Processing method, system and equipment of lung cancer bone scanning image and medium

Abstract

The invention discloses a lung cancer bone scanning image processing method, system and device and a medium, and relates to the technical field of image processing, and the method comprises the steps: obtaining a lung cancer bone scanning image; constructing a lung cancer bone imaging data model by adopting an encoder-decoder, and inputting the lung cancer bone scanning image into the lung cancer bone imaging data model for processing to obtain a processed image; in the encoder, a shape guiding strategy is used for learning specific features of the rectangular size of an image, global context information is captured in two axial directions of the image through a bidirectional pooling mechanism, focus information and focus boundary details are extracted between the encoder and the decoder through a multi-scale boundary guiding strategy, and in the decoder, focus boundary details are extracted through a multi-scale boundary guiding strategy. And reconstructing encoder output features, lesion information and lesion boundary details through convolution and up-sampling, and outputting a final image. The invention provides a technology combining a shape guide strategy, a multi-scale boundary guide module and bidirectional pooling, the problem of boundary blur is improved, and the detection effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method, system, device and medium for processing lung cancer bone scan images. Background Art

[0002] Lung cancer is one of the malignant tumors with the highest incidence rate and also one of the malignant tumors with the highest mortality rate. Therefore, early detection of lung cancer bone metastasis is not only crucial for treating the disease, but also of great significance for disease staging, outcome prediction and treatment planning.

[0003] Single Photon Emission Computed Tomography (SPECT) is a non-invasive examination method used to evaluate the functional state of the human body and obtain internal information of the organism by detecting single photons emitted by radioactive isotopes. SPECT and Positron Emission Tomography (PET) scans are commonly used to measure bone metastasis. PET uses a drug labeled with a positron radioactive substance to reflect the glucose metabolism process in the human body, but it has the disadvantages of lack of structural information, high equipment and acquisition costs. SPECT, on the other hand, can select different imaging agents according to needs to provide whole body / local functional imaging, with a flexible inspection mechanism and high cost performance.

[0004] Computer-Aided Diagnosis (CAD) systems enhance the interpretability of SPECT images through algorithms to make up for the subjectivity and efficiency bottleneck of manual interpretation. Images obtained with low resolution, motion artifacts, unbalanced data distribution and scanning parameter differences will reduce the recognition ability of CAD systems, and may misjudge normal tissues as lesions, and some lesions (such as early tumors, microcalcifications) may be missed. Especially when the image quality is poor or the lesion contrast is low, it may lead to delayed treatment. Summary of the Invention

[0005] The purpose of the present invention is to provide a method, system, device and medium for processing lung cancer bone scan images in view of the deficiencies of the above-mentioned prior art, so as to solve the problems in the prior art.

[0006] The present invention specifically provides the following technical solutions:

[0007] A method for processing lung cancer bone scan images includes the following steps:

[0008] Obtain a lung cancer bone scan image;

[0009] Construct a lung cancer bone imaging data model using an encoder, a decoder and a multi-scale boundary guidance module, and input the lung cancer bone scan image into the lung cancer bone imaging data model for processing to obtain a processed image;

[0010] Input the lung cancer bone scan image into the lung cancer bone imaging data model for processing, including:

[0011] In the encoder, using rectangular dilated convolution as the shape guidance strategy, extract the output feature map unique to the rectangular size for the lung cancer bone scan image, and capture the global context information in two axial directions of the output feature map through the bidirectional pooling mechanism to obtain the encoder output feature map;

[0012] In the multi-scale boundary guidance module, perform multi-scale extraction and Laplacian pyramid processing on the lung cancer bone scan image to obtain the Laplacian pyramid feature map, and use the feature map of each stage of the encoder, the Laplacian pyramid feature map, and the decoding prediction feature of the i+1th layer as inputs to extract the lesion information and lesion boundary details;

[0013] In the decoder, reconstruct the encoder output feature map, lesion information, and lesion boundary details through convolution and upsampling, and output the final image.

