A brain tumor image recognition diagnosis method and system for neurosurgery

Abstract

This application provides a method and system for image recognition and diagnosis of brain tumors in neurosurgery. First, the brain tumor image is preprocessed. The preprocessed brain tumor image is then decomposed into at least two feature component maps that complement each other in terms of structural and texture information. Next, each feature component map is sequentially subjected to a first-stage adaptive nonlocal mean filtering based on local statistical characteristics and a second-stage filtering based on structural correlation similarity measurement. The filtered feature component maps are then fused and reconstructed to obtain a feature-enhanced brain tumor image. This feature-enhanced brain tumor image is input into a pre-trained brain tumor recognition model, and the neurosurgical diagnosis results are obtained through the model. Using the scheme of this application, a two-stage adaptive filtering based on local scattering statistics and structural correlation measurement can be achieved, thereby enhancing tumor edge and texture features while suppressing ultrasound speckle, thus improving the accuracy of segmentation and diagnosis.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to a method and system for image recognition and diagnosis of brain tumors in neurosurgery. Background Technology

[0002] In neurosurgical clinical practice, ultrasound imaging is often used for intraoperative navigation and assessment of brain tumors due to its advantages of being real-time and radiation-free. However, the inherent speckle noise in ultrasound images severely interferes with the clarity of tumor boundaries and the identification of internal textures, directly affecting the accuracy of subsequent automated segmentation and diagnosis. Traditional image filtering methods often struggle to achieve an ideal balance between effectively suppressing noise and meticulously preserving anatomical structures when dealing with this type of noise, becoming a key bottleneck in improving diagnostic accuracy.

[0003] In existing technologies, the suppression of speckle noise in medical images mainly relies on various filtering algorithms. However, these methods generally face a core contradiction: it is difficult to simultaneously suppress speckle noise and preserve the fine structure of tumors. Specifically, speckle noise has significant multiplicative properties and spatially varying statistical characteristics. Most existing filtering methods use fixed or globally uniform parameters, which cannot adapt to the scattering statistical characteristics of local image regions. This results in insufficient denoising in homogeneous tissue regions, while excessive smoothing occurs in high-contrast regions such as tumor edges, losing key edge information defining the tumor's extent and internal texture patterns reflecting its biological characteristics. This deficiency directly impacts downstream segmentation and recognition models, causing blurred tumor boundary segmentation and distorted internal feature extraction, ultimately affecting the reliability of diagnostic conclusions such as benign or malignant tumor determination. Therefore, how to achieve two-stage adaptive filtering based on local scattering statistics and structural correlation measurements, thereby suppressing ultrasound speckle while enhancing tumor edge and texture features and improving the accuracy of segmentation and diagnosis, has become a challenge for the industry. Summary of the Invention

[0004] This application provides a brain tumor image recognition and diagnosis method and system for neurosurgery, which can realize two-stage adaptive filtering based on local scattering statistics and structural correlation measurement, thereby enhancing tumor edge and texture features while suppressing ultrasound speckle, and improving the accuracy of segmentation and diagnosis.

[0005] In a first aspect, this application provides a brain tumor image recognition and diagnosis method for neurosurgery, comprising the following steps: Images of brain tumors to be diagnosed are collected, and the brain tumor images are preprocessed to enhance image contrast and standardize image size and image intensity. The preprocessed brain tumor image is decomposed into at least two feature component maps that complement each other in terms of structural and texture information. Then, a first-stage adaptive nonlocal mean filtering based on local statistical characteristics and a second-stage filtering based on structural correlation similarity measurement are sequentially performed on each feature component map. The filtered feature component images are fused and reconstructed to obtain a feature-enhanced brain tumor image; The enhanced brain tumor image is input into a pre-trained brain tumor recognition model, which outputs a segmentation mask of the tumor region and corresponding tumor attribute diagnostic information. The segmentation mask is then refined at the edges, and combined with the tumor attribute diagnostic information, a neurosurgical auxiliary diagnostic report is generated.

[0006] Preferably, decomposing the preprocessed brain tumor image into at least two feature component maps where structural and texture information complement each other specifically includes: The preprocessed brain tumor images are subjected to orthogonal decomposition under phase consistency constraints, which projects the image information into multiple feature subspaces with complementary information. In the feature subspace, at least one structural component map, which is mainly composed of low-frequency spatial changes and is used to characterize the overall contour and anatomical structure continuity of the tumor, and at least one texture component map, which is mainly composed of high-frequency local changes and is used to characterize the edge undulations and internal texture differences of the tumor, are generated. When there are multi-scale structures or complex textures in the image, the texture components are further subdivided into multiple texture sub-components with different focuses, so as to obtain no less than two feature component sets that are distinct from each other in terms of frequency characteristics and spatial representation. Each feature component map in the feature component map set is subjected to amplitude normalization and spatial consistency correction to ensure that different feature component maps can be processed separately in the filtering process and maintain reconfigurability.

[0007] Preferably, the first-stage adaptive nonlocal mean filtering based on local statistical characteristics and the second-stage filtering based on structural correlation similarity measurement are sequentially performed on each feature component map, specifically including: For each feature component map, a search window centered on the target pixel is constructed, and the local statistical descriptors of the search window are extracted to characterize the signal distribution characteristics of the component at the corresponding spatial location. Based on the local statistical descriptor, corresponding attenuation control parameters are adaptively configured for different target pixels, and a first-stage non-local mean weighting operation is performed within the search window to obtain the first filtering result. Using the first filtering result as a reference, within the same feature component map, a structural correlation measurement model is constructed between the local features corresponding to the target pixel and the local features of the candidate pixels in the search window; The pixel weights calculated in the first stage are modulated a second time according to the structural correlation measurement model, and the first filtering result is filtered in the second stage to obtain the filtering result of the corresponding feature components.

