A SPR image optimization processing method based on image segmentation and edge enhancement

Abstract

The application discloses a SPR image optimization processing method based on image segmentation and edge enhancement, comprising the following steps: collecting SPR image data, preprocessing to generate standardized SPR image data; inputting a structure boundary extraction model constructed based on CGNet, generating a structure boundary label graph and aligning with the image; performing gradient amplitude and local entropy mutation detection to generate an artifact guide graph; inputting the standardized image into the detail and context branches of the improved CSDNet respectively to extract edge and semantic feature graphs; inputting a guide perception gate module to perform structure enhancement and artifact suppression fusion to generate a fusion feature graph; inputting a multi-scale detail restoration module to output an edge enhanced image; and performing structure similarity and edge sharpness scoring based on the original image and the enhanced image to generate an optimization result. The application realizes synchronous optimization of SPR image edge enhancement and structure preservation, and significantly improves image sharpness and diagnostic availability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to an SPR image optimization processing method based on image segmentation and edge enhancement. BACKGROUND

[0002] In the process of medical image-guided therapy and radiotherapy path planning, the SPR image, as the projection result of volume data compression, undertakes the functions of auxiliary positioning and anatomical structure reference. In the prior art, the SPR image is usually improved in visual quality by image reconstruction or image enhancement methods, including the use of super-resolution reconstruction network, edge enhancement filtering algorithm and feature fusion model based on attention mechanism. Among them, deep convolutional neural networks such as U-Net, ResNet and DenseNet are often used for structure extraction and resolution enhancement of SPR images to enhance the visibility of organ boundaries and tissue contours.

[0003] The existing image enhancement methods have obvious limitations when processing SPR images: first, the original edges of the SPR image are blurred, the contrast is low, and the structure boundaries and artifact regions overlap with each other. The traditional enhancement algorithm is easy to cause non-structure region mis-enhancement while improving the image clarity, which interferes with subsequent diagnosis and positioning judgment; second, the existing methods lack effective structure perception mechanism and cannot perform differential processing on anatomical regions and artifact regions, resulting in important boundary details being weakened or artifact signals being amplified in the enhancement results; third, the fusion strategy generally uses simple splicing or channel weighting method, without introducing a fusion path control mechanism based on structure label and artifact guide information, resulting in insufficient edge information recovery, lack of image quality controllability and structure discrimination ability.

[0004] In summary, the existing methods have obvious deficiencies in SPR image boundary extraction, structure-guided enhancement and artifact recognition control, and cannot meet the high-precision edge reconstruction and structure preservation requirements. Therefore, how to provide an SPR image optimization processing method based on image segmentation and edge enhancement is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide an SPR image optimization processing method based on image segmentation and edge enhancement. The present application uses CGNet to extract structure boundary label map, combines CSDNet and guide perception gating module to realize deep fusion of structure enhancement and artifact suppression, and has the advantages of high edge clarity, strong structure preservation and less artifact interference.

[0006] According to an SPR image optimization processing method based on image segmentation and edge enhancement according to an embodiment of the present application, the following steps are included:

[0007] The SPR image data is collected, and standardized SPR image data is generated by preprocessing. The standardized SPR image data is input into a structure boundary extraction model constructed based on a CGNet structure to extract an anatomical region boundary response map and generate a structure boundary label map. The structure boundary label map is aligned with the standardized SPR image data in a pixel dimension. Gradient direction amplitude analysis and local entropy mutation detection are performed based on the standardized SPR image data to locate non-structural mutation regions in the image and generate an artifact guide map. The standardized SPR image data is input into a detail branch and a context branch of an improved CSDNet. The detail branch extracts a local edge change feature map, and the context branch extracts a multi-scale semantic context feature map. The detail feature map and the context feature map are output. The detail feature map, the context feature map, the structure boundary label map, and the artifact guide map are input into a guided perception gating module to perform a gating weighted fusion operation based on structure enhancement and artifact suppression to generate a fusion feature map. The fusion feature map is input into a multi-scale detail restoration module containing an up-sampling structure and a skip connection unit to perform edge region reconstruction and feature restoration to output SPR image edge enhancement image data. Structure similarity analysis and edge sharpness scoring are performed based on the standardized SPR image data and the SPR image edge enhancement image data to generate an SPR image optimization result.

[0008] Optionally, the preprocessing process includes performing size normalization, grayscale standardization, and background noise suppression.

[0009] Optionally, the process of inputting the standardized SPR image data into the structure boundary extraction model constructed based on the CGNet structure to extract an anatomical region boundary response map and generate a structure boundary label map includes:

[0010] An initial feature map is generated by using a shallow convolutional backbone network to preliminarily extract features of the standardized SPR image data. Low-level gradient features, texture features, and local contrast features of the image are extracted through two consecutive convolution and normalization operations.

[0011] On the basis of the initial local feature map, a plurality of context guide modules are introduced. Each context guide module includes a local perception channel and a context modeling channel. The local perception channel uses a 3×3 convolution kernel to extract edge direction change and grayscale jump features of a pixel neighborhood. The context modeling channel constructs a context information atlas of a structure region by global pooling and 1×1 convolution to capture internal edge consistency features of a tissue region.

[0012] Dilated convolution with a dilation rate of 2 and 4 is respectively applied to different scale feature maps in the context modeling channel to construct a multi-scale receptive field. Channel attention weighted fusion operations are performed between the local channel and the context channel to generate a context-enhanced feature map.