[0014] Preferably, using rectangular dilated convolution as the shape guidance strategy, extract the output feature map unique to the rectangular size for the lung cancer bone scan image, including:

[0015] Perform 1×1 convolution processing and 7×1 convolution processing on the lung cancer bone scan image to obtain a feature cube The specific expression is:

[0016] U * =Conv7×1(Conv1×1(u),dilation=o * );

[0017] where, o * is the dilation rate;

[0018] Compress this feature cube into a 1×1×C channel description vector through the F sq (·) operation where, the specific expression of the corresponding channel description element z c ∈Z is:

[0019]

[0020] where, z c represents the output value of channel c after global average pooling; F sq (·) represents the function of the Squeeze operation, which is used to compress the spatial information of the feature map; u cDenote the two-dimensional feature map of the input feature map on channel c; H represents the height of the input feature map; W represents the width of the input feature map; m represents the row index in the feature map; n represents the column index in the feature map; u c (m, n) represents the feature value of channel c at the m-th row and the n-th column; the channel description vector is input into the activation function F ex (·, W) to obtain the activated channel description vector, where the specific expression of the activated channel description vector form is defined as:

[0021] s c = F ex (z c , W) = σ(W2δ(W1z c ));

[0022] where δ is the Sigmoid(·) function, σ is the ReLU(·) equation, r is the squeeze ratio;

[0023] Adjust the feature cube U through the activated channel description vector * to obtain the adjusted output feature map v; among them, the specific expression of the c-th output feature map v c is:

[0024] v c = F scale (u c , s c ) = s c u c ;

[0025] Select an output feature map v with a fineness higher than the threshold from the output feature map v f , where the specific expression of the output feature map v f is:

[0026]

[0027] Among them, represents element-wise multiplication, and f s (v) is a spatial function, defined as shown in the formula:

[0028] f s (v) = σ(f k×k ([AvgPool(v); MaxPool(v)]));

[0029] where δ is the Sigmoid(·) function, AvgPool(·) is the average pooling operation, MaxPool(·) is the max pooling operation, and f k×k is a convolution with a kernel size of k×k;

[0030] The output feature map v with a fineness higher than the threshold f is processed through a 1×1 convolutional layer with 2C channels and then through a 1×1 convolutional layer with C channels to obtain the final output feature map, expressed as:

[0031] v f = f 1×1,c (f 1×1，2c (v f ))

[0032] Preferably, the cross-bidirectional pooling mechanism captures global context information in two axial directions of the output feature map to obtain the encoder output feature map, including:

[0033] Extract the information in two directions of the output feature map through horizontal pooling and vertical pooling;

[0034] Merge the information obtained from the two poolings to generate a preliminary rectangular region of interest, and adjust the region shape of the preliminary rectangular region of interest through a shape self-calibration function to obtain the adjusted rectangular region of interest y, and the specific expression is:

[0035]

[0036] where the symbol δ represents the Sigmoid function, h represents horizontal pooling, and z represents vertical pooling;

[0037] Extract the local features of the output feature map x through a 3×3 depth convolution, and multiply the local features with y pixel by pixel to obtain the global feature F(x), and the specific expression is:

[0038] F(x) = conv 3×3 (x) ⊙ y(x);

[0039] Further optimize the global feature through batch normalization and a multi-layer perceptron to generate the encoder output feature map The specific expression is:

[0040] output = F(x) + MLP(BN(F(x)));

[0041] where BN represents batch normalization, MLP represents a multi-layer perceptron, and output is the encoder output feature map

[0042] Preferably, the multi-scale extraction and Laplacian pyramid processing of the lung cancer bone scan image to obtain the Laplacian pyramid feature map includes:

[0043] Perform multi-scale extraction on the lung cancer bone scan image to obtain the output feature cube I, and the specific expression is:

[0044] I = H × W × (C + C2 + C3 + CP)

[0045] Where C' = C1 + C2 + C3 + C P is the output splicing result of the four branches of the multi-scale learning block, and C1, C2, and C P respectively represent the number of channels in the feature maps corresponding to convolutional kernels with sizes of 1×1, 3×3, and 5×5;

[0046] Process the output feature cube through the Laplacian pyramid to obtain the Laplacian pyramid feature map I k , and the specific expression is:

[0047]

[0048] Where I is the lung cancer bone scan image, g is the convolutional operator with a Gaussian filter, d is the 2-fold downsampling operation, and k represents the k-th layer L of the Laplacian pyramid k , and the k-th layer L of the Laplacian pyramid k is obtained by subtracting the upsampled version u of the next top layer I k from the current layer L k+1 , and the specific expression is:

[0049] L k = I k - u(I k+1 ), f l = L1(I);

[0050] Where f l is used as the high-frequency feature, and L1(I) represents the first layer feature of the Laplacian pyramid.