[0008] Preferably, the process of inputting the enhanced brain tumor image into a pre-trained brain tumor recognition model and outputting a segmentation mask of the tumor region and corresponding tumor attribute diagnostic information specifically includes: The brain tumor images after feature enhancement are normalized to meet the input constraints of the neural network; The standardized brain tumor image is input into the brain tumor recognition model to complete feature extraction; Pixel-level segmentation masks corresponding to tumor regions are generated based on the output of neural networks; Discriminative features are extracted within the tumor region defined by the pixel-level segmentation mask, and tumor attribute diagnostic information associated with the tumor region is output.

[0009] Preferably, the process of refining the edges of the segmentation mask and generating a neurosurgical auxiliary diagnostic report by combining the tumor attribute diagnostic information specifically includes: Pixel-level boundary localization and contour continuity correction are performed on the segmentation mask to obtain refined tumor edges; The refined tumor margins are spatially aligned with the original brain tumor image to ensure consistency of anatomical structure. The alignment results are fused with tumor attribute diagnostic information to form a standardized diagnostic dataset; Diagnostic reports for neurosurgical decision support are generated based on the standardized diagnostic dataset.

[0010] Preferably, the preprocessing includes: standardizing the size and grayscale intensity of the brain tumor image and enhancing the image contrast.

[0011] Preferably, the brain tumor image is an ultrasound image.

[0012] Secondly, this application provides a brain tumor image recognition and diagnostic system for neurosurgery, comprising: A preprocessing module is used to collect brain tumor images to be diagnosed and to preprocess the brain tumor images. The preprocessing is used to enhance image contrast and standardize image size and image intensity. The filtering module is used to decompose the preprocessed brain tumor image into at least two feature component maps that complement each other in terms of structural and texture information. Then, the first-stage adaptive nonlocal mean filtering based on local statistical characteristics and the second-stage filtering based on structural correlation similarity measurement are performed on each feature component map in sequence. The fusion module is used to fuse and reconstruct the filtered feature component maps to obtain a feature-enhanced brain tumor image. The recognition module is used to input the feature-enhanced brain tumor image into a pre-trained brain tumor recognition model and output the segmentation mask of the tumor region and the corresponding tumor attribute diagnostic information. The report generation module is used to perform edge refinement post-processing on the segmentation mask and, in conjunction with the tumor attribute diagnostic information, generate a neurosurgical auxiliary diagnostic report.

[0013] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described brain tumor image recognition and diagnosis method for neurosurgery.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described brain tumor image recognition and diagnosis method for neurosurgery.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This technical solution addresses the long-standing challenge of significant speckle noise and the easy obscuring of structural information in neurosurgical ultrasound brain tumor images. It organically reconstructs key steps from the perspective of the entire image data processing chain, resulting in a clear hierarchy and progressively enhanced technical effect. First, by implementing contrast enhancement and standardized processing of size and intensity in brain tumor images, the impact of different imaging conditions, probe parameters, and individual differences on image distribution is effectively reduced. Second, the preprocessed image is decomposed into feature component maps that complement structural and textural information, allowing for differentiation between macroscopic anatomical structures such as tumor contours and fine-grained texture features reflecting internal tissue differences at the representation level, creating conditions for differentiated processing. Based on this, adaptive nonlocal processing based on local statistical properties is sequentially applied to each feature component map. The proposed method employs a two-stage filtering approach: mean filtering and a second-stage filtering based on structural correlation similarity. The first stage focuses on suppressing spatially significant ultrasound speckle noise based on local scattering statistics, while the second stage introduces structural similarity constraints to specifically enhance the continuity of tumor edges and the consistency of tissue morphology. This reduces noise interference while preventing excessive smoothing of key boundaries and texture information. Then, by fusing and reconstructing the filtered feature component images, effective features from different information dimensions are presented collaboratively in the same image, resulting in a feature-enhanced brain tumor image with controlled noise levels and clear structural expression. This image, used as input to the recognition model, significantly improves the boundary accuracy of the segmentation mask and the stability of tumor attribute discrimination. After edge refinement, it generates an auxiliary diagnostic report that better meets the needs of neurosurgical clinical interpretation. In summary, this method utilizes a two-stage adaptive filtering approach based on local scattering statistics and structural correlation metrics to enhance tumor edge and texture features while suppressing ultrasound speckle, thereby improving the accuracy of segmentation and diagnosis. Attached Figure Description

[0016] Figure 1 This is an exemplary flowchart of a brain tumor image recognition and diagnosis method for neurosurgery, according to some embodiments of this application; Figure 2 This is a schematic flowchart illustrating the decomposition of brain tumor images according to some embodiments of this application; Figure 3 This is a flowchart illustrating the determination of enhanced brain tumor images based on some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a brain tumor image recognition and diagnostic system for neurosurgery, according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device for implementing a brain tumor image recognition and diagnosis method for neurosurgery, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] refer to Figure 1 The figure is an exemplary flowchart of a brain tumor image recognition and diagnosis method for neurosurgery, according to some embodiments of this application. This brain tumor image recognition and diagnosis method for neurosurgery mainly includes the following steps: In step 101, images of brain tumors to be diagnosed are collected and preprocessed. The preprocessing is used to enhance image contrast and standardize image size and intensity.

