An acute abdomen pre-examination auxiliary optimization method and device based on CRNet

Abstract

The application provides an acute abdomen pre-examination auxiliary optimization method and device based on CRNet, relates to the field of image processing, and solves the problem that in the prior art, acute abdomen related features in an abdominal plain film X-ray image are difficult to be effectively identified. The method specifically comprises the following steps: acquiring an abdominal plain film X-ray image and a feature template library; preprocessing the abdominal plain film X-ray image to obtain a standardized image; performing spatial domain decomposition through pooling separation to obtain a global smooth component feature map and a local detail component feature map; calculating a similarity based on the local detail component feature map and the feature template library to generate a weight map; strengthening the local detail component feature map based on the weight map to obtain a local detail enhanced feature map, inputting the local detail enhanced feature map into an improved Multi-Branch Block to perform multi-scale feature fusion to obtain a fused feature map, and inputting the fused feature map into a lightweight convolution enhancement module to perform feature strengthening to obtain a feature strengthened map; and performing post-processing on the feature strengthened map to output an auxiliary optimization result. The application is used for acute abdomen pre-examination auxiliary.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and apparatus for assisting in the pre-detection of acute abdominal pain based on CRNet. Background Technology

[0002] Improving the efficiency of pre-screening analysis for acute abdomen is crucial for optimizing medical processes. Currently, clinical practice typically relies on abdominal plain X-ray images for pre-screening analysis of acute abdomen, using observation of anatomical structures and imaging features related to potential abnormalities to aid in diagnosis. However, the internal structure of the human abdomen is complex, and key imaging clues related to acute abdomen are often subtle features such as tiny free gases in the intestinal wall, slightly blurred organ outlines, and weak density differences between tissues. These features are easily obscured by background information or ignored manually in low-contrast raw X-ray images, which not only increases the difficulty of image analysis but may also lead to the omission of key imaging information, affecting the efficiency of pre-screening analysis. Therefore, existing technologies have the problem of difficulty in effectively identifying acute abdomen-related features in abdominal plain X-ray images. Summary of the Invention

[0003] This application provides a CRNet-based method and device for assisting in the pre-detection of acute abdomen, which solves the problem that the relevant features of acute abdomen in abdominal plain X-ray images are difficult to be effectively identified in the prior art.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a pre-detection optimization method for acute abdomen based on CRNet is provided, comprising: acquiring abdominal plain X-ray images and a feature template library; the feature template library is a collection of abdominal plain X-ray images of patients with acute abdomen containing feature annotations; the feature annotations include feature regions such as small free gas in the intestinal wall, gas-fluid levels, intestinal dilation, blurred organ contours, and differences in tissue density; preprocessing the abdominal plain X-ray images to obtain standardized images; the preprocessing includes noise suppression, grayscale correction, and standardization; spatially decomposing the standardized images using pooling to obtain global smooth component feature maps and local detail component feature maps; calculating the similarity between the local detail component feature maps and the feature template library to generate a weight map; enhancing the local detail component feature maps based on the weight map to obtain local detail enhancement feature maps; inputting the local detail enhancement feature maps into an improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map; and performing improved Multi-Branch... The block includes a first branch, a second branch, and a fusion unit; the fused feature map is input into the lightweight convolutional enhancement module for feature enhancement to obtain the enhanced feature map; the lightweight convolutional enhancement module includes depthwise separable convolution and ConvFFN; the enhanced feature map is post-processed to output the auxiliary optimization result; the auxiliary optimization result includes the original image, the enhanced image, and the feature annotation map.

[0005] In conjunction with the first aspect mentioned above, in one possible implementation, the abdominal plain X-ray image is preprocessed to obtain a standardized image, including: dynamically adjusting the filtering window according to the pixel neighborhood noise density, adaptively filtering the abdominal plain X-ray image to suppress noise; segmenting the abdominal plain X-ray image into sub-blocks based on the CLAHE algorithm and limiting the contrast gain to perform grayscale correction; and performing pixel value normalization processing and adjusting the image size of the grayscale-corrected abdominal plain X-ray image to obtain a standardized image.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, the standardized image is spatially decomposed by pooling to obtain a global smooth component feature map and a local detail component feature map. This includes: downsampling the standardized image using average pooling and max pooling to obtain a global smooth component feature map; upsampling the global smooth component feature map based on bilinear interpolation to obtain an upsampled feature map; and subtracting the upsampled feature map from the standardized image to obtain local detail component features.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, a weighted map is generated by calculating the similarity between the local detail component feature map and the feature template library, including: extracting local detail feature vectors of the feature-annotated regions in the feature template library; flattening the pixel features of the local detail component feature map to obtain the feature vectors corresponding to each pixel; calculating the cosine similarity between the feature vector of each pixel and the local detail feature vector; and fusing all cosine similarity results to generate a weighted map covering the local detail component feature map.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the local detail enhancement feature map is input into an improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map. This includes: performing depthwise separable convolution, local response normalization, and channel adjustment processing on the local detail enhancement feature map based on the first branch to obtain a density difference feature map; performing dilated convolution, edge-aware residual processing, and edge purification on the local detail enhancement feature map based on the second branch to obtain a contour detail feature map; concatenating the density difference feature map, contour detail feature map, and upsampled feature map through a fusion unit; and adjusting the weights of the concatenated feature map through a channel attention module, followed by convolutional dimensionality reduction processing to obtain the fused feature map.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the fused feature map is input into a lightweight convolutional enhancement module for feature enhancement to obtain an enhanced feature map. This includes: performing a 1×1 convolution operation on the fused feature map; performing a 5×5 depthwise separable convolution on the feature map after the 1×1 convolution; inputting the output of the depthwise separable convolution into a ConvFFN, and sequentially performing dimensionality-up convolution, depthwise convolution, and dimensionality-down convolution; the channel expansion ratio of the ConvFFN is 1:4; and performing a residual connection between the dimensionality-down convolutioned feature map and the 1×1 convolutioned feature map to output the enhanced feature map.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the feature enhancement map is post-processed to output auxiliary optimization results, including: mapping the feature enhancement map to the grayscale range of the abdominal plain X-ray image; using morphological opening operations to remove minor artifacts in the grayscale-mapped feature enhancement map to obtain an enhanced image; based on the density difference feature map, using adaptive threshold segmentation to label the small free gas regions of the intestinal wall, using threshold filtering to label the tissue density difference regions, and identifying and labeling stepped gas-liquid planes; based on the contour detail feature map, using edge detection to label abnormal organ contour regions, and identifying and labeling intestinal dilatation regions based on intestinal morphology features to generate a feature annotation map.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the upsampled feature map has the same pixel size as the normalized image, density difference feature map, and contour detail feature map.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, after post-processing the feature enhancement map and outputting the auxiliary optimization result, the method further includes: incorporating the feature annotation map into the feature template library.