[0051] Preferably, the extraction of lesion information and lesion boundary details includes:

[0052] Generate a reverse attention map and a boundary attention map from the decoding prediction features of the i+1-th layer and multiply the high-frequency feature f i l , the boundary attention map, and the decoding feature map with the encoding feature f i e respectively for convolution to obtain the convolved feature map f i m , and the specific expression is:

[0053]

[0054] Use the attention mask on the encoding feature f ie Process the feature map after convolution to obtain the optimized feature result f i a , and the specific expression is:

[0055]

[0056] Among them, σ represents the Sigmoid function;

[0057] Input the optimized feature result into the global context block GC for processing to obtain the features f of the lesion information and the lesion boundary details i d , and the specific expression is:

[0058]

[0059] Preferably, the acquisition process of the high-frequency feature f i l is specifically as follows:

[0060] Perform Gaussian filtering on the (i - 1)th layer, and perform 2-fold downsampling on the result of Gaussian filtering to obtain the high-frequency feature f i l , and the specific expression is:

[0061]

[0062] Among them, is the initial feature.

[0063] The present invention provides a processing system for lung cancer bone scan images, including:

[0064] An acquisition module for acquiring lung cancer bone scan images;

[0065] A processing module for constructing a lung cancer bone imaging data model by using an encoder, a decoder, and a multi-scale boundary guidance module, and inputting the lung cancer bone scan image into the lung cancer bone imaging data model for processing to obtain a processed image;

[0066] The processing module inputs the lung cancer bone scan image into the lung cancer bone imaging data model for processing, including:

[0067] In the encoder, using rectangular dilated convolution as the shape guidance strategy, extract the output feature map unique to the rectangular size for the lung cancer bone scan image, and capture the global context information in two axial directions of the output feature map through a bidirectional pooling mechanism to obtain the encoder output feature map;

[0068] In the multi-scale boundary guidance module, the lung cancer bone scan image is subjected to multi-scale extraction and Laplacian pyramid processing to obtain Laplacian pyramid feature maps, and the feature maps at each stage of the encoder, the Laplacian pyramid feature maps, and the decoding prediction features of the (i + 1)-th layer are used as inputs to extract lesion information and lesion boundary details;

[0069] In the decoder, the encoder output feature maps, lesion information, and lesion boundary details are reconstructed through convolution and upsampling to output the final image.

[0070] The present invention provides a computer device, including a memory and a processor. A program is stored in the memory. When the program is executed by the processor, the processor executes the steps of the above-mentioned method for processing lung cancer bone scan images.

[0071] The present invention provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for processing lung cancer bone scan images are implemented.

[0072] Compared with the prior art, the present invention has the following remarkable advantages:

[0073] The present invention proposes a technology combining a shape guidance strategy, a multi-scale boundary guidance module, and bidirectional pooling. The rectangular dilated convolution is used to extract output feature maps unique to the rectangular size for the lung cancer bone scan image, making up for the limitations of traditional convolution in processing non-square images. The lung cancer bone scan image is subjected to multi-scale extraction and Laplacian pyramid processing in the multi-scale boundary guidance module to obtain Laplacian pyramid feature maps, and based on this, lesion information and lesion boundary details are extracted to achieve specific processing of the edge region. The Laplacian method and multi-scale deep learning method are used to focus on the boundary features, enabling the model to generate clearer lesion boundary markings, preventing missed detection of lesion areas in the image, thereby improving the boundary blur problem and enhancing the detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is the chest region segmentation process in the embodiment of the present invention;

[0075] Figure 2 It is the model overview diagram in the embodiment of the present invention;

[0076] Figure 3 It is the shape guidance strategy diagram in the embodiment of the present invention;

[0077] Figure 4 It is the multi-scale boundary guidance module diagram in the embodiment of the present invention;

[0078] Figure 5 It is the bidirectional pooling module diagram in the embodiment of the present invention;

[0079] Figure 6 This is the overall flowchart of a method for processing lung cancer bone scan images according to the present invention;

[0080] Figure 7 This is the flowchart for processing lung cancer bone scan images by inputting them into the lung cancer bone imaging data model according to the present invention. Specific embodiments

[0081] The following combines the accompanying drawings in the present invention to clearly and completely describe the technical solutions of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all 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 scope of protection of the present invention.