[0019] It should be noted that in neurosurgical clinical practice, ultrasound imaging, due to its real-time and radiation-free advantages, is often used for intraoperative localization and boundary confirmation of brain tumors. However, the inherent speckle noise in ultrasound images, a multiplicative and structured noise generated by coherent wave interference, severely interferes with the identification of tumor fine structures (such as infiltrative edges and internal cystic changes) and significantly reduces the accuracy of deep learning-based automatic segmentation and diagnostic models. Traditional single-filtering methods often face a dilemma in suppressing speckle noise: over-smoothing leads to edge loss, while under-smoothing results in residual noise. To address these technical challenges, this application introduces a two-stage adaptive nonlocal mean filtering architecture. Its core significance lies in achieving targeted and hierarchical suppression of speckle noise: the first stage performs preliminary adaptive smoothing based on local statistical characteristics, aiming to quickly suppress speckle intensity in uniform regions; the second stage introduces a similarity metric based on structural correlation to reweight the preliminary filtering results, thereby further suppressing residual noise while keenly identifying and preserving texture details and anatomical edges representing tumor biological characteristics. This dual filtering strategy, which first suppresses globally and then refines the structure, is essentially a synergistic response to the spatial non-uniformity and structural confusion of speckle noise. It aims to provide the subsequent neural network with feature-enhanced images that significantly improve the signal-to-noise ratio and maximize the preservation of structural integrity, thereby laying the foundation for achieving high-precision automatic tumor segmentation and attribute diagnosis.

[0020] It should also be noted that the brain tumor images in this application are specifically ultrasound images containing speckle noise.

[0021] In some embodiments, the preprocessing includes: standardizing the size and grayscale intensity of brain tumor images and enhancing image contrast. In specific applications, after acquiring the brain tumor image to be diagnosed, preprocessing is performed first. The purpose is to uniformly convert original images of varying sources and qualities into a standardized format suitable for subsequent advanced analysis and recognition algorithms. This preprocessing process mainly includes the following three steps: image size standardization, which involves uniformly adjusting the pixel size of all input images to a preset fixed resolution, such as 512×512 pixels. This is typically achieved using bilinear or bicubic interpolation algorithms to ensure that all images maintain consistency in spatial dimensions, facilitating batch processing and model input; grayscale intensity standardization... Normalizing the grayscale range of an image to a standard interval, such as [0, 255], involves first calculating the minimum and maximum grayscale values ​​of a single image or the entire dataset, and then mapping the intensity of each pixel to the target interval through a linear transformation. This aims to eliminate the differences in brightness and contrast caused by different scanning devices. Image contrast enhancement, which aims to improve the visual separability between tumor tissue and normal tissue, employs contrast enhancement techniques. A preferred but not exclusive method is to limit the contrast adaptive histogram equalization. This method divides the image into several small regions, performs histogram equalization on each region, and suppresses excessive noise amplification through contrast limiting, thereby enhancing local details while maintaining the natural appearance of the image.

[0022] In step 102, the preprocessed brain tumor image is decomposed into at least two feature component maps that complement each other in terms of structural and texture information. Then, a first-stage adaptive nonlocal mean filtering based on local statistical characteristics and a second-stage filtering based on structural correlation similarity measurement are sequentially performed on each feature component map.

[0023] In some embodiments, reference Figure 2 As shown in the figure, this is a schematic diagram of the brain tumor image decomposition process in some embodiments of this application. In this embodiment, the decomposition of the preprocessed brain tumor image into at least two feature component maps whose structural and texture information complement each other can be achieved by the following steps: In step 1021, the preprocessed brain tumor image is subjected to orthogonal decomposition under phase consistency constraints, and the image information is projected to multiple feature subspaces with complementary information. In step 1022, in the feature subspace, at least one structural component map, which is mainly composed of low-frequency spatial changes and is used to characterize the overall contour and anatomical structure continuity of the tumor, and at least one texture component map, which is mainly composed of high-frequency local changes and is used to characterize the edge undulations and internal texture differences of the tumor, are generated. In step 1023, when there are multi-scale structures or complex textures in the image, the texture components are further subdivided into multiple texture sub-components with different focuses, so as to obtain no less than two feature component sets that are distinct from each other in terms of frequency characteristics and spatial representation. In step 1024, amplitude normalization and spatial consistency correction are performed on each feature component map in the feature component map set to ensure that different feature component maps can be processed separately in the filtering process and maintain reconfigurability.

[0024] It should be noted that the phase consistency constraint in this application is a mathematical condition used to ensure that key phase information of image edges and structures is preserved during the decomposition process; the feature subspace refers to a vector space that carries different frequency components and structural features in the image; the structural component map refers to a component image that characterizes low-frequency spatial changes in the image and reflects the overall contour and anatomical structure continuity of the tumor; the texture component map is a component image that characterizes high-frequency local changes in the image and reflects the undulation of tumor edges and differences in internal texture; the feature component map set is an image set that includes all structural component maps and texture sub-component maps.

[0025] In practical applications, firstly, orthogonal decomposition processing under phase consistency constraints is performed on the preprocessed brain tumor image. Projecting the image information to multiple complementary feature subspaces can be achieved in the following way: an orthogonal decomposition method based on discrete wavelet transform or curvelet transform can be applied to the preprocessed brain tumor image to decompose the image into multiple sub-bands with different frequency directions. During this process, phase consistency calculation is introduced, and edge and structural features are detected by the phase stability in the local Fourier components. Using this as a constraint, the decomposition process is guided to retain the phase information of these key regions. Finally, the image information is projected into multiple subspaces that are complementary in frequency and direction to form a set of preliminary feature sub-bands. The preprocessed brain tumor image is used as input, and the multiple decomposed sub-bands are used as feature subspaces as output.