[0013] The above scheme, through spatial domain decomposition combined with a weighted graph enhancement mechanism guided by a feature template library, can specifically highlight the specific signs of acute abdominal pain, such as gas-liquid levels and free gas. Combined with an improved multi-branch fusion and lightweight enhancement module, it meets the real-time requirements while ensuring the accuracy of feature extraction.

[0014] Secondly, a CRNet-based pre-detection and optimization device for acute abdomen is provided, comprising: a communication unit and a processing unit; the communication unit is used to acquire abdominal plain X-ray images and a feature template library; the feature template library is a collection of abdominal plain X-ray images of patients with acute abdomen containing feature annotations; the feature annotations include feature regions such as intestinal wall micro-free gas, gas-fluid levels, intestinal dilation, blurred organ contours, and tissue density differences; the processing unit is used to preprocess the abdominal plain X-ray images to obtain standardized images; the preprocessing includes noise suppression, grayscale correction, and standardization; the standardized images are spatially decomposed using pooling to obtain global smooth component feature maps and local detail component feature maps; the similarity between the local detail component feature maps and the feature template library is calculated to generate a weight map; the local detail component feature maps are enhanced based on the weight map to obtain local detail enhancement feature maps; the local detail enhancement feature maps are input into an improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map; the improved Multi-Branch... The block includes a first branch, a second branch, and a fusion unit; the fused feature map is input into the lightweight convolutional enhancement module for feature enhancement to obtain the enhanced feature map; the lightweight convolutional enhancement module includes depthwise separable convolution and ConvFFN; the enhanced feature map is post-processed to output the auxiliary optimization result; the auxiliary optimization result includes the original image, the enhanced image, and the feature annotation map.

[0015] Thirdly, this application provides a CRNet-based acute abdomen pre-detection assistance optimization device, comprising: a processor and a storage medium; the storage medium includes instructions, and the processor is used to execute the instructions to implement the method described in the first aspect and any possible implementation thereof. This CRNet-based acute abdomen pre-detection assistance optimization device can be an electronic device or a chip within an electronic device.

[0016] Fourthly, this application provides a CRNet-based pre-detection and optimization system for acute abdomen, comprising: an image acquisition module for acquiring abdominal plain X-ray images from a PACS system; a feature template library storage module for storing a set of abdominal plain X-ray images of patients with acute abdomen containing feature annotations; a CRNet inference module for processing the abdominal plain X-ray images and outputting the optimization results; and an early warning push module for triggering a priority early warning in the PACS system when the optimization results identify high-risk features.

[0017] Fifthly, this application provides a computer-readable storage medium storing instructions that, when executed on a CRNet-based acute abdominal pre-detection and optimization device, cause the CRNet-based acute abdominal pre-detection and optimization device to perform the methods described in the first aspect and any possible implementation thereof.

[0018] In a sixth aspect, this application provides a computer program product containing instructions that, when the computer program product is run on a CRNet-based acute abdominal pre-detection and optimization device, cause the CRNet-based acute abdominal pre-detection and optimization device to perform the methods described in the first aspect and any possible implementation thereof.

[0019] This application provides a CRNet-based pre-detection assistance optimization method and device for acute abdomen. First, a cosine similarity weight map guided by a feature template library provides targeted prior knowledge, solving the problem of insufficient sensitivity of purely data-driven models to minute lesions. Second, the improved Multi-Branch Block, through a dual-branch differential design, accurately captures two complementary features: air-fluid levels, density differences and contour abnormalities, and intestinal dilatation, avoiding information omissions from single feature extraction paths. Finally, the lightweight convolutional enhancement module design ensures real-time processing capabilities in emergency environments. All three are indispensable: the template library provides targeted prior knowledge, the dual branches capture complementary features of air-fluid levels and contour abnormalities, and the lightweight design ensures real-time performance. Together, they constitute a complete pre-detection assistance system that understands medical priors, can distinguish subtle signs, and responds quickly, effectively solving the problem in existing technologies where acute abdomen-related features are difficult to effectively identify and cannot meet clinical timeliness requirements.

[0020] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a CRNet-based pre-detection and optimization method for acute abdominal pain, provided in an embodiment of this application; Figure 2 A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 3A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 4 This is a schematic diagram of a frequency separation process provided in an embodiment of this application; Figure 5 A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 6 A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 7 A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 8 A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 9 A flowchart illustrating another CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment; Figure 10 This is a schematic diagram of a CRNet-based pre-detection and optimization device for acute abdominal pain, provided in an embodiment of this application. Detailed Implementation

[0022] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0023] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] To address the problem of difficulty in effectively identifying acute abdomen-related features in abdominal plain X-ray images in existing technologies, this application provides a CRNet-based pre-detection optimization method for acute abdomen. The method includes: acquiring abdominal plain X-ray images and a feature template library; preprocessing the abdominal plain X-ray images to obtain standardized images; performing spatial domain decomposition on the standardized images using pooling to obtain global smooth component feature maps and local detail component feature maps; calculating the similarity between the local detail component feature maps and the feature template library to generate a weight map; enhancing the local detail component feature maps based on the weight map to obtain enhanced local detail feature maps; inputting the enhanced local detail feature maps into an improved Multi-BranchBlock for multi-scale feature fusion to obtain a fused feature map; inputting the fused feature map into a lightweight convolutional enhancement module for feature enhancement to obtain an enhanced feature map; and post-processing the enhanced feature map to output the optimization result. Based on this, through targeted frequency separation, feature enhancement and multi-branch fusion design, it can accurately capture subtle imaging clues such as tiny free gases in the intestinal wall and blurred organ outlines, effectively eliminating interference from irrelevant information. At the same time, the lightweight module design is adapted to the deployment needs of clinical terminals, and the final intuitive auxiliary results can significantly improve the convenience and effectiveness of image analysis.