[0082] Although CAD-based medical image processing technology is of great significance in improving diagnostic efficiency and accuracy, it still faces many challenges in practical applications. These problems mainly focus on data acquisition, processing, and model performance. First, data quality and consistency issues have a significant impact on the performance of CAD systems. Medical images are often limited by noise, artifacts, and acquisition conditions, making it difficult to guarantee image quality; at the same time, differences between different devices, acquisition parameters, and processing flows result in inconsistent data distributions, directly affecting the generalization ability and cross-domain adaptability of the system. Then, the data itself has sparsity and class imbalance problems. The lesion area usually accounts for a very small proportion of the image, while the background area occupies most of it. This imbalance in data distribution causes the model to easily ignore lesion features during training, thereby reducing the detection ability for key regions.

[0083] As Figure 6 and Figure 7 shown, the following describes a method for processing lung cancer bone scan images provided by the present invention, which specifically includes the following steps:

[0084] Step S1: Obtain lung cancer bone scan images.

[0085] For SPECT whole-body bone imaging data, the data preprocessing steps specifically include: data cleaning, data denoising, blank area excision, and personalized bladder area excision. See Appendix Figure 1 .

[0086] Personalized bladder area excision:

[0087] Non-lesion hot spots (such as injection sites and the bladder) can significantly affect the accurate diagnosis of true lesions. Compared with non-lesion hot spots, the concentration of radiopharmaceuticals in the lesion area is much lower. The excessive radiation level of non-lesion hot spots can lead to the phenomenon of "the large number masking the small number", making it difficult to identify lesions, resulting in missed diagnoses. To solve this problem, non-lesion hot spots were automatically removed in the study, and the processed images were incorporated into the experimental dataset. The personalized resection process for the bladder area is shown in the appendix Figure 1 。

[0088] Data augmentation: data rotation, data translation

[0089] Data rotation means that the image will be randomly rotated by a certain angle to the left or right around its geometric center.

[0090] Data translation means that the image will be randomly translated by +t or -t pixels in the horizontal or vertical direction.

[0091] Step S2: Construct a lung cancer bone imaging data model using an encoder, a decoder, and a multi-scale boundary guidance module, and input the lung cancer bone scan image into the lung cancer bone imaging data model for processing to obtain the processed image.

[0092] Taking the SPECT thoracic bone imaging view of systemic lung cancer bone metastases as the research object, starting from the aspects of shape features, image itself information, and the difficulty of extracting context information, three strategies are combined: a shape guidance module, a multi-scale boundary guidance module, and a bidirectional pooling mechanism to solve the problem. The specific process is as follows:

[0093] Input the lung cancer bone scan image into the lung cancer bone imaging data model for processing, including:

[0094] In the encoder, using rectangular dilated convolution as the shape guidance strategy, extract the output feature map unique to the rectangular size for the lung cancer bone scan image, and capture the global context information in the two axial directions of the output feature map through the bidirectional pooling mechanism to obtain the encoder output feature map.

[0095] Set a multi-scale boundary guidance module between the encoder and the decoder to perform multi-scale extraction and Laplacian pyramid processing on the lung cancer bone scan image to obtain the Laplacian pyramid feature map, and use the feature map of each stage of the encoder, the Laplacian pyramid feature map, and the decoding prediction feature of the i + 1 layer as inputs to extract lesion information and lesion boundary details.

[0096] In the decoder, reconstruct the encoder output feature map, lesion information, and lesion boundary details through convolution and upsampling, and output the final image.

[0097] Specifically:

[0098] (1) In the encoder stage, ResNet34 is used as the backbone to extract features from lung cancer bone scan images, which includes five convolutional blocks for capturing and encoding input features. The Shape Guided strategy (SG) guides the encoder to better learn the features unique to the rectangular dimensions of the image. In addition, through the Dual Pooling mechanism (DP), it captures global context information in two axes of the image, significantly enhancing the feature representation ability and the depth of context information extraction. (2) In the decoder stage, the model reconstructs the output image from the encoded features through basic convolution and upsampling operations. (3) To address the challenges of information loss during downsampling and improve clarity, we introduce a Multi-scale Boundary Guidance (MBG) strategy between the encoder and the decoder. This strategy enhances the model's ability to extract lesion information and lesion boundary details of various sizes, and alleviates the problems related to low clarity and indistinct lesion boundaries, in order to achieve automatic diagnosis of whole-body bone imaging.

