Intelligent neurosurgery image processing system

Abstract

The invention discloses an intelligent neurosurgery image processing system, which belongs to the technical field of image processing and comprises a neurosurgery image acquisition module, a neurosurgery image denoising module, a neurosurgery image enhancement module and a neurosurgery image processing module. According to the method, focus masks are obtained through self-adaptive thresholds, the number of decomposition layers and the number of directions are dynamically set, low-frequency approximate components are filtered based on dynamically-adjusted guide filtering parameters, soft threshold processing is executed on high-frequency detail components based on self-adaptive partition thresholds, and the neurosurgery image processing effect is effectively improved. A five-layer coding and decoding structure is adopted, each coding layer comprises a region self-adaptive convolution block and a cross-region context aggregation layer, each decoding layer comprises a gating deconvolution block and a content self-adaptive attention layer, an enhanced neurosurgical image is obtained, multi-loss joint optimization is performed, and the accuracy and clinical reliability of neurosurgical image processing are improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically referring to an intelligent neurosurgical image processing system. Background Technology

[0002] Neurosurgical image processing systems utilize advanced image processing and deep learning technologies to precisely process multimodal neurosurgical images, including denoising and enhancement, highlighting lesion features while preserving normal tissue details. This provides high-quality image data for clinical diagnosis, lesion localization, and treatment planning. However, existing neurosurgical image processing systems suffer from weak grayscale differences and complex structures between lesions and normal tissues in neurosurgical images, as well as uneven noise distribution. Traditional denoising methods easily lead to missed detection of small lesions, blurred lesion edges, and incomplete noise reduction of normal tissues, resulting in poor neurosurgical image processing performance. Furthermore, existing neurosurgical image processing systems often fail to highlight lesion features and ignore contextual relationships between regions. Traditional enhancement methods are prone to insufficient enhancement of lesions or excessive enhancement of normal tissues, leading to inaccurate neurosurgical image processing. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an intelligent neurosurgical image processing system. Addressing the problems of weak grayscale differences and complex structures between lesions and normal tissues in existing neurosurgical image processing systems, as well as uneven noise distribution, traditional denoising methods are prone to missing small lesions, blurring lesion edges, and incomplete noise reduction of normal tissues, resulting in poor neurosurgical image processing performance. This solution extracts a comprehensive feature map, obtains a lesion mask through adaptive thresholding, and divides the lesion region and normal tissue region. Based on the lesion mask and local structural orientation entropy, the number of decomposition layers and orientations is dynamically set to obtain low-frequency approximation components and high-frequency detail components. Low-frequency approximation components are filtered based on dynamically adjusted guided filtering parameters, and high-frequency detail components undergo soft thresholding based on adaptive partitioning thresholds. The resulting denoised neurosurgical image significantly suppresses noise while preserving the original image. This approach preserves lesion details and normal tissue structures, providing high-quality image data and effectively improving neurosurgical image processing. Addressing the issues of insufficient lesion feature representation and neglect of inter-regional contextual relationships in existing neurosurgical image processing systems, traditional enhancement methods often result in insignificant lesion enhancement or excessive enhancement of normal tissue, leading to inaccurate neurosurgical image processing, this solution employs a five-layer encoding and decoding structure. Each encoding layer includes a region-adaptive convolutional block and a cross-regional contextual aggregation layer, while each decoding layer includes a gated deconvolutional block and a content-adaptive attention layer. The resulting enhanced neurosurgical image is obtained by weighted summation of pixel-level consistency loss, lesion region detail enhancement loss, and encoding-decoding feature consistency loss. This multi-loss synergistic constraint ensures that the enhanced image remains consistent with the original image while simultaneously enhancing lesion details, thus improving the overall accuracy and clinical reliability of neurosurgical image processing.

[0004] The present invention provides an intelligent neurosurgical image processing system, comprising a neurosurgical image acquisition module, a neurosurgical image denoising module, a neurosurgical image enhancement module, and a neurosurgical image processing module;

[0005] The neurosurgical image acquisition module acquires historical neurosurgical images and performs preprocessing, and then sends the data to the neurosurgical image denoising module;

[0006] The neurosurgical image denoising module receives data sent by the neurosurgical image acquisition module, extracts a comprehensive feature map, obtains a lesion mask through an adaptive threshold, divides the lesion area and normal tissue area, dynamically sets the number of decomposition layers and directions based on the lesion mask and local structural orientation entropy, obtains low-frequency approximation components and high-frequency detail components, filters the low-frequency approximation components based on dynamically adjusted guided filtering parameters, performs soft thresholding on the high-frequency detail components based on adaptive partitioning thresholds, reconstructs a denoised neurosurgical image, and sends the data to the neurosurgical image enhancement module.