[0026] Secondly, in the feature subspace, at least one structural component map, which is mainly composed of low-frequency spatial variations and is used to characterize the overall contour and anatomical structure continuity of the tumor, and at least one texture component map, which is mainly composed of high-frequency local variations and is used to characterize the edge undulations and internal texture differences of the tumor, can be generated in the following way: Based on the feature subspace, low-frequency components are extracted from the decomposed subbands through a low-pass filter (such as a Gaussian filter) and synthesized into a structural component map. At the same time, high-frequency components are extracted through a high-pass or band-pass filter and synthesized into a texture component map. If wavelet transform is used, the approximate subband is used to synthesize the structural component, and the detail subband is used to synthesize the texture component. The low-frequency subbands in the feature subspace are synthesized into a structural component map, and the high-frequency subbands are synthesized into a texture component map. Then, when there are multi-scale structures or complex textures in the image, the texture components are further subdivided into multiple texture sub-components with different emphases to obtain at least two feature component maps that are distinct from each other in terms of frequency characteristics and spatial representation. This can be achieved in the following way: when there are multi-scale structures or complex textures in the image, the texture component map is further decomposed into multiple texture sub-components, for example, by using a Gaussian difference filter bank or wavelet packet transform, to decompose the texture components into multiple texture sub-components according to different scales or directions; each texture sub-component emphasizes different frequency ranges or local directional features, thereby forming a feature component map that includes a structure component map and multiple texture sub-component maps. The texture component map is further subdivided into multiple texture sub-components through a multi-scale decomposition method, and finally a feature component map is formed.

[0027] Finally, amplitude normalization and spatial consistency correction are performed on each feature component map in the feature component map set to ensure that different feature component maps can be processed separately in the filtering process and maintain reconstructibility. This can be achieved in the following way: For each component map in the feature component map set, linear or histogram normalization is performed to map its pixel values ​​to the interval [0, 1] or [0, 255]. Then, spatial offset and phase difference between each component map are corrected by phase consistency alignment or edge alignment methods to ensure that they correspond to the same anatomical structure in the same spatial location.

[0028] In some embodiments, the sequential execution of a first-stage adaptive nonlocal mean filtering based on local statistical properties and a second-stage filtering based on structural correlation similarity measurement on each feature component map can be achieved using the following steps: For each feature component map, a search window centered on the target pixel is constructed, and the local statistical descriptors of the search window are extracted to characterize the signal distribution characteristics of the component at the corresponding spatial location. Based on the local statistical descriptor, corresponding attenuation control parameters are adaptively configured for different target pixels, and a first-stage non-local mean weighting operation is performed within the search window to obtain the first filtering result. Using the first filtering result as a reference, within the same feature component map, a structural correlation measurement model is constructed between the local features corresponding to the target pixel and the local features of the candidate pixels in the search window; The pixel weights calculated in the first stage are modulated a second time according to the structural correlation measurement model, and the first filtering result is filtered in the second stage to obtain the filtering result of the corresponding feature components.

[0029] It should be noted that, in this application, the search window refers to a local neighborhood region defined around the target pixel in the feature component map; the local statistical descriptor is a statistical quantity used to characterize the distribution characteristics of pixel values ​​within the search window; the decay control parameter is an adaptive parameter used to control the weight decay rate in nonlocal mean filtering; the structural correlation measurement model is a mathematical model that quantifies the structural similarity between the local features of the target pixel and the local features of the candidate pixel; the pixel weight is a numerical value used to represent the degree of contribution of the candidate pixel to the target pixel in nonlocal mean filtering; and the filtering result is used to represent the feature component map finally obtained after two-stage filtering.

[0030] It should also be noted that, compared to existing technologies, the two-stage adaptive nonlocal mean filtering method proposed in this solution fundamentally solves the technical contradiction that a single filtering framework cannot simultaneously achieve both spatial adaptive suppression of ultrasound speckle noise and complete preservation of complex anatomical structures. Specifically: the first stage dynamically configures attenuation control parameters through local statistical descriptors, achieving preliminary adaptive identification and suppression of speckle noise intensity, overcoming the shortcomings of fixed-parameter filtering, which is insufficiently smooth in uniform regions and overly smooth in textured regions; the second stage introduces a structural correlation measurement model to modulate the weights a second time. Its innovation lies in utilizing the enhanced signal-to-noise ratio after the first-stage filtering to more accurately distinguish between real tissue edges and residual noise, thereby further suppressing speckle while significantly reducing false smoothing of key diagnostic information such as tumor fine structures and invasive boundaries. The two filtering processes form a synergistic mechanism of suppression followed by correction. The final technical effect is that while significantly improving the equivalent number of looks (ENL), it achieves a lower speckle suppression index (SSI) and better structural fidelity, providing high-quality feature images with strong noise robustness and high detail integrity for neural network segmentation and diagnosis.

[0031] In practical applications, firstly, a search window centered on the target pixel is constructed, and the local statistical descriptive quantities of the search window are extracted. This can be achieved in the following way: For each pixel in the feature component map (called the target pixel), a square region of a fixed size is defined as the search window centered on the target pixel, for example, a size of 21×21 pixels. Then, within this search window, the mean, variance, standard deviation, and other statistical quantities of the gray values ​​of all pixels are calculated. These statistical quantities are the local statistical descriptive quantities.

[0032] Secondly, based on the aforementioned local statistical descriptors, the corresponding attenuation control parameters for adaptive configuration of different target pixels can be achieved using the following formula:

[0033] in, Indicates pixel position Attenuation control parameters corresponding to the target pixel; The local window standard deviation reflects the local speckle intensity. This is the local window mean, reflecting the local signal strength. This is the local variation coefficient, used to describe the relative intensity of speckle; This represents the local gradient magnitude, reflecting the edge intensity. The maximum gradient magnitude in the image is used as a normalization factor. The edge enhancement coefficient, typically set to 0.5~1.5, is used to control the edge protection strength. It should be further noted that the design in this application primarily addresses the multiplicative characteristics and spatial variability of ultrasonic speckle noise. The core of the formula consists of two parts: the local variation coefficient... With normalized gradient modulation term The local coefficient of variation directly characterizes the intensity of speckle noise because the scattering bodies in a uniform region of an ultrasound image are relatively uniformly distributed, and their signal mean is relatively low. Higher and standard deviation A lower gradient results in a smaller coefficient of variation, leading to a corresponding decrease in the attenuation control parameter h1. This results in stronger weight attenuation in nonlocal mean filtering, thus significantly suppressing speckle patterns in uniform background regions. Conversely, in structurally complex regions such as tumor edges and calcifications, local grayscale changes are drastic, increasing the coefficient of variation. Consequently, the attenuation control parameter h1 increases, weakening the filtering intensity and preventing detail blurring. The normalized gradient term further introduces edge-awareness: gradient magnitude... The response is significant at the edges of anatomical structures, as indicated by the edge enhancement coefficient. Perform weighted normalization to the maximum value This allows for additional gain at edge position h1, thereby reducing the weight decay rate of edge pixels and neighboring pixels in the filtering calculation and achieving active protection of important anatomical boundaries. Overall, this parameter design, by fusing local statistical characteristics and spatial gradient information, enables the filtering process to adaptively distinguish between noise regions and structural regions, suppressing speckle while preserving crucial structural information for the subsequent second-stage fine filtering based on structural correlation.

[0034] Furthermore, the first-stage non-local mean weighted operation within the search window to obtain the first filtering result can be implemented as follows: within the search window, for each candidate pixel, calculate the Gaussian weighted Euclidean distance between the matching window centered on the target pixel and the matching window centered on the candidate pixel as a similarity measure; then, calculate the weight of the candidate pixel based on this distance and the attenuation control parameter; after normalizing all weights, perform a weighted average on the gray values ​​of the candidate pixels to obtain the first filtering result of the target pixel; combine the first filtering results of each target pixel to form the first filtering result of the entire feature component map.

[0035] Then, with the first filtering result as a reference, within the same feature component map, the structural correlation measurement model between the local features corresponding to the target pixel and the local features of the candidate pixels in the search window can be implemented in the following way: with the first filtering result as a reference, for each target pixel, its neighborhood (e.g., a 5×5 window) is extracted as a local feature; similarly, for each candidate pixel in the search window, a neighborhood of the same size is extracted; then, the Pearson correlation coefficient between the local features of the target pixel and the local features of the candidate pixels is calculated as a measure of structural correlation; the Pearson correlation coefficient reflects the degree of linear correlation between two local regions, and the closer the value is to 1, the more similar the structures are; formulating this correlation coefficient calculation process is the structural correlation measurement model, and the local features of the target pixel and the candidate pixels are input into the structural correlation measurement model to obtain their structural similarity measurement values.

[0036] Finally, the pixel weights calculated in the first stage are modulated a second time according to the structural correlation measurement model, and the first filtering result is filtered in the second stage to obtain the filtering result of the corresponding feature component. This can be achieved in the following way: the pixel weights calculated in the first stage are modulated according to the structural correlation measurement value between the target pixel and each candidate pixel; for example, the weights in the first stage are multiplied by the exponential function of the structural correlation measurement value, so that the weights of candidate pixels with high structural correlation increase and the weights of candidate pixels with low structural correlation decrease; further, the gray values ​​of the candidate pixels (which can be the original gray values ​​or the first filtering result) are weighted and averaged using the modulated weights to obtain the second-stage filtering result of the target pixel; this operation is performed on all target pixels to obtain the final filtering result of the feature component map, and the modulated weights are applied to the gray values ​​of the candidate pixels to obtain the filtering result of the target pixel.

[0037] In step 103, the filtered feature component images are fused and reconstructed to obtain a feature-enhanced brain tumor image.

[0038] In some embodiments, reference Figure 3As shown in the figure, this is a flowchart illustrating the process of determining the feature-enhanced brain tumor image in some embodiments of this application. In this embodiment, the feature-enhanced brain tumor image is obtained by fusing and reconstructing the filtered feature component images using the following steps: In step 1031, numerical scaling and phase consistency corrections are performed on each feature component map after the two-stage filtering process, so that each feature component map is at a consistent numerical and phase reference before fusion and reconstruction. In step 1032, based on the local statistical consistency of each feature component map at the corresponding spatial location, a component confidence constraint is constructed to characterize the reliability of component reconstruction. In step 1033, under the control of the component confidence constraint, the structural component and the texture component are jointly reconstructed, and spatial consistency constraint is introduced in the reconstruction process to coordinate the joint relationship of different components at adjacent pixels. In step 1034, the multi-component reconstruction results that satisfy the component confidence constraints and spatial consistency constraints are mapped back to the unified spatial domain to generate a feature-enhanced brain tumor image.

[0039] It should be noted that the phase consistency correction in this application refers to the correction process of aligning the phase information between each feature component map to eliminate the phase shift that may be introduced during decomposition and filtering; the local statistical consistency is a statistical metric used to describe the stability and similarity of pixel value distribution at the same spatial location of each feature component map; the component confidence constraint refers to a weighted mapping model that quantifies the reliability of the feature component map during reconstruction; the joint reconstruction refers to the data fusion process of combining structural components and texture components into a complete image according to specific rules; the spatial consistency constraint refers to the rule that ensures the smooth and continuous component contribution relationship between adjacent pixels during the fusion process, avoiding block effects or discontinuous jumps; the feature-enhanced brain tumor image represents the final output diagnostic image with suppressed noise and enhanced structural and texture details.