[0099] The SG module, as Figure 3 shown, performs 1×1 convolution and 7×1 convolution on the lung cancer bone scan image to obtain a feature cube The specific expression is:

[0100] U * =Conv7×1(Conv1×1(u),dilation=o * );

[0101] where o * is the dilation rate.

[0102] The feature cube obtained after two convolutions is processed as follows: through the F sq (·) operation, the feature cube is compressed into a 1×1×C channel description vector where the corresponding channel description element z c ∈Z has the specific expression:

[0103]

[0104] where z c represents the output value after global average pooling of channel c; F sq (·) represents the function of the Squeeze operation for compressing the spatial information of the feature map; u c represents the two-dimensional feature map of the input feature map on channel c; H represents the height of the input feature map; W represents the width of the input feature map; m represents the index of the row in the feature map; n represents the index of the column in the feature map; u c(m,n) represents the eigenvalue of channel c at the m-th row and n-th column. The channel description vector is input into the excitation function F ex (·,W), and the activated channel description vector is obtained to capture channel dependencies. The specific expression of the activated channel description vector is defined as follows:

[0105] s c =F ex (z c ,W)=σ(W2δ(W1z c ));

[0106] where δ is the Sigmoid(·) function, σ is the ReLU(·) equation, r is the squeeze ratio; among them, the feature cube U is adjusted by the activated channel description vector * , and the adjusted output feature map v is obtained; the specific expression of the c-th output feature map v c is as follows:

[0107] v c =F scale (u c ,s c )=s c u c ;

[0108] The main path selects an output feature map v with a fineness higher than the threshold from the output feature map v f , where the specific expression of the output feature map v f is as follows:

[0109]

[0110] Among them, represents element-wise multiplication, and f s (v) is a spatial function, defined as shown in the formula:

[0111] f s (v)=σ(f k×k ([AvgPool(v);MaxPool(v)]));

[0112] where δ is the Sigmoid(·) function, AvgPool(·) is the average pooling operation, MaxPool(·) is the maximum pooling operation, and f k×k is a convolution with a kernel size of k×k; finally, the output feature map v f with a fineness higher than the threshold is processed by a 1×1 convolutional layer with 2C channels, and then by a 1×1 convolutional layer with C channels to obtain the final output feature map, denoted as:

[0113] vf = f 1×1，c (f 1×1，2c (v f ));

[0114] The multi-scale boundary guidance module in the MBG module consists of three different inputs, multi-scale, and attention mechanisms, and is specifically designed to emphasize lesion boundary information. As Figure 4 shown.

[0115] The first input: The first input is the feature map f i e from each encoder stage, which undergoes a 3×3 convolution operation to reduce the number of channels and obtain the output except for the first module.

[0116] The second input: The second input is to first perform multi-scale extraction on the input feature map (lung cancer bone scan image) to obtain the output feature cube I, and the specific expression is:

[0117] I = H × W × (C1 + C2 + C3 + C P );

[0118] where C′ = C1 + C2 + C3 + C P is the concatenation result of the outputs of the four branches of the multi-scale learning block, and C1, C2, and C P represent the number of channels in the feature maps corresponding to the convolution kernels of sizes 1×1, 3×3, and 5×5 respectively. Each branch is equipped with a different convolution kernel size to handle lesions of different sizes, and C P also represents the number of channels in the feature map obtained by max pooling.

[0119] The output feature cube is processed through a Laplacian pyramid to obtain the Laplacian pyramid feature map I k , and the specific expression is:

[0120]

[0121] where I is the lung cancer bone scan image, g is the convolution operator with a Gaussian filter, d is the 2-fold downsampling operation, and k represents the k-th layer L k of the Laplacian pyramid, and the k-th layer L k of the Laplacian pyramid is obtained by subtracting the upsampled version u of the next top layer I k from the current layer L k+1 , and the specific expression is:

[0122] L k = I k - u(I k+1 ), f l = L1(I);

[0123] Among them, f l As the third input, that is, as the high-frequency feature, L1(I) represents the feature of the first layer of the Laplacian pyramid.

[0124] Integration and attention mechanism: At the i-th layer, the EGA module integrates three components: the encoded feature f i e , the decoded feature from the (i + 1)-th layer and the high-frequency feature f l . Generally, the high-frequency feature of the i-th layer of the Laplacian pyramid is obtained by performing Gaussian filtering on the (i - 1)-th layer and then downsampling the result of the Gaussian filtering by a factor of 2 to obtain the high-frequency feature f i l :

[0125]

[0126] Among them, d is downsampling by a factor of 2, (d(f l )) i is downsampling by a factor of 2 for i times, that is, d(d(...d(f l ))), is the initial feature.