[0025] like Figure 1 As shown in the embodiments of this application, the CRNet-based pre-detection and optimization method for acute abdomen includes: S101. Obtain abdominal plain X-ray images and feature template library.

[0026] The feature template library is a collection of plain X-ray images of the abdomen of patients with acute abdominal pain containing feature annotations. The feature annotations include characteristic regions such as tiny free gas in the intestinal wall, gas-fluid levels, intestinal dilation, blurred organ outlines, and differences in tissue density.

[0027] In this embodiment of the application, the abdominal plain X-ray image to be analyzed is read through the hospital PACS system interface or local storage medium, and a pre-constructed feature template library is loaded at the same time. This template library compiles images of typical acute abdominal cases, covering specific signs of various pathological types such as gastrointestinal perforation and intestinal obstruction.

[0028] It should be noted that the images in the feature template library must undergo compliance review to ensure that the data is used in accordance with medical data security regulations.

[0029] As an example, raw images in DICOM format are obtained through the hospital's PACS system interface, and the feature template library is stored in a local storage unit or a cloud server with authorized access.

[0030] Based on the above steps, the basic data required for subsequent processing is obtained, providing data support for subsequent feature extraction and optimization.

[0031] S102. Preprocess the abdominal plain X-ray image to obtain a standardized image.

[0032] Preprocessing includes noise suppression, grayscale correction, and standardization.

[0033] In this embodiment, the original abdominal plain X-ray image is first subjected to a filtering algorithm adapted to the X-ray image to reduce noise interference, then the grayscale adjustment method is used to improve the uniformity of the image grayscale distribution, and finally the pixel value is normalized and the size is unified to obtain a standardized image.

[0034] As an example, the filtering intensity is increased for raw images with low signal-to-noise ratios to ensure a balance between noise suppression and detail preservation.

[0035] Based on the above steps, a standardized image with eliminated redundancy interference and uniform format is obtained, effectively reducing interference caused by image noise and format differences.

[0036] S103. Spatial domain decomposition of the standardized image is performed by pooling to obtain global smooth component feature map and local detail component feature map.

[0037] Among them, the local detail component feature map reflects the image detail information (such as tiny gases and edge contours), while the global smoothness component feature map reflects the overall image structure information.

[0038] In this embodiment, a global smooth component feature map is extracted by a combination of average pooling and max pooling downsampling. This component preserves the overall outline of the abdominal organs, the bone projection, and the large-scale tissue density distribution. Subsequently, the size is restored by interpolation upsampling, and residual subtraction is performed with the original standardized image to accurately extract the local detail component feature map containing weak signals such as free gas in the intestinal wall and blurred organ edges, thus achieving effective decoupling of structure and detail in the spatial domain.

[0039] Based on the above steps, effective separation of image features can be achieved, which can effectively distinguish between the detailed features related to acute abdomen and the background structure information.

[0040] S104. Calculate the similarity between the local detail component feature map and the feature template library, and generate a weight map.

[0041] The weight map is a two-dimensional matrix that matches the size of the local detail component feature map, and is used to characterize the degree of association between each pixel and the features related to acute abdomen.

[0042] In this embodiment, local detail feature vectors of each labeled region in the feature template library are extracted as prior anchor points. At the same time, the local detail component feature map of the current image is flattened into a pixel-level feature vector sequence. The cosine similarity between the two is calculated one by one, and all similarity results are fused and mapped to a weight distribution between 0 and 1, so that regions that are highly similar to known lesions such as gas-liquid levels and free gas receive high weights, while irrelevant background regions receive low weights.

[0043] It should be noted that the similarity calculation method can be selected according to the feature type, and is not limited to calculating cosine similarity, to ensure matching accuracy.

[0044] Based on the above steps, a weighted map that highlights the characteristics of acute abdominal pain is obtained, and pixel regions that are highly correlated with the characteristics of acute abdominal pain are accurately located.

[0045] S105. Enhance the local detail component feature map based on the weight map to obtain the local detail enhanced feature map.

[0046] Among them, enhancement refers to using a weighted graph to perform element-wise weighted modulation on the local detail component feature map, so as to achieve adaptive amplification of target features and suppression of non-target regions.

[0047] In this embodiment, the weight map and the local detail component feature map are multiplied element-wise to linearly enhance the weak lesion signal in the high-weight region, while the background noise in the low-weight region is further suppressed, thereby obtaining a local detail enhancement feature map with a significantly improved signal-to-noise ratio.

[0048] It should be noted that the enhancement process is not a simple global contrast stretching, but a spatially selective enhancement based on prior knowledge. This mechanism ensures that only details that conform to the characteristic pattern of acute abdomen are amplified, avoiding the drawback of traditional enhancement methods that amplify noise at the same time.

[0049] As an example, after enhancement, tiny free gas spots whose grayscale values ​​were originally only slightly higher than the background had their response values ​​increased to a significantly discernible level, while non-specific noise of the same grayscale level was effectively filtered out due to its extremely low weight.

[0050] Based on the above steps, the recognition of target features such as tiny free gases in the intestinal wall and abnormal organ contours is significantly improved, preventing key clues from being obscured by background information.

[0051] S106. Input the local detail enhancement feature map into the improved Multi-Branch Block for multi-scale feature fusion to obtain the fused feature map.

[0052] The improved Multi-Branch Block includes a first branch, a second branch, and a fusion unit, which are used to achieve differentiated extraction and integration of different types of features.

[0053] In this embodiment, the local detail enhancement feature map is input into the improved Multi-Branch Block, and corresponding features are extracted through different branches. Then, the features are spliced ​​together by the fusion unit, and the feature weights are adjusted by the attention mechanism before dimensionality reduction is performed to obtain the fused feature map.

[0054] It should be noted that the network structure of each branch can be designed differently according to the target feature type to ensure the targeted nature of feature extraction.

[0055] As an example, the first branch focuses on extracting density-related features, the second branch focuses on extracting contour-related features, and the fusion unit integrates features by channel splicing.

[0056] Based on the above steps, the problem of information omission in a single feature extraction path is overcome, and deep fusion and complementary enhancement of global structure and local details, density information and contour information are achieved.