[0040] In practical applications, firstly, numerical scaling and phase consistency corrections are performed on each feature component map after the two-stage filtering process to ensure that each feature component map is at a consistent numerical and phase benchmark before fusion and reconstruction. This can be achieved by the following method: linearly normalizing the pixel values ​​of each feature component map to map them to a unified numerical range; then calculating the phase consistency information of each component map within a local window and comparing it with the phase information of the original preprocessed image; and using a phase alignment algorithm to fine-tune each component map to correct the local phase deviation that may be caused by decomposition and filtering, resulting in a set of feature component maps with consistent numerical ranges and phase alignment. The feature component maps that have undergone numerical scaling and phase consistency corrections are then used as inputs at a consistent benchmark.

[0041] Secondly, based on the local statistical consistency of each feature component map at its corresponding spatial location, the component confidence constraint used to characterize the reliability of component reconstruction can be implemented as follows: For each spatial location, calculate the statistics, such as variance or standard deviation, of each feature component map within its small neighborhood at that location to measure the local consistency of the component. Then, construct a confidence function based on these local statistics: the smaller the local variance, the more stable and reliable the component information at that location, and the higher its confidence; conversely, the confidence decreases. Finally, generate a confidence weight map of the same size as the original image for each feature component map, and use the confidence weight map calculated based on local statistical consistency as the component confidence constraint. The reason for this is that the reliability of each component varies in different regions of the image. For example, structural components are more reliable in flat regions, while detail components are more reliable in textured regions. Constructing confidence constraints can give higher weights to components in reliable regions during fusion, improving the reconstruction quality.

[0042] Then, under the control of the component confidence constraints, the structural components and texture components are jointly reconstructed. Spatial consistency constraints are introduced during the reconstruction process to coordinate the joint relationship of different components at adjacent pixels. This can be achieved in the following way: using the component confidence constraints as the basis for the fusion weights of each component at the pixel level, the structural components and texture components are weighted and averaged to achieve preliminary joint reconstruction. In this process, spatial smoothness constraints are introduced, for example, by adding a regularization term for the difference in weights between adjacent pixels to the reconstruction objective function, which forces the weight distribution of adjacent pixels to change smoothly. This is only an example and is not intended to limit the specific scope of the invention. This process is solved by iterative optimization algorithms (such as gradient descent) to finally obtain a fused image that conforms to both the confidence distribution of each component and the spatial smoothness requirement. The obtained fused image is then used as the joint reconstruction result. The reason for this is that simple weighted averaging may cause visual defects in the weight abrupt change region of the fusion result, while introducing spatial consistency constraints can ensure that the fusion weights transition smoothly in space, thereby generating a visually coherent and natural fused image.

[0043] Finally, the multi-component reconstruction results that satisfy the component confidence constraints and spatial consistency constraints are mapped back to the unified spatial domain to generate the feature-enhanced brain tumor image. This can be achieved by mapping the obtained multi-component reconstruction results back to the gray value range of the original image through linear inverse transformation. This mapping process needs to maintain the relative contrast relationship between each pixel. Finally, an image that is spatially aligned with the original image, has a standardized gray value range, and has enhanced visual features is generated. The image mapped back to the unified spatial domain and whose gray value recovery is completed is used as the feature-enhanced brain tumor image.

[0044] In step 104, the enhanced brain tumor image is input into a pre-trained brain tumor recognition model, which outputs a segmentation mask of the tumor region and corresponding tumor attribute diagnostic information.

[0045] It should be noted that the brain tumor recognition model in this application is a deep neural network capable of simultaneously outputting a pixel-level segmentation mask of the tumor and its attribute diagnosis. Its pre-training process involves: first, constructing an encoder-decoder architecture, such as U-Net. The encoder consists of 4-5 stages, each containing two 3×3 convolutional layers (using He initialization parameters) followed by ReLU activation and 2×2 max pooling, with the number of channels doubling sequentially; the decoder symmetrically upsamples through 2×2 transposed convolutions, and concatenates the corresponding stage features from the encoder with skip connections, followed by two 3×3 convolutions. The model output is connected in parallel to the segmentation head (1×1 convolution followed by Sigmoid) and the classification head (global pooling followed by a fully connected layer). During training, a multi-task loss function (such as Dice loss + cross-entropy loss) is used. The Adam optimizer (initial learning rate 0.001, β1=0.9, β2=0.999) is used for end-to-end training on the labeled dataset. Data augmentation such as random rotation and flipping is used to improve generalization ability. Training is terminated and the optimal parameters are saved when the loss on the validation set no longer decreases. This is only an example and is not intended to limit the specific scope of the invention.

[0046] In some embodiments, the feature-enhanced brain tumor image is input into a pre-trained brain tumor recognition model, and the segmentation mask of the tumor region and the corresponding tumor attribute diagnostic information are output. This can be achieved by the following steps: The brain tumor images after feature enhancement are normalized to meet the input constraints of the neural network; The standardized brain tumor image is input into the brain tumor recognition model to complete feature extraction; Pixel-level segmentation masks corresponding to tumor regions are generated based on the output of neural networks; Discriminative features are extracted within the tumor region defined by the pixel-level segmentation mask, and tumor attribute diagnostic information associated with the tumor region is output.

[0047] It should be noted that the normalization processing in this application refers to adjusting the size, data type, and numerical range of the brain tumor image after feature enhancement to meet the standard operation requirements of the neural network model input; the pixel-level segmentation mask is a labeled image that accurately marks the location of the tumor region in the image in binary or probability value form with pixels as the smallest unit; the tumor attribute diagnostic information refers to the final judgment conclusion or descriptive result about the tumor category or nature obtained based on discriminant feature analysis.