[0013] The context-enhanced feature map is input into a boundary response prediction unit, and a prediction probability of each pixel point being a boundary point is calculated by using pixel-by-pixel 1*1 convolution under the condition that the image size is kept unchanged, and a boundary response map is output;

[0014] A pixel-by-pixel classification operation is performed on the boundary response map, and pixels higher than a threshold value are classified as boundary pixels and pixels lower than the threshold value are classified as non-boundary pixels according to a set boundary probability threshold value, and a structure boundary label map is formed by combining area, boundary continuity and gradient amplitude distribution judgment;

[0015] The structure boundary label map is resampled by a bilinear interpolation method and mapped to a pixel size and a coordinate system completely consistent with the standardized SPR image data.

[0016] Optionally, the artifact guide map is generated, comprising:

[0017] A gradient direction calculation operation is performed on the standardized SPR image data, and a Sobel operator is used to calculate a gradient amplitude map in a horizontal direction and a vertical direction, respectively, to construct a primary gradient image set containing gradient direction amplitude information;

[0018] The gradient variance and the direction change rate of each pixel point in a local window are calculated in the primary gradient image set, a gradient direction mutation area is identified, and a candidate artifact area response map is generated;

[0019] A local entropy analysis is performed on the standardized SPR image data, a fixed window is used to slide to statistically analyze the local entropy value distribution of each pixel area, an entropy value mutation area is located, and a local entropy mutation map is generated;

[0020] The candidate artifact area response map and the local entropy mutation map are normalized and superimposed pixel by pixel, a joint response map is calculated by a weighted fusion method, a fusion intensity map is generated, and the fusion weight is dynamically adjusted according to the gradient fluctuation and the entropy mutation degree of each area;

[0021] A hierarchical threshold segmentation is performed on the fusion intensity map, an area exceeding a set upper threshold value is marked as a high-intensity artifact area, and an area lower than a set lower threshold value is marked as a background area, an initial artifact guide map is constructed, a boundary closing and noise area filtering operation is performed on the initial artifact guide map, and a final artifact guide map is generated, which is one-to-one corresponding to the standardized SPR image data in the pixel dimension.

[0022] Optionally, the improved CSDNet structure includes a detail branch and a context branch, and has the following characteristics:

[0023] The detail branch is composed of three consecutive layers of depth separable convolution modules, each layer contains a standard 3*3 depth convolution, a point-by-point convolution and a batch normalization operation, and outputs a detail feature map.

[0024] The context branch is composed of four-layer pyramid semantic encoding units, each layer containing a convolution kernel size of 5*5 and a channel attention enhancement module, and a multi-scale context expansion receptive field is realized by adopting a hollow structure with a hollow rate of 2, 4, 6, and 8;

[0025] A cross-layer feature residual connection structure is adopted between the outputs of each level of the context branch to enhance the semantic consistency expression ability, and a deep context perception map is formed by fusing the semantic feature maps of the current layer and the previous layer;

[0026] The detail branch and the context branch respectively output a detail feature map and a context feature map at the end, and the two are aligned in feature dimension through linear transformation of unified channel number, and are respectively input into a guided perception gating module.

[0027] Optionally, the guided perception gating module has the following structure and operation process:

[0028] The detail feature map, the context feature map, the structure boundary label map and the artifact guide map are received as input data, and channel dimension expansion and spatial dimension standardization processing are respectively performed to construct an input feature set with unified size and channel number;

[0029] The guided perception gating module includes a structure enhancement gating unit and an artifact suppression gating unit. The structure enhancement gating unit uses the structure boundary label map as a guide to generate a structure enhancement weight map using a pixel-by-pixel multiplication operation, and applies the structure enhancement weight map to the detail feature map.

[0030] The artifact suppression gating unit uses the artifact guide map as a guide to generate a suppression mask map using inverse normalization and a Sigmoid gating function, and performs a pixel-by-pixel multiplication operation on the suppression mask map and the context feature map to suppress non-structure region feature interference.

[0031] After the structure enhancement weight map and the suppression mask map are applied, the enhanced detail feature map and the suppressed context feature map are subjected to channel dimension weighted fusion, and the fusion weight is dynamically generated according to the structure region confidence map, and a fused feature map is output.

[0032] The fused feature map is superimposed with the shallow feature map of the original standardized SPR image data through a residual connection before being output, so as to maintain the structural consistency of the fusion result and the input image.

[0033] Optionally, the multiscale detail recovery module comprises two up-sampling units and a skip connection unit, each up-sampling unit adopts bilinear interpolation and 3*3 convolution to jointly construct a multiscale feature up-sampled image, the skip connection unit connects the intermediate layer feature map of the CSDNet detail branch and the up-sampled feature map of the corresponding scale, compensates the edge detail information through a channel alignment and element addition fusion mode, and outputs an edge enhanced image data with a resolution consistent with the standardized SPR image data.

[0034] Optionally, the structural similarity analysis comprises: taking the standardized SPR image data as a reference image and the SPR image edge enhanced image data as an image to be evaluated, performing a fixed window sliding calculation operation in a unified resolution space, extracting the brightness mean value, the contrast standard deviation and the structural gradient vector in each window respectively, calculating the brightness similarity, the contrast similarity and the structural similarity components correspondingly, and generating a structural similarity index score through weighted fusion for analysis.

[0035] Optionally, the edge sharpness score takes the SPR image edge enhanced image data as an input image, adopts a multiscale Laplacian edge response function to extract an edge contour response map, constructs a comprehensive score function based on the gradient amplitude distribution of the edge region, the edge length continuity and the structural boundary label map matching degree, and outputs an edge sharpness score result.