[0057] S107. Input the fused feature map into the lightweight convolutional enhancement module for feature enhancement and obtain the enhanced feature map.

[0058] The lightweight convolution enhancement module is based on depthwise separable convolution and ConvFFN, balancing feature enhancement effect and computational efficiency.

[0059] In this embodiment, the fused feature map is input into the lightweight convolution enhancement module. First, the channel dimension is adjusted through convolution operation, then multi-scale features are extracted through lightweight depthwise separable convolution, and finally the feature expression is further enhanced through ConvFFN feature transformation and residual connection to obtain the feature enhancement map.

[0060] As an example, depthwise separable convolutions with small kernel sizes are used, and ConvFFN sets a reasonable channel expansion ratio to reduce hardware resource consumption.

[0061] Based on the above steps, a feature enhancement map with richer feature representation and less redundant information is obtained, thereby reducing interference from redundant information and improving feature discriminativeness.

[0062] S108. Perform post-processing on the feature enhancement map and output auxiliary optimization results.

[0063] The auxiliary optimization results include the original abdominal plain X-ray image, enhanced image, and feature annotation map, which are used to assist in the pre-detection analysis.

[0064] In this embodiment, the enhanced image is obtained by performing grayscale mapping and artifact removal on the feature enhancement map. Then, based on the feature maps output by the first and second branches in S106, a feature annotation map is generated by thresholding or edge detection. Finally, the three types of images are integrated to output the auxiliary optimization result.

[0065] Based on the above steps, auxiliary optimization results that can intuitively present key information are output, improving the targeting of image analysis.

[0066] Based on the above technical solution, the key subtle features related to acute abdomen in abdominal plain X-ray images are effectively highlighted. The output of multiple types of auxiliary images not only conforms to clinical image reading habits, but also reduces the difficulty of image analysis, taking into account both processing efficiency and practicality. This solves the problem that the features related to acute abdomen in abdominal plain X-ray images are difficult to identify effectively in existing technologies.

[0067] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above S102 can be specifically implemented through the following S201, S202 and S203, which are explained in detail below: S201. The filtering window is dynamically adjusted according to the noise density of the pixel neighborhood to perform adaptive filtering on the abdominal plain X-ray image and suppress noise.

[0068] Among them, pixel neighborhood noise density refers to the proportion of noisy pixels within a preset range centered on the target pixel.

[0069] In this embodiment, each pixel (x, y) of the abdominal plain X-ray image I is traversed, and the noise density within its 3×3 neighborhood is calculated. (N_noise is the number of noisy pixels in the neighborhood), then according to Dynamically adjust the filter window size k to satisfy ( =3 is the base window. =4 is the maximum adjustment value), and finally, noise suppression is performed on pixel (x,y) using an adaptive median filtering algorithm.

[0070] It should be noted that the window size k should be limited to the range of 3×3 to 7×7 to avoid excessive adjustment that could lead to blurred image details.

[0071] As an example, when ρ(x,y)=0.2, k=3+round(0.2×4)=4, and an adaptive filtering is performed using a 4×4 window.

[0072] Based on the above steps, random noise in abdominal plain X-ray images can be suppressed.

[0073] S202. Based on the CLAHE algorithm, the abdominal plain X-ray image is segmented into sub-blocks and the contrast gain is limited for grayscale correction.

[0074] Among them, CLAHE refers to a contrast-limited adaptive histogram equalization algorithm used to improve local contrast in images.

[0075] In this embodiment of the application, the noise-suppressed image The image is divided into M×N sub-blocks. A gray-level histogram H is calculated for each sub-block. A contrast gain threshold T = 0.02×S (where S is the total number of pixels in the sub-block) is set. When H(i) > T, the excess is evenly distributed to other gray levels to obtain a corrected histogram H'. The sub-block is then subjected to equalization processing using H'. Finally, all sub-blocks are stitched together to obtain the gray-level corrected image. .

[0076] As an example, the sub-block size is set to 8×8, S=64, T=1.28, and the grayscale frequencies exceeding the threshold are evenly distributed to the other 255 grayscale levels.

[0077] Based on the above steps, the local contrast of the image is effectively improved, making subtle features easier to reveal.

[0078] S203. Perform pixel value normalization processing on the grayscale-corrected abdominal plain X-ray image and adjust the image size to obtain a standardized image.

[0079] In this embodiment of the application, the image after grayscale correction Pixel value normalization is performed using the following formula: =( (x,y)- ) / ( - ),in , They are respectively The minimum and maximum pixel values ​​are used to map the pixel values ​​to the [0,1] interval, and then a bilinear interpolation algorithm is used to... Adjust to the preset size (height H × width W × number of channels C) to obtain a standardized image. .

[0080] As an example, with the preset size set to 512×512, the bilinear interpolation algorithm calculates the target pixel value by weighted average of the four neighboring pixels around the pixel.

[0081] Based on the above steps, a standardized image with uniform pixel values ​​and size is obtained, which is suitable for the feature extraction requirements of subsequent pooling separation.

[0082] Based on the above technical solution, the three-step collaborative processing of adaptive filtering to accurately suppress noise, CLAHE algorithm to improve local contrast, and normalization and size adjustment to unify the image format not only preserves the subtle features related to acute abdomen, but also lays a high-quality data foundation for subsequent feature separation and key feature enhancement, making it easier for the target features such as tiny free gas in the intestinal wall and tissue density differences to be captured in subsequent processes.

[0083] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above S103 can be implemented by the following S301, S302 and S303, which are explained in detail below: S301. The standardized image is downsampled by average pooling and max pooling to obtain the global smooth component feature map.

[0084] Among them, average pooling focuses on preserving the overall grayscale distribution of the image, while max pooling focuses on preserving local peak features.

[0085] In this embodiment of the application, for standardized images (Height H × Width W × Number of Channels C), using a 2×2 pooling kernel and a stride of 2, average pooling (AvgPool) and max pooling (MaxPool) are performed respectively, with the following formulas: , Then, the global smooth component feature map is obtained by averaging the two elements. The size after downsampling is (H / 2)×(W / 2)×C.