[0007] The neurosurgical image enhancement module receives data sent by the neurosurgical image denoising module and adopts a five-layer encoding and decoding structure. Each encoding layer contains a region-adaptive convolutional block and a cross-regional context aggregation layer. Each decoding layer contains a gated deconvolutional block and a content-adaptive attention layer to obtain an enhanced neurosurgical image. The total loss is obtained based on pixel-level consistency loss, lesion region detail enhancement loss and encoding-decoding feature consistency loss, and the data is sent to the neurosurgical image processing module.

[0008] The neurosurgical image processing module receives data sent by the neurosurgical image enhancement module, and obtains the processing results of real-time neurosurgical images through the neurosurgical image denoising module and the neurosurgical image enhancement module.

[0009] Furthermore, the neurosurgical image denoising module includes a multi-feature fusion lesion pre-detection unit, a structure-guided decomposition unit, a low-frequency guided filtering unit, a high-frequency adaptive thresholding unit, and a denoised image reconstruction unit, specifically comprising the following:

[0010] Multi-feature fusion lesion pre-detection unit: Extract grayscale features, texture features and contrast features from neurosurgical images, normalize the three features respectively, and then fuse them by weight to obtain a comprehensive feature map. The adaptive threshold is calculated by the maximum inter-class variance method to generate candidate lesion region labels. Morphological operations are performed on the candidate lesion region labels to generate lesion masks. The comprehensive feature map is divided into lesion regions and normal tissue regions.

[0011] Structure-guided decomposition unit: For each pixel in a neurosurgical image, the proportion of gradient magnitude in each direction within its local window to the total gradient magnitude is calculated to obtain the local structural orientation entropy of the pixel. The number of decomposition layers is dynamically set according to the lesion mask, and the number of decomposition directions is dynamically set according to the local structural orientation entropy. Non-subsampled shear wave transform is performed on the neurosurgical image according to the number of decomposition layers and the number of decomposition directions to obtain low-frequency approximate components and high-frequency detail components.

[0012] Low-frequency guided filtering unit: For low-frequency approximate components, the guided filtering parameters are dynamically adjusted by combining lesion mask and local structural orientation entropy. The low-frequency approximate component itself is used as the guided image. The local mean and local variance are calculated in the filtering window centered on each pixel. The linear coefficients are solved to obtain the filtered low-frequency approximate optimized component.

[0013] High-frequency adaptive thresholding unit: Based on the basic noise standard deviation and global mean of high-frequency detail components, combined with the deviation of each pixel gray value from the mean, the skewness and kurtosis of the noise distribution are calculated respectively to obtain the noise statistical correction factor. The adaptive partitioning threshold of each pixel is determined according to the lesion mask, and soft thresholding is performed on the high-frequency detail components to obtain the high-frequency detail optimized components.

[0014] The denoising image reconstruction unit performs non-subsampled shear wave inverse transform on the low-frequency approximation optimization component and the high-frequency detail optimization component to reconstruct a denoised neurosurgical image.

[0015] Furthermore, the neurosurgical image enhancement module includes a region context coding unit, a gated attention decoding unit, and a multi-loss joint optimization unit, specifically comprising the following:

[0016] The region context encoding unit concatenates the denoised neurosurgical image with the lesion mask along the channel dimension to form the initial input feature map. It adopts a five-layer encoding structure, with each layer containing a region adaptive convolutional block and a cross-region context aggregation layer. The region adaptive convolutional block applies different depthwise separable convolutional weights to the lesion region and the normal tissue region according to the lesion mask to achieve region adaptive feature extraction. The cross-region context aggregation layer obtains the global context vector through global pooling operation, reconstructs the region adaptive features, and obtains a condensed context feature map.

[0017] The gated attention decoding unit concatenates the condensed context feature map output from the last layer of the encoder with the lesion mask along the channel dimension to form a mask-enhanced feature map. A five-layer decoding structure is adopted, with each layer containing a gated deconvolution block and a content-adaptive attention layer. The gated deconvolution block upsamples the mask-enhanced feature map through transposed convolution and performs skip connections with the corresponding layer features of the encoder, using a gating mechanism to obtain fused features. The content-adaptive attention layer uses the lesion mask as a prior and performs spatial and channel modulation on the fused features to obtain an optimized enhanced feature map. A 1×1 convolution and sigmoid activation are performed on the feature map output from the last layer of the decoder to generate a preliminary enhanced image, which is then restored to the original grayscale range through linear mapping to obtain an enhanced neurosurgical image.

[0018] The multi-loss joint optimization unit calculates the total loss by weighting and summing the pixel-level consistency loss, lesion region detail enhancement loss, and encoder-decoder feature consistency loss. The Adam optimizer is used to calculate the gradient of the total loss with respect to all trainable parameters in the region context encoding unit and gated attention decoding unit in each iteration. The parameters are updated by gradient descent to minimize the total loss, and the iteration continues until the loss converges.