[0036] Optionally, the generation of the SPR image optimization result is based on the structural similarity index score map and the edge sharpness score result, a joint score function is used to comprehensively evaluate the structure preservation and edge recognizability of the edge enhanced image data, whether the fusion feature map meets the image quality requirements is judged according to a preset score threshold, when the score result meets the structural consistency requirement and the edge sharpness requirement at the same time, the SPR image optimization result is output, otherwise the structure enhancement weight and the artifact suppression factor in the guided perception gating module are updated in parameters, and the fusion and reconstruction process is re-executed.

[0037] The beneficial effects of the present application are:

[0038] (1) Realize the accurate extraction of anatomical structure boundary: the structural boundary extraction model based on CGNet structure is adopted in the present application, the boundary response map is extracted on the standardized SPR image data and the structural boundary label map is generated, the spatial positioning ability of boundary recognition is effectively enhanced, and the clear structural guidance basis is provided for feature fusion.

[0039] (2) The selectivity and stability of the edge enhancement process are improved: the artifact guide map is introduced, the detail feature map and the context feature map are input into the guide perception gate module, the structure enhancement factor and the artifact suppression factor in the fusion path are dynamically adjusted, the effective area of the edge is strengthened, and the non-structure interference area is suppressed, which can improve the definition and reliability of the enhancement result.

[0040] (3) The quality controllability and structure preservation of the enhanced edge optimization result are improved: the joint quality evaluation mechanism is constructed based on the structure similarity analysis and the edge definition score, the SPR image optimization result is output combined with the score result, the end-to-end closed loop control ability is possessed, and the original structure information is preserved while the edge interpretability is improved in the image enhancement process. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0042] Figure 1 A whole flow chart of an SPR image optimization processing method based on image segmentation and edge enhancement is provided for the application;

[0043] Figure 2 A network structure schematic diagram of a structure boundary extraction model based on CGNet structure construction of an SPR image optimization processing method based on image segmentation and edge enhancement is provided for the application;

[0044] Figure 3 A feature extraction path schematic diagram of a detail branch and a context branch in an improved CSDNet structure of an SPR image optimization processing method based on image segmentation and edge enhancement is provided for the application. DETAILED DESCRIPTION

[0045] The application will now be further described in detail in conjunction with the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0046] REFERENCE Figures 1-3 An SPR image optimization processing method based on image segmentation and edge enhancement, comprising the following steps:

[0047] The SPR image data is collected, and standardized SPR image data is generated by preprocessing. The standardized SPR image data is input into a structure boundary extraction model constructed based on a CGNet structure to extract an anatomical region boundary response map and generate a structure boundary label map. The structure boundary label map is aligned with the standardized SPR image data in a pixel dimension. Gradient direction amplitude analysis and local entropy mutation detection are performed based on the standardized SPR image data to locate non-structural mutation regions in the image and generate an artifact guide map. The standardized SPR image data is input into a detail branch and a context branch of an improved CSDNet, respectively. The detail branch extracts a local edge change feature map, and the context branch extracts a multi-scale semantic context feature map. The detail feature map and the context feature map are output. The detail feature map, the context feature map, the structure boundary label map, and the artifact guide map are input into a guided perception gating module to perform a gated weighted fusion operation based on structure enhancement and artifact suppression to generate a fusion feature map. The fusion feature map is input into a multi-scale detail restoration module containing an up-sampling structure and a skip connection unit to perform edge region reconstruction and feature restoration to output SPR image edge enhancement image data. Structure similarity analysis and edge sharpness scoring are performed based on the standardized SPR image data and the SPR image edge enhancement image data to generate an SPR image optimization result.

[0048] In this embodiment, the preprocessing process includes performing size normalization, grayscale standardization, and background noise suppression.

[0049] In this embodiment, the process of inputting the standardized SPR image data into the structure boundary extraction model constructed based on the CGNet structure to extract an anatomical region boundary response map and generate a structure boundary label map includes:

[0050] The standardized SPR image data is preliminarily feature-extracted by a shallow convolution backbone network. Low-level gradient features, texture features, and local contrast features of the image are extracted through two consecutive convolution and normalization operations to generate an initial local feature map.

[0051] On the basis of the initial local feature map, a plurality of context guide modules are introduced. Each context guide module includes a local perception channel and a context modeling channel. The local perception channel extracts edge direction change and grayscale jump features of a pixel neighborhood by using a 3×3 convolution kernel. The context modeling channel constructs a context information atlas of a structure region by using global pooling and a 1×1 convolution to capture internal edge consistency features of a tissue region.

[0052] The cavitation convolution with the cavitation rate of 2 and 4 is respectively applied to the different scale feature maps in the context modeling channel to construct a multi-scale receptive field, to enhance the perception ability of the structure boundary of different tissue scales, and to perform a channel attention weighting fusion operation between the local channel and the context channel to generate a context enhanced feature map;

[0053] The context enhanced feature map is input into a boundary response prediction unit, a prediction probability of each pixel point being a boundary point is calculated by using a pixel-by-pixel 1*1 convolution under the premise of keeping the image size unchanged, and a boundary response map is output, the prediction probability being a real value between 0 and 1;

[0054] A pixel-by-pixel classification operation is performed on the boundary response map, and pixels higher than a threshold value are classified as boundary pixels and pixels lower than the threshold value are classified as non-boundary pixels according to a set boundary probability threshold, and a structure boundary label map is formed by combining area, boundary continuity and gradient amplitude distribution judgment;

[0055] The structure boundary label map is resampled by a bilinear interpolation method and mapped to a pixel size and a coordinate system completely consistent with the standardized SPR image data, so that the structure boundary label map and the standardized SPR image data are aligned point by point in the pixel dimension.