[0086] It should be noted that the pooling kernel size and step size must be matched with the subsequent upsampling operation to ensure size reversibility; for grayscale abdominal plain X-ray images, the number of channels (C) = 1. A channel is a single feature dimension that stores grayscale values, and only this one channel carries the brightness and darkness information of the image.

[0087] As an example, the normalized image size is 512×512×1, and the size of the global smooth component feature map after downsampling is 256×256×1.

[0088] Based on the above steps, the global smooth component feature map of the overall image structure is extracted in a comprehensive manner, avoiding feature loss caused by single pooling.

[0089] S302. Upsample the global smooth component feature map based on bilinear interpolation to obtain the upsampled feature map.

[0090] Bilinear interpolation restores the size by weighting the four neighboring pixels around the target pixel.

[0091] In this embodiment of the application, the global smoothing component feature map is... Perform bilinear interpolation for the target pixel (x, y) using the formula... Calculate pixel values ​​( Using interpolation weights, the image is restored to its original size H×W×C of the normalized image to obtain the upsampled feature map. .

[0092] It should be noted that the upsampling process must maintain the same number of channels in the feature map as the original image.

[0093] As an example, a 256×256×1 global smooth component feature map is upsampled to obtain a 512×512×1 upsampled feature map.

[0094] Based on the above steps, a background structure estimate aligned with the original image space is obtained, providing an accurate subtraction benchmark for extracting minute lesion features from the original image.

[0095] S303. Subtract the upsampled feature map from the standardized image to obtain the local detail component feature map.

[0096] In the embodiments of this application, such as Figure 4 As shown, for standardized images With upsampled feature map Perform pixel-by-pixel subtraction, the calculation formula is: When the pixel value is negative, the absolute value is taken, and the final result is the local detail component feature map that reflects the detailed information. (Dimensions H×W×C).

[0097] It should be noted that the pixel value needs to be normalized after the subtraction operation to avoid numerical overflow affecting subsequent processing.

[0098] As an example, A certain pixel has a value of 0.6. The corresponding pixel value is 0.4, after calculation The corresponding pixel value is 0.2.

[0099] Based on the above steps, local detail components are accurately separated, providing core objects for subsequent targeted enhancement.

[0100] Based on the above technical solution, precise feature separation is achieved, which not only preserves the global information of the overall image structure, but also efficiently extracts details related to acute abdomen.

[0101] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 5 As shown, the above S104 can be specifically implemented through the following S501 to S504, which are explained in detail below: S501. Extract the local detail feature vectors of the feature-annotated regions in the feature template library.

[0102] Among them, the local detail feature vector is a one-dimensional vector obtained by dimensional compression of the gray-level detail information of the feature-annotated region.

[0103] In this embodiment, all images in the feature template library are traversed, and the feature annotation region of each image is extracted. For each annotation region (size...), Perform global average pooling to obtain feature values, and then concatenate the feature values ​​of all labeled regions to form a local detail feature vector V with dimension K×1 (K is the total number of labeled regions in the feature template library).

[0104] It should be noted that the labeled area must have complete boundaries to avoid missing key features during the truncation process.

[0105] As an example, the template library contains 1000 labeled regions, K=1000, and the local detail feature vector V is a one-dimensional vector of 1000×1.

[0106] S502. Flatten the pixel features of the local detail component feature map to obtain the feature vector corresponding to each pixel.

[0107] The flattening process converts the two-dimensional grayscale features in the neighborhood of a pixel into a one-dimensional vector while preserving the correlation of local features.

[0108] In this embodiment of the application, the local detail component feature map For each pixel (x,y) of a (H×W×1) matrix, extract its 5×5 neighborhood feature block. (x,y) (5×5×1), flattened into a 1×25 one-dimensional vector v(x,y) in row priority order, and after traversing all pixels, we get a set of pixel feature vectors of H×W×25.

[0109] As an example, the local detail component feature map has a size of 512×512×1. The neighborhood feature blocks of each pixel are flattened into 1×25 vectors, forming a vector set of 512×512×25.

[0110] Based on the above steps, the local features of pixels are transformed into a computable vector form, laying the foundation for similarity comparison.

[0111] S503. Calculate the cosine similarity between the feature vector of each pixel and the local detail feature vector.

[0112] Cosine similarity is used to measure the directional consistency between two vectors, and its value ranges from [-1, 1]. The closer the value is to 1, the higher the similarity.

[0113] In this embodiment, for each pixel vector v(x,y) (1×25) and local detail feature vector V (K×1), V is first expanded to K×25 (with repeated padding), and then the cosine similarity is calculated. The average of all sim(x,y) is taken as the final similarity s(x,y) of the pixel.

[0114] Optionally, the cosine similarity is calculated according to the following formula:

[0115] Where i is the index of the labeled region, i = 1, 2, ..., K. The norm of a vector.

[0116] Based on the above steps, the correlation between each pixel and the features of acute abdomen is accurately quantified, providing data support for weight allocation.

[0117] S504. Fuse all cosine similarity results to generate a weighted map that covers the feature maps of local detail components.

[0118] The weight map maps the similarity results to weight values ​​in the [0,1] interval, and is used to characterize the priority of pixel for feature enhancement.

[0119] In this embodiment of the application, the similarity s(x,y) of all pixels is normalized by Min-Max, and after traversing all pixels, a weight map W (H×W×1) with the same size as the local detail component feature map is generated.

[0120] Based on the above technical solution, accurate matching of acute abdominal pain-related features and local detail component feature maps is achieved. The generated weight map can accurately highlight highly correlated pixel regions, providing targeted guidance for subsequent local detail component enhancement and effectively improving the recognition accuracy of key subtle features.

[0121] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 6 As shown, the above S106 can be specifically implemented through the following S601 to S604, which are explained in detail below: S601. Based on the first branch, perform depthwise separable convolution, local response normalization and channel adjustment on the local detail enhancement feature map to obtain the density difference feature map.

[0122] Local Response Normalization (LRN) is used to suppress local feature redundancy and enhance feature discriminative power; channel adjustment achieves dimension adaptation through convolution.

[0123] In this embodiment of the application, the feature map for enhancing local details is... Perform a 3×3 depthwise separable convolution (stride 1, padding=1) on a single-channel grayscale feature map (H×W×1) to output the feature map. Then, each pixel is normalized using its local response; finally, a 1×1 convolution is used to adjust the number of channels to 4, resulting in a density difference feature map. (H×W×4).