[0019] Furthermore, the neurosurgical image processing module acquires real-time neurosurgical images, preprocesses them, and then sequentially inputs them into the neurosurgical image denoising module and the neurosurgical image enhancement module to obtain the processing results of the real-time neurosurgical images.

[0020] The beneficial effects achieved by the present invention using the above solution are as follows:

[0021] (1) In view of the problems in the existing neurosurgical image processing system, such as weak gray-level difference and complex structure between lesions and normal tissues in neurosurgical images, and uneven noise distribution, traditional denoising methods are prone to missing small lesions, blurring of lesion edges and incomplete denoising of normal tissues, resulting in poor neurosurgical image processing effect. This scheme extracts comprehensive feature map, obtains lesion mask through adaptive threshold, divides lesion area and normal tissue area, and enhances the distinction between lesion and normal tissue. According to the lesion mask and local structural direction entropy, the number of decomposition layers and directions are dynamically set to obtain low-frequency approximate components and high-frequency detail components, ensuring that the decomposed components accurately reflect the image features. The low-frequency approximate components are filtered based on dynamically adjusted guided filtering parameters to achieve differential denoising of low-frequency approximate components. The high-frequency detail components are processed by soft thresholding based on adaptive partition threshold to accurately screen effective details of lesion edges, suppress background noise, and enhance lesion contour features. The denoised neurosurgical image is reconstructed. The reconstructed image significantly suppresses noise while completely preserving lesion details and normal tissue structure, providing high-quality image data and effectively improving the neurosurgical image processing effect.

[0022] (2) To address the problems in existing neurosurgical image processing systems, such as the lack of prominent expression of lesion features in neurosurgical images and the neglect of contextual relationships between regions, traditional enhancement methods are prone to insignificant enhancement of lesions or excessive enhancement of normal tissues, leading to inaccurate neurosurgical image processing, this scheme adopts a five-layer encoding and decoding structure. Each encoding layer contains a region-adaptive convolutional block and a cross-regional context aggregation layer, which assigns differentiated weights to lesions and normal tissues, strengthens lesion features, and integrates global context to provide high-quality encoded features for accurate enhancement. Each decoding layer contains a gated deconvolutional block and a content-adaptive attention layer to obtain enhanced neurosurgical images. Combined with skip connections to preserve details, it accurately focuses on lesions, achieving significant enhancement of lesion regions and moderate preservation of normal tissues. The pixel-level consistency loss, lesion region detail enhancement loss, and encoding-decoding feature consistency loss are weighted and summed to obtain the total loss. The model parameters are iteratively optimized through the Adam optimizer. The multi-loss collaborative constraint ensures that the enhanced image is consistent with the original image, while enhancing the lesion detail enhancement effect, thus improving the overall accuracy and clinical reliability of neurosurgical image processing. Attached Figure Description

[0023] Figure 1 A schematic diagram of an intelligent neurosurgical image processing system provided by the present invention;

[0024] Figure 2 A schematic diagram of a neurosurgical image denoising module;

[0025] Figure 3 This is a schematic diagram of a neurosurgical image enhancement module.

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] Example 1, see Figure 1 The present invention provides an intelligent neurosurgical image processing system, comprising a neurosurgical image acquisition module, a neurosurgical image denoising module, a neurosurgical image enhancement module, and a neurosurgical image processing module;

[0030] The neurosurgical image acquisition module acquires historical neurosurgical images and performs preprocessing, and then sends the data to the neurosurgical image denoising module;

[0031] The neurosurgical image denoising module receives data sent by the neurosurgical image acquisition module, extracts a comprehensive feature map, obtains a lesion mask through an adaptive threshold, divides the lesion area and normal tissue area, dynamically sets the number of decomposition layers and directions based on the lesion mask and local structural orientation entropy, obtains low-frequency approximation components and high-frequency detail components, filters the low-frequency approximation components based on dynamically adjusted guided filtering parameters, performs soft thresholding on the high-frequency detail components based on adaptive partitioning thresholds, reconstructs a denoised neurosurgical image, and sends the data to the neurosurgical image enhancement module.

[0032] The neurosurgical image enhancement module receives data sent by the neurosurgical image denoising module and adopts a five-layer encoding and decoding structure. Each encoding layer contains a region-adaptive convolutional block and a cross-regional context aggregation layer. Each decoding layer contains a gated deconvolutional block and a content-adaptive attention layer to obtain an enhanced neurosurgical image. The total loss is obtained based on pixel-level consistency loss, lesion region detail enhancement loss and encoding-decoding feature consistency loss, and the data is sent to the neurosurgical image processing module.

[0033] The neurosurgical image processing module receives data sent by the neurosurgical image enhancement module, and obtains the processing results of real-time neurosurgical images through the neurosurgical image denoising module and the neurosurgical image enhancement module.

