

[0077] In step 7, since the stacking of channel dimensions requires the resolution of feature maps to be consistent, the resolutions of feature maps of different scales in the encoding path are first unified; the feature maps with more underlying information are and After 4 times non-overlapping maximum pooling downsampling and 2 times downsampling respectively, the feature maps after pooling are further extracted by 3×3 convolution; the feature maps of the same scale are After only 3×3 convolution, the decoding path feature map with small resolution but relatively high-level semantic information is transformed into and The five scale feature maps are upsampled by four and two times through bilinear interpolation, and each undergoes a 3×3 convolution. The resulting five scale feature maps with unified resolution are stacked in the channel dimension and uniformly passed through a 3×3 convolution layer for information fusion and channel interaction. This aggregates detailed information such as target boundaries contained in the underlying feature maps with target category and positioning information contained in the higher-level feature maps.

[0078] In the encoding stage, steps 1 to 5 are repeated until five preliminary valid feature layers with resolutions of 256×256, 128×128, 64×64, 32×32, and 16×16 are obtained; in the decoding stage, step 6 is repeated, and step 7 runs through all steps to obtain the final valid feature layer, and the final segmentation result is obtained after a layer of convolution and activation.

[0079] The technology of the present invention is implemented in the dataset ISLES2022, which provides a total of 250 annotated multimodal MR image training sets, including three modalities: fluid-attenuated inversion recovery (FLAIR), magnetic resonance diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC). Among them, DWI utilizes the difference in the diffusivity of intracellular and extracellular water, and can detect cerebral ischemia earlier than other modalities. It is the most sensitive sequence for detecting acute ischemic stroke, also known as the "stroke sequence". It is highly sensitive to ischemic areas within 15 to 30 minutes after the onset of symptoms, clearly showing high-intensity signal areas. The lesions in the image have a high contrast with normal tissue, and the lesions are clearly visible. This embodiment uses DWI modality images in the case for segmentation, and each case contains more than 22 to 70 two-dimensional slices. Regarding the model process, the number of network iterations used for training is set to 300 times, the batch size is 24 times, Adam is used as the optimizer, and the initial learning rate is set to 0.0001.

[0080] In order to verify the effectiveness of the technology of the present invention, Figure 3 The segmentation results of our technique on the ISLES2022 dataset are presented and compared with the U-Net, U-Net++, and U-Net3++ methods. Red arrows indicate incorrectly segmented areas. Our technique can segment lesions of varying sizes and effectively distinguish lesions in inputs with multiple lesions, achieving performance improvements over the baseline network methods. These results demonstrate the effectiveness of our technique.

[0081] To verify the effectiveness of the present invention, Table 1 compares the present invention with existing technologies using the metrics Dice (the intersection of twice the predicted area and the true area divided by the pixel sum of the predicted and true areas), Precision (accuracy), Recall (measures the ratio of the area correctly predicted by the model to the pixel sum of the true area), and Intersection over Union (IoU), which measures the ratio of the area correctly predicted to the area that is the union of the predicted and true values. As shown in Table 1, the present invention achieves superior results across multiple metrics, thus demonstrating its effectiveness.

[0082] Table 1 Quantitative comparison results between the invention technology and the existing technology

[0083] Method Dice IoU Recall PRE SegNet 0.7863 0.6522 0.7614 0.7712 PSPNet 0.8213 0.7004 0.8006 0.8080 DeepLabv3+ 0.8301 0.7124 0.7346 0.8333 U-Net 0.8366 0.7250 0.7393 0.8646 U-Net++ 0.8553 0.7511 0.8374 0.8448 U-Net3+ 0.8618 0.7614 0.8306 0.8655 The technology of the present invention 0.8863 0.8005 0.8781 0.9154

[0084] In order to solve the problem of information loss caused by downsampling, the present invention adopts a method of lossless two-dimensional discrete wavelet pooling of feature maps, which reduces the spatial dimension of the feature maps, thereby achieving the effect of increasing the network receptive field of the traditional pooling layer, while retaining all feature map information by increasing the number of feature map channels. In addition, since stroke images have individual differences according to different patients, resulting in different lesion locations, shapes, and sizes, the present invention effectively improves the network's adaptability to inputs with lesions of different sizes, and can make full use of the contextual information between pixels from a global perspective to make more accurate judgments on segmentation edges, taking into account the segmentation of lesions of different sizes, and can make good lesion distinctions for inputs with multiple lesion areas.