[0124] Optionally, the local response is normalized to satisfy the following formula:

[0125] in, These are the normalized eigenvalues. This is the bias constant (to avoid a denominator of 0). is the scaling factor (controlling the normalization intensity to prevent over-suppression of features), j is the channel index, and C is the total number of channels in the input LRN feature map. The exponential coefficient is used to adjust the degree of nonlinearity and enhance the distinguishability of features.

[0126] It should be noted that the 3×3 kernel of depthwise separable convolution must be strictly matched with padding=1 to ensure... and Consistent spatial dimensions; LRN hyperparameters (k=2, = , =0.75, n=5) Referencing the common values ​​of LRN in the field of deep learning (such as AlexNet), n=5 in single channel can reserve space for subsequent multi-channel feature association; the number of channels is adjusted to 4 because the density difference of acute abdomen needs to cover 4 core scenarios (intestinal wall-micro free gas, normal tissue-abnormal tissue, organ-surrounding tissue, other subtle density fluctuations), and 4 channels can be adapted and stitched with the second branch contour detail feature map (H×W×4).

[0127] Based on the above steps, the core features of density differences related to acute abdomen are accurately extracted, providing targeted feature support for the subsequent post-processing to annotate the intestinal wall micro-free gas and tissue density difference regions, while taking into account the feature purity and subsequent multi-channel fusion requirements.

[0128] S602. Based on the second branch, perform dilated convolution, edge-aware residual processing, and edge purification on the local detail enhancement feature map to obtain the contour detail feature map.

[0129] Among them, dilated convolution expands the receptive field through dilation rate, and edge-aware residuals are used to preserve the integrity of edge features.

[0130] In the embodiments of this application, for Perform a 3×3 dilated convolution with a dilation rate of 2 (padding=2) to obtain The original features are fused using an edge-aware residual module; finally, the gradient magnitude is calculated using the Sobel operator. , Let be the square of the gradient in the x-direction. The square of the gradient in the y-direction is used to filter pixels whose gradient G is higher than the threshold T to complete edge purification and obtain the contour detail feature map. (H×W×4).

[0131] Optionally, the fused original features satisfy the following formula:

[0132] in, This is the output feature map of the edge-aware residual module. This indicates a 1×1 convolution.

[0133] As an example, with T=0.15, pixels with a gradient magnitude greater than 0.15 are preserved as edge features.

[0134] Based on the above steps, the contour edges of organs and subtle contour changes are enhanced to avoid loss of details caused by downsampling.

[0135] S603: The density difference feature map, contour detail feature map and upsampled feature map are stitched together by the fusion unit.

[0136] The splicing operation is performed based on the channel dimension, and it is necessary to ensure that the pixel size of the three types of feature maps is completely consistent (H×W).

[0137] In this embodiment of the application, a density difference feature map is obtained. (H×W×4) Outline detail feature map (H×W×4) and upsampled feature map (H×W×1), the spliced ​​feature map is obtained by splicing along the channel dimension. The output size is H×W×(4+4+1), which is H×W×9.

[0138] It should be noted that the upsampled feature map has the same pixel size as the normalized image, density difference feature map, and contour detail feature map.

[0139] As an example, all three feature maps are 512×512 pixels in size, and after stitching... It is 512×512×9.

[0140] Based on the above steps, density differences, contour details, and overall structural features are integrated to achieve complete preservation of multi-dimensional features.

[0141] S604. The weights of the spliced ​​feature maps are adjusted by the channel attention module, and the fused feature maps are obtained after convolutional dimensionality reduction.

[0142] The channel attention module is used to adaptively allocate channel weights, highlighting the contribution of key feature channels.

[0143] In the embodiments of this application, for Perform global average pooling (H×W×9) to obtain a 9-dimensional channel feature vector. Channel attention weights are calculated using two fully connected layers (9→4→9). ;Will and Multiplying by channel one by one yields the weighted feature map. Finally, a 3×3 convolution (stride 1, padding=1) is used to reduce the number of channels to 1, and the fused feature map is output. (H×W×1).

[0144] It should be noted that the number of neurons in a fully connected layer can be adjusted according to the number of channels.

[0145] As an example, The value range is [0.1, 0.9], and the key feature channel weight is close to 0.9.

[0146] Based on the above steps, the key channel features related to acute abdomen are enhanced, while the dimensionality is reduced to decrease the amount of subsequent computation.

[0147] Based on the above technical solution, density difference and contour detail features are extracted by dual-branch differential extraction. Combined with intelligent weight allocation and multi-feature stitching fusion using channel attention mechanism, the overall structural information of the image is preserved, while the core features related to acute abdomen are accurately enhanced. The resulting fused feature map is both complete and targeted.

[0148] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 7 As shown, the above S107 can be specifically implemented through the following S701 to S704, which are explained in detail below: S701, Perform a 1×1 convolution operation on the fused feature map.

[0149] The 1×1 convolution is used to adjust the channel dimension of the feature map and achieve cross-channel feature aggregation without changing the spatial size.

[0150] In this embodiment of the application, for the fused feature map (H×W×1), using a 1×1 convolution kernel (output channels = 8), with stride = 1 and padding = 0, using the formula =1 ( , The calculation (s=1, p=0) outputs a feature map with dimensions still H×W×8. ,in, This indicates that the spatial dimension of the convolution kernel is 1 row × 1 column, which is the smallest convolution kernel size. s is the stride and p is the padding.

[0151] It should be noted that the number of output channels, 8, is an empirical value adapted to subsequent depthwise separable convolutions and ConvFFN, balancing feature representation and computational cost.

[0152] As an example, The size is 512×512×1, which is convolved by 1×1 to obtain a size of 512×512×8. .

[0153] Based on the above steps, the dimensionality expansion from single-channel to multi-channel is completed, laying the foundation for subsequent multi-scale feature extraction.

[0154] S702. Perform a 5×5 depthwise separable convolution on the feature map after the 1×1 convolution.

[0155] Among them, depthwise separable convolution splits spatial convolution and channel convolution, which reduces the amount of computation while retaining the ability to extract spatial features.