[0034] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the neurosurgical image acquisition module, historical neurosurgical images are acquired and preprocessed. The neurosurgical images include CT images, MRI images, X-ray images, and cerebral angiography images. The preprocessing includes image format unification and geometric correction. The image format unification is to convert all modal neurosurgical images into NIfTI-1 format to ensure data compatibility. The geometric correction is to achieve spatial alignment of multimodal neurosurgical images based on the Elastix registration toolkit to eliminate geometric deviations between modalities.

[0035] Example 3, see Figure 1 and Figure 2 This embodiment, based on the above embodiment, includes a multi-feature fusion lesion pre-detection unit, a structure-guided decomposition unit, a low-frequency guided filtering unit, a high-frequency adaptive thresholding unit, and a denoised image reconstruction unit in the neurosurgical image denoising module, specifically comprising the following:

[0036] A multi-feature fusion lesion pre-detection unit is used. In neurosurgical images, the grayscale difference between lesions and normal tissue may be subtle. A single feature may easily miss small lesions or misjudge normal structures, making it difficult to accurately distinguish the target area from the background. Grayscale features, texture features, and contrast features are extracted from neurosurgical images. Grayscale features are directly expressed as pixel grayscale, texture features are measured using the entropy value of a 5×5 local window, and contrast features are calculated using the difference between the maximum and minimum values ​​of a 5×5 local window. After normalizing the three features to the [0, 1] interval, they are weighted and fused with weights of 0.4, 0.3, and 0.3 to obtain a comprehensive feature map. The maximum inter-class squared error is then used. An adaptive threshold is calculated using a difference method to generate candidate lesion region labels. Candidate lesion regions in the comprehensive feature map with gray values ​​greater than the adaptive threshold are labeled as 1, while those in other regions are labeled as 0. Morphological operations are performed on the candidate lesion region labels to generate lesion masks, dividing the comprehensive feature map into lesion regions and normal tissue regions. Regions with a lesion mask equal to 1 are lesion regions, and the rest are normal tissue regions. Multi-feature fusion compensates for the limitations of single features, and the adaptive threshold and morphological operations improve labeling accuracy, providing precise region guidance for subsequent differential processing and reducing unnecessary computation. The formulas used are as follows:

[0037] ;

[0038] ;

[0039] In the formula, T th It is an adaptive threshold, t is a candidate threshold value, and μ T It is the global average gray value of the comprehensive feature map. It is the lesion mask at pixel (i,j). It is the label value of the candidate lesion region at pixel (i,j). and These represent the percentage of pixels classified as normal tissue area and lesion area when the threshold is t, respectively. and These are the average gray values ​​of the normal tissue area and the lesion area when the threshold is t, respectively, and K. 3×3 and K 2×2 These are square structural elements of sizes 3×3 and 2×2, respectively. and These are the morphological closing operator and the morphological opening operator, respectively.

[0040] Structure-guided decomposition unit: Neurosurgical images have complex structures, and fixed decomposition parameters cannot simultaneously preserve lesion details and improve normal tissue processing efficiency, easily resulting in the loss of information about minute lesions or increased computational redundancy. The gradient direction range from 0° to 180° is uniformly divided into 8 directions, each spaced 22.5° apart. For each pixel in the neurosurgical image, the proportion of gradient magnitude in each direction within its 5×5 local window to the total gradient magnitude is calculated to obtain the local structural direction entropy of the pixel. The number of decomposition layers is dynamically set according to the lesion mask. Lesion regions with a lesion mask of 1 are decomposed into 5 layers to preserve details of minute lesions, while normal tissue regions with a lesion mask of 0 are decomposed into 4 layers. Layer decomposition is employed to strike a balance between computational efficiency and detail preservation. The number of decomposition directions is dynamically set based on the local structural direction entropy: 32 directions are used when the local structural direction entropy > 1.2, 16 directions are used when the local structural direction entropy ≤ 0.8 ≤ 1.2, and 8 directions are used when the local structural direction entropy < 0.8. Non-subsampled shear wave transform is performed on the neurosurgical images based on the number of decomposition layers and directions to obtain low-frequency approximate components and high-frequency detail components. More decomposition layers are used in lesion areas to preserve details, while the number of layers is optimized for normal tissue to balance efficiency. The number of directions is adaptively adjusted according to structural complexity to ensure targeted decomposition and lay the foundation for accurate noise reduction. The formulas used are as follows:

[0041] ;

[0042] ;

[0043] In the formula, It represents the proportion of the gradient magnitude at pixel (i,j) in the d-th direction to the total gradient magnitude of the local window, where d is the direction index. It is a 5×5 local window centered at pixel (i,j). and These are the gradient magnitude and gradient direction angle at pixel (x, y), respectively. It is the local structural orientation entropy at pixel (i,j). It is an indicator function, when When it belongs to the d-th direction, ,otherwise , It is a smoothing term. ;