[0056] In the embodiment, the artifact guide map is generated, including:

[0057] A gradient direction calculation operation is performed on the standardized SPR image data, and a Sobel operator is used to calculate gradient amplitude maps in the horizontal direction and the vertical direction to construct a primary gradient image set containing gradient direction amplitude information;

[0058] The gradient variance and the direction change rate of each pixel point in the local window are calculated in the primary gradient image set, the gradient direction mutation area is identified, and a candidate artifact area response map is generated;

[0059] A local entropy analysis is performed on the standardized SPR image data, a fixed window is used to slide to statistically analyze the local entropy value distribution of each pixel area, the area with a sudden change in entropy value is located, and a local entropy mutation map is generated;

[0060] The candidate artifact region response map and the local entropy mutation map are normalized and superimposed pixel by pixel, a joint response map is calculated by a weighted fusion method, a fusion intensity map is generated, and the fusion weight is dynamically adjusted according to the gradient fluctuation and the entropy mutation degree of each region; specifically, in the process of constructing the artifact guide map, in order to improve the distinguishing ability of the fusion intensity map to the non-structural mutation region, the dynamic weight adjustment mechanism based on the gradient fluctuation and the entropy mutation response value is adopted when the candidate artifact region response map and the local entropy mutation map are fused, for each pixel position, the directional gradient amplitude of the pixel position in the candidate artifact region response map is calculated, the standard deviation of the gradient amplitude in a fixed window centered on the pixel is calculated, and the gradient fluctuation factor is defined to describe the local directional change intensity; the entropy response value of the pixel in the local entropy mutation map is extracted synchronously, which represents the gray level complexity and information mutation degree of the region, and according to the normalized results of the gradient fluctuation factor and the entropy response value, a fusion weight distribution function is constructed, and the two responses are mapped into the fusion weight factor according to the proportional relationship, wherein the gradient dominant region is given a higher gradient channel fusion weight, and the entropy mutation dominant region is given a higher entropy channel fusion weight; the calculation method of the weight factor ensures that the fusion weight dynamically changes in the overall dimension of the image, accurately adapts to the difference in the dependence of different structural regions on artifact judgment features, and finally generates a fusion intensity map through weighted linear combination.

[0061] A hierarchical threshold segmentation is performed on the fusion intensity map, regions exceeding a set upper threshold are marked as high-intensity artifact regions, and regions below a set lower threshold are marked as background regions, an initial artifact guide map is constructed, boundary closure and noise region filtering operations are performed on the initial artifact guide map, and a final artifact guide map is generated, which is one-to-one corresponding to the standardized SPR image data in the pixel dimension.

[0062] In this embodiment, the improved CSDNet structure includes a detail branch and a context branch, and has the following characteristics:

[0063] The detail branch is composed of three consecutive layers of depth separable convolution modules, each layer containing a standard 3x3 depth convolution, a point-by-point convolution and a batch normalization operation, for extracting edge directional change features, low-level texture response features and local gradient mutation region features in the SPR image, and outputting a detail feature map;

[0064] The context branch is composed of four layers of pyramid semantic encoding units, each layer containing a convolution kernel size of 5x5 of a hollow convolution and a channel attention enhancement module, and a hollow structure with a hollow rate of 2, 4, 6 and 8 is adopted to realize multi-scale context expansion receptive field;

[0065] A cross-layer feature residual connection structure is adopted between the outputs of each level of the context branch to enhance the semantic consistency expression ability, and a deep context perception map is formed by fusing the semantic feature maps of the current layer and the previous layer.

[0066] The detail branch and the context branch respectively output a detail feature map and a context feature map at the end, both of which are linearly transformed by a unified channel number to align the feature dimensions, and are respectively input into a guided perception gating module for structure enhancement and artifact suppression fusion operation.

[0067] The improved CSDNet structure adopts a shallow lightweight architecture as a whole, avoids full connection layers, and keeps the spatial resolution of feature maps unchanged throughout the whole process, ensuring that edge details are not compressed or lost during processing. Compared with the original CSDNet model, the structure is optimized in terms of feature extraction path, receptive field construction method and edge preservation ability, which brings the following beneficial effects:

[0068] By introducing a multi-layer depth separable convolution module in the detail branch, the response capability to edge direction mutation, low-contrast texture and small structure disturbance in the SPR image is improved, the edge sensitivity of the feature extraction process is enhanced, and the network parameter quantity and computational complexity are reduced.

[0069] The context branch uses multi-level hole convolution combined with channel attention mechanism to expand the context features, which significantly enhances the multi-scale semantic modeling capability of anatomical structure regions, enabling the model to have stronger semantic consistency and regional continuity expression capability when identifying complex organ structure contours, and effectively suppressing background interference.

[0070] By constructing a cross-layer feature residual connection channel inside the context branch, the information transmission efficiency between different semantic levels is improved, the structure misplacement problem that may occur in deep feature expression is alleviated, and the controllability of context semantic information in spatial positioning accuracy is enhanced.

[0071] The feature maps throughout the whole process maintain the original spatial resolution in the detail path and the context path, avoiding the use of full connection layers and spatial compression structure, which further guarantees the restoration capability of edge details and structure position in the SPR image, and provides higher fidelity feature input for fusion operation.