[0085] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A method for ischemic stroke image segmentation based on wavelet pooling neural network, characterized by: The following steps are involved: Step 1: Using the U-Net convolutional neural network as the basic network architecture, add the convolutional residual module Identity Block and the convolutional residual block Conv Block to each convolution block in the encoding path; Step 2: The preprocessed 256×256×4 feature map is input into the convolution block of the encoding path. After the feature map passes through the first double-layer convolution residual block Conv Block, a 256×256×64 feature map with inconsistent dimensions is obtained; it then passes through the convolution residual module Identity Block once more; Get the preliminary effective feature map of the first layer, at this time the dimensions before and after convolution are consistent; Step 3: Increase the number of feature map channels to 64 and obtain the first preliminary effective feature layer, which is then losslessly sampled through the wavelet pooling module; Step 4: Connect the channel attention SE module after the wavelet pooling layer to redirect channel features through squeezing, excitation and recalibration; Step 5: After each downsampling operation, the feature map is input into the pyramid dilation coordinate attention module to achieve multi-scale feature fusion through void convolution and coordinate attention; Step 6: In the decoding path, the feature map after each upsampling operation is input to the spatial attention module; Step 7: A full-scale skip connection module is introduced between each encoder layer and decoder layer. The full-scale skip connection module fuses features of different scales to capture fine-grained and coarse-grained semantics at all scales and obtain the final segmentation result.

2. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 1 is characterized in that: In step 1, the residual module algorithm is described as follows: the residual module consists of two consecutive convolutional layers and a skip connection; the input features pass through the first convolutional layer, then through the ReLU activation function, and the output result is passed to the second convolutional layer; the output of the second convolutional layer is added to the input features to obtain the output of the residual module.

3. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 1, characterized in that: The step 3 specifically includes the following steps: Step 301: In the wavelet pooling module, the discrete wavelet decomposition process is described as follows: the input image is first subjected to a one-dimensional wavelet transform in the horizontal direction, i.e., the input image is simultaneously subjected to high-pass and low-pass filters to extract the high and low frequency information in the horizontal direction; the high and low frequency components obtained in the first step are respectively subjected to a one-dimensional wavelet transform in the vertical direction, to obtain four image components after two-dimensional wavelet decomposition; Step 302: input the four image components with half the resolution obtained by the two-dimensional discrete wavelet transform into the pooling layer for processing; Step 303: The sub-images obtained by the two downsampling methods of maximum pooling and discrete wavelet decomposition are stacked in the channel dimension and used as the input of the next layer of network.

4. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 3 is characterized in that: In step 301, the four image components are: the component HH filtered by horizontal high frequency and vertical high frequency represents the details of the diagonal direction of the image, the component HL filtered by horizontal high frequency and vertical low frequency represents the horizontal edge information of the image, the component LH filtered by horizontal low frequency and vertical high frequency represents the vertical edge information of the image, and the component LL filtered by horizontal low frequency and vertical low frequency represents the approximate information of the image.

5. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 4 is characterized in that: The step 4 comprises the following steps: Step 401: Given a feature map x with an input size of H×W×C, where H is the height of a single feature map, W is the width, and C is the number of channels; a squeezing operation compresses the input feature map along the spatial length and width dimensions, i.e., a global average pooling operation is performed, compressing each two-dimensional feature map into a real number and arranging them along the channel direction, thus obtaining a global angular feature vector 1×1×C with a spatial direction, which represents the description of each feature map on the channel. The formula for the channel is as follows: Among them, x c represents the cth channel of the input feature x; z c is the global average value of the c-th channel; Step 402: In the excitation operation, the 1×1×C global feature vector obtained by the squeeze operation is first passed through a fully connected layer with a small number of neurons for inter-channel information interaction learning, while reducing the channel dimension to C / r; then it is activated by a ReLU layer; and then it is further passed through a fully connected layer of length C for further channel information fusion and weight learning, and activated by a sigmoid function to reduce the weight range to between 0 and 1. The size of the weight represents the importance of the corresponding channel feature map. The excitation formula is as follows: s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)), Where W1 is a fully connected layer with dimension C / r, r is the drop ratio, W2 is a fully connected layer with dimension C, δ is the ReLU activation function, σ is the sigmoid activation function; s is the obtained attention weight; Step 403: In the channel feature recalibration operation, the attention weight s obtained through the fully connected layer and nonlinear learning is multiplied by the feature map x before the squeeze operation. The formula is: Where, This is the final feature map with channel attention applied.

6. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 5, characterized in that: The step 5 comprises the following steps: Step 501: Perform multiple convolution operations with different dilation rates on the preliminary effective feature map. The dilation rates selected by the module are 1, 2, 3, and 4, respectively, to obtain a multi-scale feature map. Step 502: The multi-scale feature maps obtained by the four dilated convolutions are concatenated according to the channel dimension, and then a 1×1 convolution is performed to interactively fuse the multi-scale information of each channel; Step 503: The above feature map is subjected to coordinate attention again to redistribute the attention of channel information and spatial information, and finally an effective feature layer with multi-scale robustness and coordinate sensitivity is obtained.

7. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 6, characterized in that: In step 6: first, global feature extraction is performed on the input feature map of size H×W×C in the channel dimension, that is, the values ​​on all channels corresponding to each feature point are averaged and the extreme value is calculated to obtain two H×W×1 feature maps; the two feature maps representing global information are stacked in the channel direction, and then a 7×7 convolution operation is performed to reduce the number of channels to 1, and then a sigmoid function is used for activation to transform the value of the feature map to the range of 0 to 1, that is, the attention weight map in the spatial direction is obtained; its formula is: Among them, f 7×7 represents a convolution kernel of size 7×7, and σ represents the sigmoid activation function; Represents the two-dimensional feature map obtained by taking the average value of the channel dimension, Represents the two-dimensional feature map obtained by taking the maximum value of the channel dimension; By performing matrix multiplication on the spatial attention weight map and the input feature map, we can obtain the output feature map recalibrated in the spatial direction; then we go through a ConvBlock to deepen the feature extraction, and then repeat the above process to obtain the final effective feature layer, and then get the preliminary segmentation result after a layer of convolution and activation.

8. The ischemic stroke image segmentation method based on wavelet pooling neural network according to claim 7, characterized in that: In step 7, the resolution of the feature maps of different scales in the coding path is first unified; the feature maps with more underlying information are and After 4 times non-overlapping maximum pooling downsampling and 2 times downsampling respectively, the feature maps after pooling are further extracted by 3×3 convolution; the feature maps of the same scale are After only 3×3 convolution, the decoding path feature map with small resolution but relatively high-level semantic information is transformed into and The five scale feature maps are upsampled by four and two times through bilinear interpolation, and each undergoes a 3×3 convolution. The resulting five scale feature maps with unified resolution are stacked in the channel dimension and uniformly passed through a 3×3 convolution layer for information fusion and channel interaction. This aggregates detailed information such as target boundaries contained in the underlying feature maps with target category and positioning information contained in the higher-level feature maps. In the encoding stage, steps 1 to 5 are repeated until five preliminary valid feature layers with resolutions of 256×256, 128×128, 64×64, 32×32, and 16×16 are obtained; In the decoding stage, step 6 is repeated, and step 7 runs through all steps to obtain the final effective feature layer, and the final segmentation result is obtained after a layer of convolution and activation.

Images