[0044] Low-frequency guided filtering unit: In low-frequency components, lesion details are fragile, and noise distribution in normal tissue is uneven. Fixed filtering parameters can easily lead to blurred lesions or incomplete noise reduction of normal tissue. For low-frequency approximate components, the guided filtering parameters are dynamically adjusted based on the lesion mask and local structural orientation entropy to achieve differentiated smoothing. For lesion regions with a lesion mask of 1, the filtering window radius r=3 is set, and the regularization parameter... To preserve detailed structure; for normal tissue areas with a lesion mask of 0, set the filter window radius. Regularization parameters To smooth noise while preserving structure, the method uses the low-frequency approximation component itself as a guide image. Within a filtering window centered on each pixel, its local mean and local variance are calculated, and the linear coefficients are solved to obtain the optimized low-frequency approximation component where noise is suppressed and structure is preserved. Small windows and low regularization parameters in the lesion region ensure complete detail, while normal tissue parameters are dynamically optimized with structural entropy, achieving a precise balance between noise suppression and structure preservation. The formulas used are as follows:

[0045] ;

[0046] ;

[0047] ;

[0048] In the formula, and These are the linear coefficients within a window centered at pixel (i,j) with a radius of r. It is the regularization parameter at pixel (i,j). and These are the local mean and local variance within a filtering window centered at pixel (i,j) with radius r, respectively. It rounds down. It is the gray value of the low-frequency approximation component of the neurosurgical image at pixel (i,j). It is the gray value of the low-frequency approximation optimization component at pixel (i,j);

[0049] High-frequency adaptive thresholding unit: In high-frequency components, lesion details and noise signals are superimposed. Fixed thresholds are prone to misremoving lesion edge information or residual noise, and the noise distribution varies greatly in different regions. Based on the basic noise standard deviation and global mean of the high-frequency detail components, combined with the deviation of each pixel's grayscale value from the mean, the skewness and kurtosis of the noise distribution are calculated to obtain a noise statistical correction factor. The adaptive partition threshold of each pixel is determined according to the lesion mask. Soft thresholding is performed on the high-frequency detail components, retaining detail coefficients with absolute values ​​greater than the threshold and suppressing noise coefficients with absolute values ​​less than the threshold, resulting in optimized high-frequency detail components. The noise statistical correction factor adapts to complex noise distributions, and the partition threshold is differentiated for lesions and normal tissues, preserving lesion edge details while effectively suppressing background noise. The formulas used are as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] In the formula, α is the noise statistical correction factor. It is the standard deviation of the fundamental noise of the high-frequency detail components. , It is the global median of the high-frequency detail components, and S and C are the skewness and kurtosis of the noise distribution in the high-frequency detail components, respectively. , , and These are the gray values ​​of the high-frequency detail components of the neurosurgical image at pixels (i,j) and (x,y), respectively. and N H These are the global mean of the high-frequency detail components and the total number of pixels, respectively. It is the adaptive partitioning threshold at pixel (i,j). and These are the maximum value function and the variance function, respectively. It is the grayscale value of the high-frequency detail optimization component at pixel (i,j). It is a symbolic function;

[0054] The denoised image reconstruction unit performs non-subsampled shear wave inverse transform on the low-frequency approximation optimization component and the high-frequency detail optimization component to reconstruct a denoised neurosurgical image in which noise is initially suppressed while clinically critical structures are preserved.

[0055] By performing the above operations, this solution addresses the problems in existing neurosurgical image processing systems, such as weak gray-level differences and complex structures between lesions and normal tissues in neurosurgical images, uneven noise distribution, and the tendency of traditional denoising methods to miss small lesions, blur lesion edges, and incomplete noise reduction of normal tissues, resulting in poor neurosurgical image processing performance. This solution extracts a comprehensive feature map, obtains a lesion mask through adaptive thresholding, and divides lesion and normal tissue regions to enhance the distinction between lesions and normal tissues. Based on the lesion mask and local structural orientation entropy, the number of decomposition layers and directions are dynamically set to obtain low-frequency approximate components and high-frequency detail components, ensuring that the decomposed components accurately reflect image features. Low-frequency approximate components are filtered based on dynamically adjusted guided filtering parameters to achieve differentiated denoising of these components. Soft thresholding is performed on high-frequency detail components based on adaptive partitioning thresholds to accurately select effective details at lesion edges, suppress background noise, and enhance lesion contour features. The resulting denoised neurosurgical image significantly suppresses noise while fully preserving lesion details and normal tissue structures, providing high-quality image data and effectively improving the neurosurgical image processing performance.