[0072] In this embodiment, the guided perception gating module has the following structure and operation process:

[0073] The detail feature map, the context feature map, the structure boundary label map and the artifact guide map are received as input data, and channel dimension expansion and spatial dimension standardization processing are respectively performed to construct an input feature set with unified size and channel number;

[0074] The guided perception gating module includes a structure enhancement gating unit and an artifact suppression gating unit. The structure enhancement gating unit uses a structure boundary label map as guidance to generate a structure enhancement weight map using a pixel-wise multiplication operation, and applies the structure enhancement weight map to the detail feature map to enhance the response intensity of the edge region.

[0075] The artifact suppression gating unit uses an artifact guide map as guidance to generate a suppression mask map using inverse normalization and a Sigmoid gating function, and performs a pixel-wise multiplication operation on the suppression mask map and the context feature map to suppress non-structure region feature interference.

[0076] After the structure enhancement weight map and the suppression mask map are applied, the enhanced detail feature map and the suppressed context feature map are subjected to channel dimension weighted fusion, and the fusion weight is dynamically generated according to the structure region confidence map, and the fusion feature map is output.

[0077] The fusion feature map is stacked with the shallow feature map of the original standardized SPR image data through residual connection before output to enhance information stability and maintain the structural consistency of the fusion result and the input image.

[0078] In this embodiment, the multi-scale detail recovery module includes two up-sampling units and a jump connection unit. Each up-sampling unit uses bilinear interpolation and 3x3 convolution to jointly construct a multi-scale feature up-sampled image. The jump connection unit connects the intermediate layer feature map of the CSDNet detail branch and the up-sampled feature map of the corresponding scale, compensates for the edge detail information through channel alignment and element addition fusion, and outputs an edge enhanced image data with the same resolution as the standardized SPR image data.

[0079] In this embodiment, the structure similarity analysis includes: taking the standardized SPR image data as the reference image and the SPR image edge enhanced image data as the to-be-evaluated image, performing a fixed window sliding calculation operation in a unified resolution space, extracting the brightness mean value, contrast standard deviation and structure gradient vector in each window, calculating the brightness similarity, contrast similarity and structure similarity components, and generating a structure similarity index score through weighted fusion for analysis.

[0080] In this embodiment, the edge sharpness score takes the SPR image edge enhanced image data as the input image, adopts a multi-scale Laplacian edge response function to extract an edge contour response map, constructs a comprehensive score function based on the gradient amplitude distribution of the edge region, the edge length continuity, and the matching degree of the structure boundary label map, and outputs the edge sharpness score result. Specifically, in the edge sharpness score process of the SPR image edge enhanced image data, a multi-scale Laplacian edge response function is used to construct an edge contour response map to accurately represent the edge contour intensity and detail continuity at different scales. The multi-scale Laplacian edge response function captures the second-order gray change information at the edge by performing high-pass edge convolution operations at multiple scales on the input image, and maintains the integrity of the edge response corresponding to different structure sizes.

[0081] In specific implementation, first, a scale pyramid sequence is constructed for the SPR image edge enhanced image data, and a Gaussian blur kernel is used to generate low, medium, and high scale version images. On each scale image, a standard Laplacian kernel function is applied:

[0082] ;

[0083] A two-dimensional convolution is performed on the image to extract the edge response map at this scale;

[0084] Subsequently, the edge response maps at different scales are normalized and fused by pixel-by-pixel maximum response to construct the final multi-scale Laplacian edge contour response map. This response map can retain both small edge texture and large-scale structure boundary, and improve the ability to distinguish the integrity of image details in the edge sharpness score;

[0085] Based on the response map, three types of structure indicators of the edge region are further calculated: (1) edge gradient amplitude distribution statistics, used to evaluate the overall level of edge intensity; (2) edge pixel connection length, used to evaluate the edge curve continuity; (3) spatial overlap degree of edge response region and structure boundary label map, used to quantify the ability of edge response to maintain anatomical structure.

[0086] Finally, the above three indicators are input into a weighted score function to output the edge sharpness score value as a quantitative basis for the effectiveness of SPR image enhancement.

[0087] In this embodiment, the generation of the SPR image optimization result is based on the structural similarity index score map and the edge sharpness score result, a joint scoring function is used to comprehensively evaluate the structural consistency and edge recognizability of the edge enhanced image data, and whether the fusion feature map meets the image quality requirements is judged according to a preset scoring threshold. When the scoring result meets the structural consistency requirement and the edge sharpness requirement at the same time, the SPR image optimization result is output, otherwise the structural enhancement weight and the artifact suppression factor in the guided perception gating module are updated in parameters, and the fusion and reconstruction process is re-executed. Specifically as follows:

[0088] First, calculate the structural similarity score , which is derived from the average value of the structural similarity index score map of the standardized SPR image data and the SPR image edge enhanced image data in the region defined by the structure boundary label map, and is used to quantify the fidelity of the anatomical structure;

[0089] Second, calculate the edge sharpness index , a multi-scale Laplacian edge response function is used to extract the edge response map, and the gradient amplitude distribution is counted in the local window in the response map, and the overall edge sharpness level is measured by weighted average of edge intensity of different scales;

[0090] Third, introduce the structure alignment index , which is calculated by the pixel coincidence degree between the structure boundary label map and the edge response map, and is used to evaluate whether the enhanced edge is accurately on the boundary of the anatomical region;

[0091] The above three indexes are input into the joint scoring function:

[0092] ;