[0048] In practical applications, firstly, the feature-enhanced brain tumor image is scaled or cropped spatially to a fixed input size preset by the neural network model, such as 224 pixels × 224 pixels; then, the pixel values ​​of the image are normalized to the specific numerical range used during model training. Common practices include subtracting the dataset mean and dividing by the standard deviation, or performing a linear mapping to the [0, 1] interval; simultaneously, it is ensured that the image data is converted to the data type specified by the model, obtaining an image that fully meets the model requirements in terms of size, numerical range, and data type. This processed image is then used as a normalized image that satisfies the neural network input constraints. Secondly, the normalized brain tumor image is input into a pre-trained encoding... In encoder-decoder neural network models, such as U-Net, the encoder extracts features from the input image in a progressively deeper manner through multiple cascaded convolutional and downsampling layers, resulting in a series of feature maps with progressively decreasing resolution but increasing semantic information. In the decoder, the model gradually restores the spatial resolution of the feature maps through upsampling operations. Simultaneously, a skip connection mechanism concatenates the corresponding feature maps at the same scale from the encoder stage with the decoder feature maps along the channel dimension, thereby achieving multi-level feature fusion. This effectively combines shallow detail information with deep semantic information, using the high-dimensional feature map obtained after encoder deep feature extraction and decoder feature fusion as the model's representation of the internal features of the input image. The structured representation is then generated. The final output layer of the neural network model typically generates a single-channel feature map with the same spatial size as the input image. The value of each pixel in this feature map represents the probability that the pixel is classified as belonging to a tumor region. A preset threshold (e.g., 0.5) is applied to this probability map for binarization decision processing: pixels with a probability value greater than or equal to the threshold are classified as tumor regions and assigned a label value of 1; pixels with a probability value less than the threshold are classified as non-tumor regions and assigned a label value of 0. This generates a binary image that spatially corresponds strictly to the original input image. The binary image after thresholding is used as the final pixel-level segmentation mask for the tumor region. Finally, in the generation... Within the region defined by the pixel-level segmentation mask, a set of predefined quantitative features are extracted from the original feature-enhanced image or the intermediate layer feature map of the model. These features may include morphological features (such as area, perimeter, and roundness), gray-level statistical features (such as mean, variance, and skewness), and texture features (such as contrast, energy, and homogeneity based on the gray-level co-occurrence matrix). These extracted discriminative features are then input into a classifier, such as a fully connected layer or a support vector machine, which is jointly trained with the segmentation network or trained independently. The classifier outputs one or more labels, such as the benign or malignant classification of tumors, the World Health Organization classification, or pathological type prediction. The judgment label calculated by the classifier based on the discriminative features is used as the final tumor attribute diagnostic information.

[0049] In step 105, the segmentation mask is subjected to edge refinement post-processing, and combined with the tumor attribute diagnostic information, a neurosurgical auxiliary diagnostic report is generated.

[0050] In some embodiments, the generation of a neurosurgical auxiliary diagnostic report by refining the edges of the segmentation mask and combining it with the tumor attribute diagnostic information can be achieved through the following steps: Pixel-level boundary localization and contour continuity correction are performed on the segmentation mask to obtain refined tumor edges; The refined tumor margins are spatially aligned with the original brain tumor image to ensure consistency of anatomical structure. The alignment results are fused with tumor attribute diagnostic information to form a standardized diagnostic dataset; Diagnostic reports for neurosurgical decision support are generated based on the standardized diagnostic dataset.

[0051] In practical application, firstly, a contour tracking algorithm is used to extract the tumor boundary pixels from the segmentation mask, and a smoothing algorithm is used to smooth and repair the gaps in the boundary to obtain a smooth and refined tumor edge. Secondly, according to the scaling or cropping ratio of the image during preprocessing, the coordinate mapping relationship is calculated to transform the coordinates of the refined edge and map them to the actual position in the original image, thus completing spatial alignment. Then, the aligned tumor contour coordinates, together with the tumor attribute diagnostic information (such as type, size, and grade) output by the neural network, are integrated into a structured record to form a standardized diagnostic dataset. Finally, based on this dataset, a schematic diagram is generated by overlaying the tumor contour onto the original image, and the diagnostic information is filled into a fixed-format text report template, merging and outputting a diagnostic report document with both text and images.

[0052] On the other hand, in some embodiments, this application provides a brain tumor image recognition and diagnostic system for neurosurgery, referring to... Figure 4 The figure is a schematic diagram of the structure of a brain tumor image recognition and diagnosis system for neurosurgery according to some embodiments of this application. The brain tumor image recognition and diagnosis system 400 for neurosurgery includes: a preprocessing module 401, a filtering module 402, a fusion module 403, a recognition module 404, and a report generation module 405, which are described below: The preprocessing module 401 is used to collect brain tumor images to be diagnosed and to preprocess the brain tumor images. The preprocessing is used to enhance image contrast and standardize image size and image intensity. The filtering module 402 is used to decompose the preprocessed brain tumor image into at least two feature component maps that complement each other in terms of structural and texture information, and then sequentially perform a first-stage adaptive nonlocal mean filtering based on local statistical characteristics and a second-stage filtering based on structural correlation similarity measurement on each feature component map. The fusion module 403 is used to fuse and reconstruct the filtered feature component maps to obtain a feature-enhanced brain tumor image. The recognition module 404 is used to input the feature-enhanced brain tumor image into a pre-trained brain tumor recognition model and output the segmentation mask of the tumor region and the corresponding tumor attribute diagnostic information. The report generation module 405 is used to perform edge refinement post-processing on the segmentation mask and, in conjunction with the tumor attribute diagnostic information, generate a neurosurgical auxiliary diagnostic report.

[0053] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described brain tumor image recognition and diagnosis method for neurosurgery.

[0054] In some embodiments, reference Figure 5 This figure is a schematic diagram of the structure of a computer device implementing a brain tumor image recognition and diagnosis method for neurosurgery, according to some embodiments of this application. The brain tumor image recognition and diagnosis method for neurosurgery described in the above embodiments can be... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0055] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0056] The communication bus 502 can be used to transmit information between the aforementioned components.