[0056] Example 4, see Figure 1 and Figure 3This embodiment, based on the above embodiment, includes a region context coding unit, a gated attention decoding unit, and a multi-loss joint optimization unit in the neurosurgical image enhancement module, specifically comprising the following:

[0057] The system employs a region context encoding unit. Traditional encoding uses uniform convolutional weights for lesions and normal tissue, which fails to highlight lesion features and easily overlooks the contextual relationships between regions. The denoised neurosurgical image and lesion mask are concatenated along the channel dimension to form the initial input feature map. A five-layer encoding structure is used, with each layer containing a region-adaptive convolutional block and a cross-region context aggregation layer. The region-adaptive convolutional block applies different depthwise separable convolutional weights to the lesion region and normal tissue region according to the lesion mask, achieving region-adaptive feature extraction. The cross-region context aggregation layer obtains the global context vector through global pooling, reconstructs the region-adaptive features, and obtains a condensed context feature map. The region-adaptive weights enhance the expression of lesion features, and the context aggregation integrates global information, reducing the one-sidedness of local features and providing semantically rich feature support for decoding. The formulas used are as follows:

[0058] ;

[0059] ;

[0060] in, It is the region adaptive feature output by the region adaptive convolution block of layer l, where l is the layer index. It is a 3×3 convolution operation, where M is the lesion mask and W... lesion and W normal These are the depth-separable convolution weights corresponding to the lesion region and the normal tissue region, respectively. It is a depthwise separable convolution operation. It is Hadamaji. It is a batch of normalization, It is the Sigmoid activation function, and g is the global context description vector. , It is global average pooling, and T is the transpose operation. It is the Softmax function. It is layer normalization. and These are the condensed context feature maps output from layer l and layer (l-1), respectively.

[0061] The gated attention decoding unit addresses the issue that feature upsampling during decoding can easily lose details and struggle to focus on the lesion area, resulting in insignificant enhancement at the lesion site or excessive enhancement of normal tissue. The unit concatenates the condensed contextual feature map output from the encoder's last layer with the lesion mask along the channel dimension to form a mask-enhanced feature map. A five-layer decoding structure is employed, with each layer containing a gated deconvolution block and a content-adaptive attention layer. The gated deconvolution block upsamples the mask-enhanced feature map through transposed convolution and performs skip connections with the corresponding layer features from the encoder, utilizing a gating mechanism to obtain fused features. The content-adaptive attention layer uses the lesion mask as a priori to spatially and channel-modulate the fused features, resulting in an optimized enhanced feature map. A 1×1 convolution and sigmoid activation are applied to the output feature map of the decoder's last layer to generate a preliminary enhanced image, which is then restored to the original grayscale range through linear mapping to obtain an enhanced neurosurgical image. Skip connections supplement detailed information, and the gating mechanism and attention modulation precisely focus on the lesion, avoiding excessive enhancement of normal tissue and improving the visual recognition of the lesion area. The formulas used are as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula, It is the upsampling feature of the l-th layer decoder. It is a feature concatenation operation. It is a 1×1 convolution operation. It is the ReLU activation function. It is the fused feature of the output of the l-th gated deconvolution block. It is a clipping function. It is a fully connected layer. and These are the optimized and enhanced feature maps obtained from layers 1 and 5 after content-adaptive attention modulation, respectively. min and V max These are the minimum and maximum gray values ​​of the original denoised neurosurgical image, I. enh It is enhanced neurosurgical imaging, L max It is the total number of encoder layers. It is the Lth max -l+1 layer output condensed context feature map;