[0127] Generate the reverse attention map and the boundary attention map from the decoded prediction feature of the (i + 1)-th layer The calculation expression of the reverse attention map is:

[0128]

[0129] The boundary attention map is obtained by applying the Laplacian operator, and the specific calculation expression is:

[0130]

[0131] Multiply the high-frequency feature f i l , the boundary attention map and the decoded feature map with the encoded feature f i e respectively for convolution to obtain the convolved feature map f i m , and the specific expression is:

[0132]

[0133] Process the encoded feature f i e and the convolved feature map using the attention mask to obtain the optimized feature result f ia , the specific expression is:

[0134]

[0135] Among them, σ represents the Sigmoid function.

[0136] The optimized feature results are input into the global context block GC for processing to obtain the feature f of the lesion information and the lesion boundary details i d , the specific expression is:

[0137]

[0138] Among them, GC(.) represents the global context block.

[0139] The third input: At the i-th layer of the boundary guidance module, the third input is the decoded prediction feature from the (i + 1)-th layer, denoted as This decoded prediction feature is derived from the decoded feature and is generated by the decoder at the (i + 1)-th layer.

[0140] DP module: As Figure 5 shown, the bidirectional pooling module aims to extract context information through horizontal pooling and vertical pooling. The output feature map x first undergoes horizontal pooling and vertical pooling to extract information in two directions of the output feature map. Then, the information obtained from the two poolings is combined to generate a preliminary rectangular region of interest, and the region shape of the preliminary rectangular region of interest is adjusted through a shape self-calibration function to obtain the adjusted rectangular region of interest y, that is, the region shape is adjusted through the shape self-calibration function to make it closer to the foreground object. The specific expression is:

[0141]

[0142] Among them, the symbol δ represents the Sigmoid function, h represents horizontal pooling, and z represents vertical pooling.

[0143] The output feature map x is further used to extract the local feature details of the output feature map through a 3×3 depth convolution, and the local feature is multiplied by y pixel by pixel to obtain the global feature F(x). The specific expression is:

[0144] F(x) = conv 3×3 (x) ☉ y(x);

[0145] The global feature is further optimized through batch normalization and a multi-layer perceptron to generate the encoder output feature map The specific expression is:

[0146] output = F(x) + MLP(BN(F(x)));

[0147] Among them, BN represents batch normalization, and MLP represents a multi-layer perceptron. Here, output is the output feature map of the encoder

[0148] Based on the above method, the present invention provides a processing system for lung cancer bone scan images, including: an acquisition module and a processing module.

[0149] Among them, the acquisition module is used to obtain lung cancer bone scan images; the processing module is used to construct a lung cancer bone imaging data model using an encoder-decoder, and input the lung cancer bone scan images into the lung cancer bone imaging data model for processing to obtain the processed images; the processing module inputs the lung cancer bone scan images into the lung cancer bone imaging data model for processing, including: in the encoder, using rectangular dilated convolution as the shape guidance strategy to extract the output feature map unique to the rectangular size for the lung cancer bone scan images, and capturing global context information in two axial directions of the output feature map through a bidirectional pooling mechanism to obtain the encoder output feature map; setting a multi-scale boundary guidance module between the encoder and the decoder to perform multi-scale extraction and Laplacian pyramid processing on the lung cancer bone scan images to obtain Laplacian pyramid feature maps, and using the feature maps of each stage of the encoder, the Laplacian pyramid feature maps, and the decoding prediction features of the (i + 1)-th layer as inputs to extract lesion information and lesion boundary details; in the decoder, reconstructing the encoder output feature map, lesion information, and lesion boundary details through convolution and upsampling to output the final image.

[0150] The present invention also provides a computer device, including a memory and a processor. When a program stored in the memory is executed by the processor, the processor executes the steps of a method for processing lung cancer bone scan images.

[0151] According to the disclosed embodiments, the computer device can communicate with one or more external devices (such as a keyboard, a pointing device, Bluetooth communication, etc.), or communicate with any device (such as a router, a demodulator, etc.) that enables the computing device to communicate with one or more other computing devices.