[0093] Wherein, the weighted coefficients , , meet the sum of 1, and can be adaptively adjusted according to the image structure complexity and the sparsity of anatomical region;

[0094] According to the preset quality threshold , it is judged whether the scoring function output meets the image optimization standard:

[0095] If , it is determined that the quality of the current edge enhanced image data meets the optimization requirements, and it is output as the SPR image optimization result;

[0096] If , the parameters in the guided perception gating module are updated;

[0097] The feedback update strategy is as follows:

[0098] structure enhancement weight in the structure enhancement path is updated as follows:

[0099]

[0100] artifact suppression factor in the artifact suppression path is updated as follows:

[0101]

[0102] wherein, , is a learning rate coefficient, is a stabilization term to prevent division by zero error, denotes taking the maximum value of ;

[0103] After parameter update, the fusion feature map generation and multi-scale detail reconstruction process are re-executed until the scoring result meets the image optimization requirements, ensuring the overall improvement of the final output image in structure consistency and edge quality.

[0104] Embodiment 1:

[0105] In order to verify the feasibility of the application in implementation, the application is applied to treatment plan verification and real-time positioning in treatment process. Since the edge structure of SPR image itself is fuzzy, low contrast features are obvious, and the conventional enhancement method cannot highlight the key edges while maintaining the accuracy of the structure, which affects the doctor's judgment of organ interface and the accurate positioning of radiotherapy target area. The scheme of the application is based on image segmentation and edge enhancement technology, inputs the standardized SPR image into the small segmentation network based on CGNet, accurately extracts the organ boundary response map and generates the structure boundary label map; generates the artifact guide map through gradient direction amplitude analysis and local entropy mutation detection; constructs the improved CSDNet network, the detail branch and the context branch extract features in parallel, and adjusts the enhancement and suppression weights in the guide perception gate module with the structure label and the artifact map respectively; finally, the image reconstruction is completed through the multi-scale detail restoration module, and the closed-loop evaluation of structure similarity and edge sharpness is relied on the joint scoring function to ensure that the optimization result is clear and real.

[0106] ​​In clinical application, the SPR projection image of the patient is sent into the processing flow of the application by the radiotherapy physicist before the treatment plan is formulated. The edge definition of the processed image is significantly improved, and the organ contour visibility is enhanced, so that the radiation physician can check the treatment positioning accuracy in real time. For 10 different cases of the same part, the conventional super-resolution reconstruction, the edge enhancement method based on U-Net and the method of the application are compared respectively, and all the enhanced images are evaluated by a joint scoring function. The results show that the method of the application is superior to other schemes in terms of structural similarity index and edge definition score.

[0107] To verify the beneficial effects, the improvement amplitudes of the methods in the three indexes of PSNR, SSIM and edge definition index are compared. The average PSNR of the conventional super-resolution method is improved by 2.3 dB, the SSIM is improved by 0.04, and the edge definition index is improved by 12%; the average PSNR of the U-Net method is improved by 3.8 dB, the SSIM is improved by 0.06, and the edge definition index is improved by 18%; the average PSNR of the method of the application is improved by 5.2 dB, the SSIM is improved by 0.09, and the edge definition index is improved by 27%. These data prove that the method of the application can highlight the real structure edge while suppressing artifacts, and provide a reliable image basis for subsequent registration and target positioning.

[0108] In complex tissue regions, such as the lung field junction and the liver-diaphragm junction region, the method of the application effectively retains the small wrinkle and organ edge curvature information by aligning the multi-scale Laplacian edge response and the structure label, while the conventional method often leads to edge discontinuity or misplacement in these regions. In summary of the image segmentation and enhancement effect in each case, the method of the application obtains a good rating rate of more than 90% in the doctor's subjective score, proving that the method has good clinical applicability and promotional value.

[0109] Table 1: Quality comparison of SPR images before and after optimization (based on 10 patients)

[0110] Case No. Original PSNR (dB) Original SSIM Original Edge Index (%) Invention PSNR (dB) Invention SSIM Invention Edge Index (%) Similarity Boost (SSIM Delta) Edge Boost (%) 1 24.6 0.72 45 29.8 0.81 58 0.09 13 2 23.9 0.7 43 29.0 0.79 55 0.09 12 3 25.1 0.74 47 30.3 0.83 61 0.09 14 4 24.3 0.71 44 29.6 0.8 59 0.09 15 5 23.7 0.69 42 28.9 0.78 54 0.09 12 6 24.8 0.73 46 30.1 0.82 60 0.09 14 7 24.2 0.71 44 29.4 0.8 57 0.09 13 8 23.5 0.68 41 28.7 0.77 53 0.09 12 9 25.4 0.75 48 30.6 0.84 62 0.09 14 10 24.0 0.7 43 29.2 0.79 56 0.09 13

[0111] As can be seen from Table 1, the method of the application improves in terms of signal-to-noise ratio, structure fidelity and edge definition. First, the PSNR value after processing is increased by about 5.2 dB on average compared with the original image in all patients, specifically from 23.5 dB-25.4 dB to 28.7 dB-30.6 dB, which indicates that the image noise is significantly reduced and the detail restoration degree is significantly enhanced. Second, the structural similarity index (SSIM) is improved from 0.68-0.75 to 0.77-0.84, with an increment of 0.09 in all cases, which indicates that the boundary continuity and overall structure preservation are reliably guaranteed, and the enhanced image is highly consistent with the standard SPR image in anatomical structure.