[0066] A multi-loss joint optimization unit is used. A single loss function can easily lead to large deviations between the enhanced image and the original image, insufficient enhancement of lesion details, or disconnection between encoding and decoding features, affecting clinical applicability. Based on the lesion mask, regional differential weights are set, with the weight of the lesion region being 1.2 and the weight of the normal tissue region being 0.8. The denoised neurosurgical image and the enhanced neurosurgical image are normalized to the [0, 1] interval, and the squared error is calculated pixel by pixel. Combined with the regional weights, the pixel-level consistency loss is obtained. The lesion regions in the denoised neurosurgical image and the enhanced neurosurgical image are filtered out by the lesion mask. The gray-level variance of the 5×5 local window of each pixel in the lesion region is calculated. Then, the mean of the local variances of all lesion pixels is taken to generate the respective lesions. The average local variance of the region is used to obtain the lesion region detail enhancement loss. L2 normalization is applied to the condensed context feature map output from the last layer of the encoder and the optimized enhanced feature map output from the last layer of the decoder, and the absolute error between them is calculated pixel-by-pixel and channel-by-channel to obtain the encoder-decoder feature consistency loss. The pixel-level consistency loss, lesion region detail enhancement loss, and encoder-decoder feature consistency loss are weighted and summed to obtain the total loss. The Adam optimizer is used, and in each iteration, the gradient of the total loss with respect to all trainable parameters in the region context encoding unit and the gated attention decoding unit is calculated. Parameters are updated using gradient descent to minimize the total loss, iterating until the loss converges (total loss fluctuation is less than 10 for 30 consecutive rounds). -5 This method achieves global optimization of neurosurgical image enhancement. Trainable parameters include the weights and biases of various convolutional and fully connected layers in the region context coding unit and gated attention decoding unit. Region-differentiated weights prioritize lesion optimization, while multi-loss synergy ensures image realism, lesion detail clarity, and feature consistency, thereby enhancing the clinical diagnostic value of the enhanced images. The formulas used are as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] In the formula, L pixel L lesion L feat and L total These represent pixel-level consistency loss, lesion region detail enhancement loss, encoder-decoder feature consistency loss, and total loss, respectively. Q, U, and Z are the number of height pixels, width pixels, and channels, respectively. q, u, and z are the height pixel index, width pixel index, and channel index, respectively. ω1 and ω0 are the weights of the lesion region and the normal tissue region, respectively. It is the lesion mask at pixel (q,u). and These are the grayscale values ​​at pixel (q,u) of the enhanced neurosurgical image and the denoised neurosurgical image, respectively. and These are the mean local variances of the lesion region in denoised neurosurgical images and enhanced neurosurgical images, respectively. and These are the condensed context feature map output from the last layer of the encoder and the optimized and enhanced feature map output from the last layer of the decoder, respectively, after L2 normalization, at pixel (q,u) and channel z.

[0072] By performing the above operations, this solution addresses the problems in existing neurosurgical image processing systems, such as the lack of prominent lesion features in neurosurgical images, the neglect of contextual relationships between regions, and the tendency of traditional enhancement methods to result in insignificant lesion enhancement or excessive enhancement of normal tissue, leading to inaccurate neurosurgical image processing. This solution employs a five-layer encoding and decoding structure. Each encoding layer includes a region-adaptive convolutional block and a cross-regional context aggregation layer, assigning differentiated weights to lesions and normal tissues to enhance lesion features and integrate global context, providing high-quality encoded features for precise enhancement. Each decoding layer includes a gated deconvolutional block and a content-adaptive attention layer, resulting in an enhanced neurosurgical image. Combined with skip connections to preserve details, it precisely focuses on lesions, achieving significant enhancement of lesion regions and moderate preservation of normal tissue. The pixel-level consistency loss, lesion region detail enhancement loss, and encoding-decoding feature consistency loss are weighted and summed to obtain the total loss. The model parameters are iteratively optimized using the Adam optimizer, and the multi-loss collaborative constraints ensure that the enhanced image remains consistent with the original image while enhancing lesion details, thus improving the overall accuracy and clinical reliability of neurosurgical image processing.

[0073] Example 5, see Figure 1 This embodiment is based on the above embodiment. In the neurosurgical image processing module, real-time neurosurgical images are acquired, preprocessed, and then sequentially input into the neurosurgical image denoising module and the neurosurgical image enhancement module to obtain the final neurosurgical image processing result.

[0074] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0076] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent neurosurgical image processing system, characterized in that: It includes a neurosurgical image acquisition module, a neurosurgical image denoising module, a neurosurgical image enhancement module, and a neurosurgical image processing module; The neurosurgical image acquisition module acquires historical neurosurgical images and performs preprocessing, and then sends the data to the neurosurgical image denoising module; The neurosurgical image denoising module receives data sent by the neurosurgical image acquisition module, extracts a comprehensive feature map, obtains a lesion mask through an adaptive threshold, divides the lesion area and normal tissue area, dynamically sets the number of decomposition layers and directions based on the lesion mask and local structural orientation entropy, obtains low-frequency approximation components and high-frequency detail components, filters the low-frequency approximation components based on dynamically adjusted guided filtering parameters, performs soft thresholding on the high-frequency detail components based on adaptive partitioning thresholds, reconstructs a denoised neurosurgical image, and sends the data to the neurosurgical image enhancement module. The neurosurgical image enhancement module receives data sent by the neurosurgical image denoising module and adopts a five-layer encoding and decoding structure. Each encoding layer contains a region-adaptive convolutional block and a cross-regional context aggregation layer. Each decoding layer contains a gated deconvolutional block and a content-adaptive attention layer to obtain an enhanced neurosurgical image. The total loss is obtained based on pixel-level consistency loss, lesion region detail enhancement loss and encoding-decoding feature consistency loss, and the data is sent to the neurosurgical image processing module. The neurosurgical image processing module receives data sent by the neurosurgical image enhancement module, and obtains the processing results of real-time neurosurgical images through the neurosurgical image denoising module and the neurosurgical image enhancement module.