[0156] In the embodiments of this application, for (H×W×8) Perform 5×5 depthwise separable convolution. First, extract spatial features through depthwise convolution (5×5 convolution per channel), and then fuse channel information through pointwise convolution (1×1 convolution). Set stride=1 and padding=2 to ensure that the output size is consistent with the input to obtain the feature map. (H×W×8).

[0157] It should be noted that padding=2 matches the 5×5 core, avoiding the loss of spatial features and conforming to the original intention of lightweight design.

[0158] As an example, 512×512×8 After processing, Maintain the dimensions of 512×512×8.

[0159] Based on the above steps, multi-scale spatial features (such as the spatial distribution of tiny gas regions) can be extracted efficiently, with a computational cost of only 1 / 8 that of a regular 5×5 convolution.

[0160] S703. Input the output of the depthwise separable convolution into ConvFFN, and perform dimensionality-upgrading convolution, depthwise convolution, and dimensionality-reducing convolution in sequence.

[0161] Among them, ConvFFN (Convolutional Feedforward Network) enhances feature representation through channel dimension transformation. A channel expansion ratio of 1:4 means that the number of channels after dimensionality enhancement is 4 times that of the input.

[0162] In the embodiments of this application, for A 1×1 upscaling convolution is performed on (H×W×8) channels, expanding the number of channels from 8 to 32 (8×4); then, a 3×3 depthwise convolution is used to extract intra-channel features; finally, a 1×1 downscaling convolution is used to restore the number of channels to 8, yielding the feature map. (H×W×8), with a total step length of 1 and padding of 1, keeping the dimensions constant.

[0163] It should be noted that depthwise convolution only operates on a single channel to avoid redundant computation across channels.

[0164] As an example, after dimensionality upscaling, the number of channels is 32. After depthwise convolution, the number of channels is reduced back to 8. The feature dimensions are dynamically adjusted to enhance the expression.

[0165] Based on the above steps, the semantic information of features is enriched through a process of channel dimensionality enhancement, depth extraction, and dimensionality reduction compression, thus avoiding the problem of simplistic feature representation.

[0166] S704. Perform residual concatenation between the dimensionality-reduced convolutional features and the 1×1 convolutional feature map to output a feature enhancement map.

[0167] Among them, residual connections fuse shallow and deep features through pixel-by-pixel addition, which alleviates the gradient vanishing problem and preserves the original feature information.

[0168] In the embodiments of this application, (H×W×8) and S701 output Perform pixel-by-pixel addition operations (H×W×8) to output a feature enhancement map. (H×W×8).

[0169] It should be noted that both must maintain identical dimensions and number of channels to ensure that addition operations are feasible.

[0170] As an example, 512×512×8 and After adding them together, we get 512 × 512 × 8. .

[0171] Based on the above steps, features from different levels are integrated to improve the robustness of key features of acute abdomen and avoid feature degradation caused by deep networks.

[0172] Based on the above technical solution, while significantly reducing computational complexity and adapting to clinical terminal deployment, it fully enhances the subtle features related to acute abdominal pain.

[0173] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 8 As shown, the above S108 can be specifically implemented through the following S801 to S803, which are explained in detail below: S801. Map the feature enhancement map to the grayscale range of the abdominal plain X-ray image.

[0174] Among them, grayscale mapping restores the feature values ​​output by the model to the grayscale range of routine clinical X-ray images, adapting to the reading habits.

[0175] In this embodiment of the application, for feature enhancement maps (H×W×8), first average the pixel values ​​of each channel to obtain the single-channel feature map. (H×W×1), then through the linear mapping formula (x,y)= (x,y)×( )+ Calculation, where =0、 =4095 (standard grayscale range for X-ray images), output mapped image (H×W×1).

[0176] It should be noted that before mapping, it is necessary to ensure Pixel values ​​have been normalized to [0,1] to avoid exceeding the target grayscale range.

[0177] Based on the above steps, the abstract feature map is transformed into a clinically recognizable grayscale image.

[0178] S802. Morphological opening operations are used to remove minor artifacts in the feature enhancement map after grayscale mapping to obtain an enhanced image.

[0179] Among them, the morphological opening operation consists of "erosion + dilation", and its core function is to remove tiny bright spot artifacts while preserving the shape and size of the target features.

[0180] In this embodiment of the application, a 3×3 rectangular structural element SE is selected, and first... Perform corrosion operation = SE (eliminates tiny artifacts smaller than 3×3), then performs dilation. = SE (original contour restoration of target features) ultimately yields the enhanced image. (H×W×1), For erosion calculation, This is for the expansion operation.

[0181] Based on the above steps, enhanced images with no redundant artifacts and prominent key features are obtained, improving the clarity of clinical image interpretation.

[0182] S803. Based on the density difference feature map, the micro-free gas region of the intestinal wall is labeled by adaptive threshold segmentation, the region of tissue density difference is labeled by threshold screening, and the stepped gas-liquid plane is identified and labeled based on the density difference feature map; based on the contour detail feature map, the abnormal region of organ contour is labeled by edge detection, and the region of intestinal dilation is identified and labeled by combining the morphological features of the intestinal tube, thus generating a feature label map.

[0183] Among them, adaptive threshold segmentation can automatically adapt to the grayscale distribution of the image, threshold filtering is used to accurately capture density differences, and edge detection focuses on contour anomaly recognition.

[0184] In this embodiment of the application, the density difference feature map is first analyzed. (H×W×4) Adaptive threshold calculated using the Otsu algorithm. , pixel value greater than The area is marked as a tiny free gas in the intestinal wall; then a fixed threshold is set. =0.2, filter Medium pixel value greater than The regions are labeled as differences in tissue density. Based on the density difference feature map, channel averaging is first applied to obtain a single-channel density feature map. Then, horizontal gradient filtering is performed to extract linear features of alternating high and low densities. Regions exhibiting a stepped distribution, with at least two stepped layers and a continuous length meeting the standard are identified as stepped gas-liquid planes and labeled accordingly. Finally, the contour detail feature map is... (H×W×4) Perform Canny edge detection, set low and high thresholds, mark areas with continuous edge breaks or irregularities as organ contour abnormalities, extract complete and continuous intestinal contours based on contour detail feature maps, calculate the maximum transverse diameter of intestinal segments using morphological algorithms in combination with intestinal morphology features, identify areas with diameters exceeding the clinical normal intestinal diameter threshold as intestinal dilatation areas and mark them, and integrate the three types of markings to obtain a feature annotation map.