[0057] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0058] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The brain tumor image recognition and diagnosis method for neurosurgery in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0059] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0060] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0061] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0062] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described brain tumor image recognition and diagnosis method for neurosurgery.

[0063] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0064] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for image recognition and diagnosis of brain tumors in neurosurgery, characterized in that, include: Images of brain tumors to be diagnosed are collected, and the brain tumor images are preprocessed to enhance image contrast and standardize image size and image intensity. The preprocessed brain tumor image is decomposed into at least two feature component maps that complement each other in terms of structural and texture information. Then, a first-stage adaptive nonlocal mean filtering based on local statistical characteristics and a second-stage filtering based on structural correlation similarity measurement are sequentially performed on each feature component map. The filtered feature component images are fused and reconstructed to obtain a feature-enhanced brain tumor image; The enhanced brain tumor image is input into a pre-trained brain tumor recognition model, which outputs a segmentation mask of the tumor region and corresponding tumor attribute diagnostic information. The segmentation mask is then refined at the edges, and combined with the tumor attribute diagnostic information, a neurosurgical auxiliary diagnostic report is generated.

2. The method as described in claim 1, characterized in that, The preprocessed brain tumor image is decomposed into at least two feature component maps that complement each other in terms of structural and textural information. Specifically, this includes: The preprocessed brain tumor images are subjected to orthogonal decomposition under phase consistency constraints, which projects the image information into multiple feature subspaces with complementary information. In the feature subspace, at least one structural component map, which is mainly composed of low-frequency spatial changes and is used to characterize the overall contour and anatomical structure continuity of the tumor, and at least one texture component map, which is mainly composed of high-frequency local changes and is used to characterize the edge undulations and internal texture differences of the tumor, are generated. When there are multi-scale structures or complex textures in the image, the texture components are further subdivided into multiple texture sub-components with different focuses, so as to obtain no less than two feature component sets that are distinct from each other in terms of frequency characteristics and spatial representation. Each feature component map in the feature component map set is subjected to amplitude normalization and spatial consistency correction to ensure that different feature component maps can be processed separately in the filtering process and maintain reconfigurability.

3. The method as described in claim 1, characterized in that, The first-stage adaptive nonlocal mean filtering based on local statistical properties and the second-stage filtering based on structural correlation similarity measurement are performed sequentially on each feature component map. Specifically, this includes: For each feature component map, a search window centered on the target pixel is constructed, and the local statistical descriptors of the search window are extracted to characterize the signal distribution characteristics of the component at the corresponding spatial location. Based on the local statistical descriptor, corresponding attenuation control parameters are adaptively configured for different target pixels, and a first-stage non-local mean weighting operation is performed within the search window to obtain the first filtering result. Using the first filtering result as a reference, within the same feature component map, a structural correlation measurement model is constructed between the local features corresponding to the target pixel and the local features of the candidate pixels in the search window; The pixel weights calculated in the first stage are modulated a second time according to the structural correlation measurement model, and the first filtering result is filtered in the second stage to obtain the filtering result of the corresponding feature components.

4. The method as described in claim 1, characterized in that, The enhanced brain tumor image is input into a pre-trained brain tumor recognition model, and the output of the tumor region segmentation mask and corresponding tumor attribute diagnostic information specifically includes: The brain tumor images after feature enhancement are normalized to meet the input constraints of the neural network; The standardized brain tumor image is input into the brain tumor recognition model to complete feature extraction; Pixel-level segmentation masks corresponding to tumor regions are generated based on the output of neural networks; Discriminative features are extracted within the tumor region defined by the pixel-level segmentation mask, and tumor attribute diagnostic information associated with the tumor region is output.

5. The method as described in claim 1, characterized in that, The process of refining the edges of the segmentation mask and combining it with the tumor attribute diagnostic information to generate a neurosurgical auxiliary diagnostic report specifically includes: Pixel-level boundary localization and contour continuity correction are performed on the segmentation mask to obtain refined tumor edges; The refined tumor margins are spatially aligned with the original brain tumor image to ensure consistency of anatomical structure. The alignment results are fused with tumor attribute diagnostic information to form a standardized diagnostic dataset; Diagnostic reports for neurosurgical decision support are generated based on the standardized diagnostic dataset.

6. The method as described in claim 1, characterized in that, The preprocessing includes: standardizing the size and grayscale intensity of the brain tumor image and enhancing the image contrast.

7. The method as described in claim 1, characterized in that, The brain tumor image is an ultrasound image.

8. A brain tumor image recognition and diagnostic system for neurosurgery, characterized in that, include: A preprocessing module is used to collect brain tumor images to be diagnosed and to preprocess the brain tumor images. The preprocessing is used to enhance image contrast and standardize image size and image intensity. The filtering module is used to decompose the preprocessed brain tumor image into at least two feature component maps that complement each other in terms of structural and texture information. Then, the first-stage adaptive nonlocal mean filtering based on local statistical characteristics and the second-stage filtering based on structural correlation similarity measurement are performed on each feature component map in sequence. The fusion module is used to fuse and reconstruct the filtered feature component maps to obtain a feature-enhanced brain tumor image. The recognition module is used to input the feature-enhanced brain tumor image into a pre-trained brain tumor recognition model and output the segmentation mask of the tumor region and the corresponding tumor attribute diagnostic information. The report generation module is used to perform edge refinement post-processing on the segmentation mask and, in conjunction with the tumor attribute diagnostic information, generate a neurosurgical auxiliary diagnostic report.

9. A computer device comprising a memory and a processor, the memory storing code, characterized in that, The processor is configured to acquire the code and execute the brain tumor image recognition and diagnosis method for neurosurgery as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the brain tumor image recognition and diagnosis method for neurosurgery as described in any one of claims 1 to 7.

Images