2. The intelligent neurosurgical image processing system according to claim 1, characterized in that: The neurosurgical image denoising module includes a multi-feature fusion lesion pre-detection unit, a structure-guided decomposition unit, a low-frequency guided filtering unit, a high-frequency adaptive thresholding unit, and a denoised image reconstruction unit, specifically comprising the following: Multi-feature fusion lesion pre-detection unit: Extract grayscale features, texture features and contrast features from neurosurgical images, normalize the three features respectively, and then fuse them by weight to obtain a comprehensive feature map. The adaptive threshold is calculated by the maximum inter-class variance method to generate candidate lesion region labels. Morphological operations are performed on the candidate lesion region labels to generate lesion masks. The comprehensive feature map is divided into lesion regions and normal tissue regions. Structurally guided decomposition unit; Low-frequency steerable filter unit; High-frequency adaptive threshold unit; The denoising image reconstruction unit performs non-subsampled shear wave inverse transform on the low-frequency approximation optimization component and the high-frequency detail optimization component to reconstruct a denoised neurosurgical image.

3. The intelligent neurosurgical image processing system according to claim 2, characterized in that: The structure-guided decomposition unit calculates the proportion of gradient magnitude in each direction within a local window to the total gradient magnitude for each pixel in the neurosurgical image, obtaining the local structural orientation entropy of the pixel. It dynamically sets the number of decomposition layers and the number of decomposition directions based on the lesion mask and the local structural orientation entropy. Based on the number of decomposition layers and the number of decomposition directions, it performs a non-subsampled shear wave transform on the neurosurgical image to obtain low-frequency approximate components and high-frequency detail components.

4. The intelligent neurosurgical image processing system according to claim 3, characterized in that: The low-frequency guided filtering unit dynamically adjusts the guided filtering parameters based on the lesion mask and local structural orientation entropy for the low-frequency approximation component. Using the low-frequency approximation component itself as the guiding image, it calculates the local mean and local variance within the filtering window centered on each pixel, solves for the linear coefficients, and obtains the filtered low-frequency approximation optimized component.

5. The intelligent neurosurgical image processing system according to claim 4, characterized in that: The high-frequency adaptive thresholding unit is based on the basic noise standard deviation and global mean of the high-frequency detail components. It combines the deviation of each pixel's gray value from the mean to calculate the skewness and kurtosis of the noise distribution, obtains the noise statistical correction factor, determines the adaptive partition threshold of each pixel according to the lesion mask, and performs soft thresholding on the high-frequency detail components to obtain the high-frequency detail optimized components.

6. The intelligent neurosurgical image processing system according to claim 5, characterized in that: The neurosurgical image enhancement module includes a region context coding unit, a gated attention decoding unit, and a multi-loss joint optimization unit, specifically comprising the following: The region context encoding unit concatenates the denoised neurosurgical image with the lesion mask along the channel dimension to form the initial input feature map. It adopts a five-layer encoding structure, with each layer containing a region adaptive convolutional block and a cross-region context aggregation layer. The region adaptive convolutional block applies different depth-separable convolutional weights to the lesion region and the normal tissue region according to the lesion mask to achieve region adaptive feature extraction. The cross-regional context aggregation layer obtains the global context vector through global pooling operations, reconstructs the regional adaptive features, and obtains a condensed context feature map; Gated attention decoding unit; Multi-loss joint optimization unit.

7. The intelligent neurosurgical image processing system according to claim 6, characterized in that: The gated attention decoding unit concatenates the condensed context feature map output from the last layer of the encoder with the lesion mask along the channel dimension to form a mask-enhanced feature map. It adopts a five-layer decoding structure, with each layer containing a gated deconvolution block and a content-adaptive attention layer. The gated deconvolution block upsamples the mask-enhanced feature map through transposed convolution and makes skip connections with the corresponding layer features of the encoder to obtain fused features using a gating mechanism. The content-adaptive attention layer uses the lesion mask as a prior and performs spatial and channel modulation on the fused features to obtain an optimized enhanced feature map. The enhanced neurosurgical image is obtained based on the feature map output from the last layer of the decoder.

8. The intelligent neurosurgical image processing system according to claim 7, characterized in that: The multi-loss joint optimization unit is to sum the pixel-level consistency loss, lesion region detail enhancement loss and encoder-decoder feature consistency loss in a weighted manner to obtain the total loss. The Adam optimizer is used to calculate the gradient of the total loss with respect to all trainable parameters in the region context encoding unit and the gated attention decoding unit in each iteration. The parameters are updated by gradient descent to minimize the total loss, and the iteration continues until the loss converges.

9. The intelligent neurosurgical image processing system according to claim 8, characterized in that: The neurosurgical image processing module acquires real-time neurosurgical images, which are then preprocessed and sequentially input into the neurosurgical image denoising module and the neurosurgical image enhancement module to obtain the processing results of the real-time neurosurgical images.

Images