[0112] In terms of edge definition, the original 41% - 48% is improved to 53% - 62% by the method of the present application, with an average improvement of about 13%, and the improvement of case 4 and case 9 is particularly prominent, reaching 15% and 14% respectively, fully verifying the enhancement ability of the present application to the tiny edge details in the complex structure area. The SPR image optimization comparison table shows that the method of the present application has high consistency and robustness in different cases and different organ parts, meeting the dual requirements of image quality and structure accuracy in clinic.

[0113] The above merely describes the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes within the technical range disclosed by the present application according to the technical solution and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for SPR image optimization processing based on image segmentation and edge enhancement, characterized in that, Includes the following steps: The process involves: acquiring SPR image data and preprocessing it to generate standardized SPR image data; inputting the standardized SPR image data into a structural boundary extraction model built on the CGNet structure to extract anatomical region boundary response maps and generate structural boundary label maps, which are then aligned with the standardized SPR image data in the pixel dimension; performing gradient direction magnitude analysis and local entropy mutation detection based on the standardized SPR image data to locate non-structural mutation regions in the image and generate artifact guidance maps; inputting the standardized SPR image data into the detail branch and context branch of an improved CSDNet, respectively, where the detail branch extracts local edge change feature maps and the context branch extracts multi-scale semantic context feature maps, outputting detail feature maps and context feature maps; inputting the detail feature maps, context feature maps, structural boundary label maps, and artifact guidance maps into a guidance perception gating module to perform a gated weighted fusion operation based on structural enhancement and artifact suppression to generate a fused feature map; the guidance perception gating module has the following structure and operation flow: Receive detail feature maps, context feature maps, structural boundary label maps, and artifact guide maps as input data, and perform channel dimension expansion and spatial dimension standardization processing respectively to construct an input feature set with a uniform size and number of channels; The guided perception gating module includes a structure enhancement gating unit and an artifact suppression gating unit. The structure enhancement gating unit uses the structure boundary label map as a guide and generates a structure enhancement weight map using pixel-wise multiplication operations, and applies the structure enhancement weight map to the detail feature map. The artifact suppression gating unit uses the artifact guide map as a guide, and uses inverse normalization and the Sigmoid gating function to generate a suppression mask map. The suppression mask map is then multiplied pixel by pixel with the context feature map to suppress feature interference from unstructured regions. After the structural enhancement weight map and the suppression mask map are applied, channel-dimensional weighted fusion is performed on the enhanced detail feature map and the suppression context feature map. The fusion weights are dynamically generated based on the structural region confidence map, and the fused feature map is output. Before output, the fused feature map is superimposed with a shallow feature map of the original normalized SPR image data through residual connections to maintain the consistency of the fusion result with the input image structure. The fused feature map is then input into a multi-scale detail recovery module containing upsampling structures and skip connection units to perform edge region reconstruction and feature restoration, outputting SPR image edge enhancement image data. Based on the normalized SPR image data and the SPR image edge enhancement image data, structural similarity analysis and edge sharpness scoring are performed to generate SPR image optimization results. The SPR image optimization result is generated based on the structural similarity index score map and the edge sharpness score map. A joint scoring function is used to comprehensively evaluate the structural preservation and edge recognizability of the edge-enhanced image data. A preset scoring threshold is used to determine whether the fused feature map meets the image quality requirements. When the scoring result simultaneously meets the structural consistency and edge sharpness requirements, the SPR image optimization result is output; otherwise, the parameters of the structural enhancement weights and artifact suppression factors in the guided perception gating module are updated, and the fusion and reconstruction process is re-executed, as follows:

1. Calculate the structural similarity score , The value is derived from the average of the structural similarity index score maps of the standardized SPR image data and the SPR image edge enhancement image data within the region defined by the structural boundary label map, and is used to quantify the fidelity of the anatomical structure; Second, calculate the edge sharpness index. The edge response map is extracted using the multi-scale Laplacian edge response function, and the gradient magnitude distribution is statistically analyzed through a local window in the edge response map. The edge intensity at different scales is weighted and averaged to measure the overall edge sharpness level. Third, introduce structural alignment index. The pixel overlap between the structural boundary label map and the edge response map is used to evaluate whether the enhanced edge accurately falls on the boundary of the anatomical region. Input the above three indicators into the joint scoring function: ； Among them, the weighting coefficient , , The summation is 1, and it can be adaptively adjusted according to the complexity of the image structure and the sparsity of the anatomical region. Based on preset quality thresholds Determine whether the output of the scoring function meets the image optimization criteria: like If the current edge enhancement image data quality meets the optimization requirements, the result is output as the SPR image optimization result. like Then the feedback-driven mechanism will update the parameters in the perception gating module; The feedback update strategy is as follows: Structural enhancement weights in structural enhancement paths Update using the following formula: ； Artifact suppression factor in the artifact suppression pathway Update using the following formula: ； in, , This is the learning rate coefficient. To prevent division by zero error, Indicates taking The maximum value; After the parameters are updated, the fusion feature map generation and multi-scale detail reconstruction process is re-executed until the scoring results meet the image optimization requirements, ensuring a comprehensive improvement in structural consistency and edge quality in the final output image.

2. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 1, characterized in that, The preprocessing process includes performing size normalization, grayscale normalization, and background noise suppression.

3. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 2, characterized in that, The process of inputting standardized SPR image data into a structural boundary extraction model based on the CGNet structure to extract anatomical region boundary response maps and generate structural boundary label maps includes: A shallow convolutional backbone network is used to perform preliminary feature extraction on standardized SPR image data. Low-level gradient features, texture features and local contrast features of the image are extracted through two consecutive convolutional and normalization operations to generate an initial local feature map. Based on the initial local feature map, several context guidance modules are introduced. Each context guidance module contains a local perception channel and a context modeling channel. The local perception channel uses a 3×3 convolution kernel to extract the edge direction change and gray-level jump features of the pixel neighborhood. The context modeling channel constructs the context information map of the structural region through global pooling and 1×1 convolution to capture the edge consistency features inside the organization region. Dilated convolutions with dilation rates of 2 and 4 are applied to feature maps at different scales within the context modeling channel to construct a multi-scale receptive field. Channel attention weighted fusion is then performed between the local channel and the context channel to generate context-enhanced feature maps. The context-enhanced feature map is input into the boundary response prediction unit. While keeping the image size unchanged, the prediction probability of each pixel being a boundary point is calculated by pixel-wise 1×1 convolution, and the boundary response map is output. A pixel-by-pixel classification operation is performed on the boundary response map. Pixels above the set boundary probability threshold are classified as boundary pixels, and pixels below the threshold are classified as non-boundary pixels. The structural boundary label map is formed by combining the region area, boundary continuity and gradient magnitude distribution. The structural boundary label map is resampled using bilinear interpolation and mapped to pixel size and coordinate system that are completely consistent with the standardized SPR image data.

4. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 3, characterized in that, The generation of the artifact guide map includes: Gradient direction calculation is performed on the standardized SPR image data. The Sobel operator is used to calculate the gradient magnitude maps in the horizontal and vertical directions respectively, and a primary gradient image set containing gradient direction magnitude information is constructed. In the primary gradient image set, calculate the gradient variance and direction change rate of each pixel within the local window, identify regions with abrupt gradient direction changes, and generate response maps of candidate artifact regions. Local entropy analysis is performed on standardized SPR image data. The distribution of local entropy values ​​in each pixel region is statistically analyzed by sliding a fixed window to locate regions of abrupt entropy changes and generate a local entropy change map. The candidate artifact region response map and the local entropy mutation map are normalized pixel by pixel and superimposed. The joint response map is calculated by weighted fusion to generate a fusion intensity map. The fusion weight is dynamically adjusted according to the gradient fluctuation and entropy mutation degree of each region. A hierarchical threshold segmentation is performed on the fused intensity map. Regions exceeding the set upper threshold are marked as high-intensity artifact regions, and regions below the set lower threshold are marked as background regions. An initial artifact guide map is constructed, and boundary closure and noise region filtering operations are performed on the initial artifact guide map to generate the final artifact guide map, which maintains a one-to-one correspondence with the standardized SPR image data in the pixel dimension.

5. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 4, characterized in that, The improved CSDNet structure includes detail branches and context branches, and has the following characteristics: The detail branch consists of three consecutive depthwise separable convolutional modules. Each layer contains a standard 3×3 depthwise convolution, a pointwise convolution, and a batch normalization operation, which outputs a detail feature map. The context branch consists of four layers of pyramid-shaped semantic coding units. Each layer contains a 5×5 dilated convolution kernel and a channel attention enhancement module. The dilated structure with dilation rates of 2, 4, 6 and 8 is used to achieve multi-scale context expansion of the receptive field. A cross-layer feature residual connection structure is adopted between the outputs of each level of the context branch to enhance the semantic consistency expression capability. By fusing the semantic feature maps of the current layer and the previous layer, a deep context-aware map is formed. The detail branch and the context branch output detail feature maps and context feature maps respectively at the end. The feature dimensions are aligned by a linear transformation with a unified number of channels, and then input into the guided perception gating module.

6. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 5, characterized in that, The multi-scale detail restoration module includes two upsampling units and one skip connection unit. Each upsampling unit uses bilinear interpolation and 3×3 convolution to jointly construct a multi-scale feature upsampled image. The skip connection unit connects the intermediate layer feature map of the CSDNet detail branch with the upsampled feature map of the corresponding scale. It compensates for edge detail information through channel alignment and element-wise addition fusion, and outputs edge-enhanced image data with the same resolution as the standardized SPR image data.

7. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 6, characterized in that, The structural similarity analysis includes: using standardized SPR image data as the reference image and SPR image edge enhancement image data as the image to be evaluated, performing a fixed window sliding calculation operation in a uniform resolution space, extracting the mean brightness, standard deviation of contrast, and structural gradient vector within each window, calculating the brightness similarity, contrast similarity, and structural similarity components accordingly, and generating a structural similarity index score through weighted fusion for analysis.

8. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 7, characterized in that, The edge sharpness score uses SPR image edge enhancement image data as input image, extracts edge contour response map using multi-scale Laplacian edge response function, constructs comprehensive scoring function based on gradient magnitude distribution of edge region, edge length continuity and structural boundary label map matching degree, and outputs edge sharpness score result.

9. The SPR image optimization processing method based on image segmentation and edge enhancement according to claim 8, characterized in that, The SPR image optimization result is generated based on the structural similarity index score map and the edge sharpness score map. A joint scoring function is used to comprehensively evaluate the structural preservation and edge recognizability of the edge-enhanced image data. Based on the preset scoring threshold, it is determined whether the fused feature map meets the image quality requirements. When the scoring result meets both the structural consistency requirement and the edge sharpness requirement, the SPR image optimization result is output. Otherwise, the parameters of the structural enhancement weight and artifact suppression factor in the guided perception gating module are updated, and the fusion and reconstruction process is re-executed.

Images