[0152] The present invention also provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for processing lung cancer bone scan images are implemented.

[0153] According to the disclosed embodiments, the storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present invention, the storage medium may be any tangible medium that contains or stores a program, and the program can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0154] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A processing method for lung cancer bone scan images, characterized in that, Including: Obtain a lung cancer bone scan image; Construct a lung cancer bone imaging data model using an encoder, a decoder, and a multi-scale boundary guidance module, and input the lung cancer bone scan image into the lung cancer bone imaging data model for processing to obtain a processed image; Input the lung cancer bone scan image into the lung cancer bone imaging data model for processing, including: In the encoder, using rectangular dilated convolution as a shape guidance strategy, extract an output feature map unique to the rectangular size for the lung cancer bone scan image, and capture global context information in two axial directions of the output feature map through a bidirectional pooling mechanism to obtain an encoder output feature map; In the multi-scale boundary guidance module, perform multi-scale extraction and Laplacian pyramid processing on the lung cancer bone scan image to obtain Laplacian pyramid feature maps, and use the feature maps of each stage of the encoder, the Laplacian pyramid feature maps, and the decoding prediction features of the (i + 1)-th layer as inputs to extract lesion information and lesion boundary details; In the decoder, reconstruct the encoder output feature map, lesion information, and lesion boundary details through convolution and upsampling, and output the final image.

2. The processing method of a lung cancer bone scan image according to claim 1, wherein Using rectangular dilated convolution as a shape guidance strategy to extract an output feature map unique to the rectangular size for the lung cancer bone scan image, including: For lung cancer bone scan images Perform 1×1 convolution processing and 7×1 convolution processing to obtain a feature cube The specific expression is as follows: U * = Conv7×1(Conv1×1(u), dilation = o * ); where, o * is the expansion ratio; Through F sq (·) operation compresses the feature cube into a 1×1×C channel description vector Among them, the specific expression of the corresponding channel description element z c ∈Z is: Among them, z c represents the output value of channel c after global average pooling; F sq (·) represents the function of the Squeeze operation, which is used to compress the spatial information of the feature map; u c represents the two-dimensional feature map of the input feature map on channel c; H represents the height of the input feature map; W represents the width of the input feature map; m represents the index of the row in the feature map; n represents the index of the column in the feature map; u c (m, n) represents the feature value of channel c at the m-th row and the n-th column; the channel description vector is input into the excitation function F ex (·, W), and the activated channel description vector is obtained. The specific expression of the form of the activated channel description vector is defined as follows: s c = F ex (z c , W) = σ(W2δ*(W1z c )); where δ is the Sigmoid(·) function, σ is the ReLU(·) equation, r is the extrusion ratio; Adjust the feature cube U through the activated channel description vector * , and obtain the adjusted output feature map v; among them, the specific expression of the c-th output feature map v c is as follows: v c = F scale (u c , s c ) = s c u c ; Select an output feature map \(v\) from the output feature maps with a fineness higher than the threshold f , where the output feature map \(v\) f has the specific expression as follows: Among them, represents element-wise multiplication, and f s (v) is a spatial function defined as shown in the formula: f s (v) = σ(f k×k ([AvgPool(v); MaxPool(v)])); where δ is the Sigmoid(·) function, AvgPool(·) is the average pooling operation, MaxPool(·) is the max pooling operation, and f k×k is a convolution with a kernel size of k×k; The output feature map v with a fineness higher than the threshold f is processed through a 1×1 convolutional layer with 2C channels and then through a 1×1 convolutional layer with C channels to obtain the final output feature map, expressed as: v f = f 1×1，c (f 1×1，2c (v f ))。 3. The processing method of a lung cancer bone scan image according to claim 1, characterized in that, The process of capturing global context information in two axial directions of the output feature map through a bidirectional pooling mechanism to obtain an encoder output feature map, including: Extract information in two directions of the output feature map through horizontal pooling and vertical pooling; Merge the information obtained from the two poolings to generate a preliminary rectangular region of interest, and adjust the region shape of the preliminary rectangular region of interest through a shape self-calibration function to obtain an adjusted rectangular region of interest y, and the specific expression is: where the symbol δ represents the Sigmoid function, h represents horizontal pooling, and z represents vertical pooling; Further extract the local features of the output feature map by passing the output feature map x through a 3×3 depth convolution, and multiply the local features with y pixel by pixel to obtain the global feature F(x), and the specific expression is: F(x) = conv 3×3 (x) ⊙ y(x); Further optimize the global features through batch normalization and multi-layer perceptron to generate the encoder output feature map The specific expression is as follows: output = F(x) + MLP(BN(F(x))); Among them, BN represents batch normalization, MLP represents multi-layer perceptron, and output is the output feature map of the encoder 4. The processing method of a lung cancer bone scan image according to claim 3, characterized in that, The process of performing multi-scale extraction and Laplacian pyramid processing on the lung cancer bone scan image to obtain Laplacian pyramid feature maps, including: Perform multi-scale extraction on the lung cancer bone scan image to obtain an output feature cube I, and the specific expression is: I = H × W × (C1 + C2 + C3 + C P ); Among them, C′ = C1 + C2 + C3 + C P is the output splicing result of the four branches of the multi-scale learning block, and C1, C2, and C P respectively represent the number of channels in the feature maps corresponding to the convolution kernels with sizes of 1×1, 3×3, and 5×5; The output feature cube is processed through a Laplacian pyramid to obtain the Laplacian pyramid feature map I k , and the specific expression is as follows: Where, I is the lung cancer bone scan image, g is the convolution operator with a Gaussian filter, d is the 2-fold downsampling operation, and k represents the k-th layer L of the Laplacian pyramid k , the k-th layer L of the Laplacian pyramid k is obtained by subtracting the upsampled version u of the next top layer I k from the current layer L k+1 , and the specific expression is: L k = I k - u(I k+1 ), f l = L1(I); Among them, f l As a high-frequency feature, L1(I) represents the first-layer feature of the Laplacian pyramid.

5. The processing method of a lung cancer bone scan image according to claim 4, wherein, The process of extracting lesion information and lesion boundary details, including: Generate the decoding prediction features of the (i + 1)-th layer to generate the reverse attention map and the boundary attention map and multiply the high-frequency feature f i l , the boundary attention map, and the decoded feature map with the encoded feature f i e respectively for convolution to obtain the convolved feature map f i m , and the specific expression is as follows: Process the encoded feature f using an attention mask i e and the feature map after convolution to obtain the optimized feature result f i a , and the specific expression is: where σ represents the Sigmoid function; Input the optimized feature results into the global context block GC for processing to obtain the feature f of lesion information and lesion boundary details i d , and the specific expression is:

6. The processing method of a lung cancer bone scan image according to claim 5, wherein, The acquisition process of the high-frequency feature f i l is specifically as follows: By performing Gaussian filtering on the (i - 1)-th layer and downsampling the result of the Gaussian filtering by a factor of 2, the high-frequency feature f is obtained. i l , and the specific expression is as follows: Among them, where f l = L1(I), is the initial feature, d is 2x downsampling, (d(f l )) i is the i-th 2x downsampling.

7. A processing system for lung cancer bone scan images, characterized in that, Including: An acquisition module for obtaining a lung cancer bone scan image; A processing module for constructing a lung cancer bone imaging data model using an encoder, a decoder, and a multi-scale boundary guidance module, and inputting the lung cancer bone scan image into the lung cancer bone imaging data model for processing to obtain a processed image; The processing module inputs the lung cancer bone scan image into the lung cancer bone imaging data model for processing, including: In the encoder, using rectangular dilated convolution as a shape guidance strategy, extract an output feature map unique to the rectangular size for the lung cancer bone scan image, and capture global context information in two axial directions of the output feature map through a bidirectional pooling mechanism to obtain an encoder output feature map; In the multi-scale boundary guidance module, the lung cancer bone scan image is subjected to multi-scale extraction and Laplacian pyramid processing to obtain Laplacian pyramid feature maps, and the feature maps of each stage of the encoder, the Laplacian pyramid feature maps, and the decoding prediction features of the (i + 1)-th layer are used as inputs to extract lesion information and lesion boundary details; In the decoder, the encoder output feature maps, lesion information, and lesion boundary details are reconstructed through convolution and upsampling to output the final image.

8. A computer device, characterized in that, It includes a memory and a processor. A program is stored in the memory. When the program is executed by the processor, the processor is caused to execute the steps of the method for processing a lung cancer bone scan image according to any one of claims 1 to 6.

9. A storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by the processor, the steps of the method for processing a lung cancer bone scan image according to any one of claims 1 to 6 are implemented.

Images