[0185] It should be noted that the threshold needs to be calibrated based on clinical samples to ensure a balance between annotation sensitivity and specificity; in Canny edge detection, the low threshold is used for preliminary edge screening, and the high threshold is used to identify strong edges, with the high threshold being 3 times the low threshold.

[0186] Based on the above technical solution, the enhanced features of the model are transformed into clinically usable enhanced images and feature annotation maps, which not only highlight subtle key features but also conform to the logic of image reading, significantly reducing the difficulty of image analysis.

[0187] In one possible implementation, combining Figure 1 ,like Figure 9 As shown, following S108, the CRNet-based pre-detection and optimization method for acute abdomen provided in this application embodiment further includes S901: S901. Add the feature annotation map to the feature template library.

[0188] Among them, the feature-annotated images included are effective grayscale images containing three types of annotations: tiny free gases in the intestinal wall, differences in tissue density, and abnormal organ contours, which are consistent with the existing sample format in the template library.

[0189] In this embodiment of the application, the feature annotation map output in S108 is stored in the feature template library, and the template library retrieval index is updated.

[0190] Based on the above steps, the feature template library can be dynamically expanded to enrich the coverage of feature samples for acute abdominal pain.

[0191] In some embodiments, to further verify the practicality and effectiveness of the present invention in actual clinical environments, a nighttime emergency Triage workflow application scenario based on the above method is provided. This application scenario deeply embeds the image processing algorithms of the aforementioned embodiments into the hospital's existing radiology information system, aiming to solve the clinical pain points of missed or delayed diagnosis of acute abdominal conditions due to the shortage of medical resources and high fatigue of radiologists during nighttime emergency periods, and to achieve a leap from simple image enhancement processing to intelligent clinical decision support.

[0192] Specifically, in this application scenario, the CRNet-based acute abdominal pre-screening and optimization device is deployed on the backend server of the hospital's PACS system or a standalone AI inference workstation, and establishes a real-time communication connection with the image acquisition equipment and reading terminal via the DICOM protocol. When the emergency department completes the acquisition of abdominal X-ray images and uploads them to the PACS system, the communication unit automatically listens to and captures the image data stream, instantly triggering the processing unit to execute the complete analysis process described above without manual intervention. This silent background triggering mechanism ensures the parallel execution of the pre-screening assistance process and the regular image transmission process, without increasing the workload of technicians or changing their existing work habits, demonstrating the seamless integration of the technology into the existing clinical workflow.

[0193] After obtaining the auxiliary optimization results, which include the original image, enhanced image, and feature annotation map, the system does not simply store or display them, but further executes intelligent triage logic based on feature content. The processing unit parses the key semantic information in the feature annotation map, especially prioritizing high-risk signs identified by S803. For example, when a free gas area under the diaphragm is detected in the feature annotation map, or when more than a preset threshold number of stepped air-fluid levels (such as three or more) is identified, or when a significantly abnormally large dilated intestinal tract area is marked, the system automatically marks the case as high-risk for acute abdomen. Conversely, if only slight organ outline blurring or non-specific density differences are detected, it is marked as ordinary observation level. This triage logic directly utilizes the precise feature expression capability established in the spatial domain decomposition and multi-branch feature fusion stages of this invention, transforming the abstract algorithm output into triage criteria with clear clinical meaning.

[0194] Based on the above classification results, the system automatically adjusts the sorting strategy of the PACS image reading list. For cases marked as high-risk, their auxiliary optimization results are forcibly pushed to the top of the on-call doctor's image reading queue, and a classification warning is triggered with a prominent visual indicator (such as a flashing red border or pop-up prompt), reminding the doctor to pay priority attention. At the same time, the system overlays enhanced images and feature annotation maps in the preview interface by default, allowing doctors to intuitively see the location and shape of lesions enhanced by the algorithm the moment they open the image, without having to manually adjust the window width and level or switch layers. For ordinary-risk cases, they are arranged in the usual chronological order, with auxiliary view options only provided when doctors actively access them. This dynamic queue management mechanism essentially transforms the image processing capabilities of this invention into a time-dimensional medical resource optimization tool, ensuring that in the face of critical and severe cases requiring the fastest decision-making, the limited attention of doctors can be accurately guided to the most critical targets.

[0195] This invention, through the synergistic effect of prior knowledge injection guided by the feature template library, dual-branch differentiated feature extraction, and lightweight real-time inference techniques described in the foregoing embodiments, can automatically screen and enhance the visualization of high-risk features within seconds, essentially providing on-call doctors with a tireless digital assistant. This not only shortens the time window from image acquisition to preliminary diagnosis but also saves time, thereby improving the overall quality and safety of emergency medical services.

[0196] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as the CRNet-based acute abdominal pre-detection assistance optimization device, includes at least one of the hardware structures and software modules corresponding to each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0197] This application embodiment can divide the CRNet-based acute abdominal pain pre-detection auxiliary optimization device into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0198] When using integrated units, Figure 10 A possible structural schematic diagram of the CRNet-based acute abdominal pre-detection auxiliary optimization device (referred to as CRNet-based acute abdominal pre-detection auxiliary optimization device 100) involved in the above embodiments is shown. The CRNet-based acute abdominal pre-detection auxiliary optimization device 100 includes a processing unit 1001 and a communication unit 1002, and may also include a storage unit 1003. Figure 10 The schematic diagram shown can be used to illustrate the structure of the CRNet-based acute abdominal pre-detection and optimization device involved in the above embodiments.

[0199] when Figure 10 The schematic diagram shown illustrates the structure of the CRNet-based acute abdominal pre-detection and optimization device involved in the above embodiments. The processing unit 1001 is used to control and manage the operation of the CRNet-based acute abdominal pre-detection and optimization device, the communication unit 1002 is used for the CRNet-based acute abdominal pre-detection and optimization device to communicate with other devices, and the storage unit 1003 is used to store the program code and data of the CRNet-based acute abdominal pre-detection and optimization device.

[0200] For example, the communication unit 1002 is used to acquire abdominal plain X-ray images and a feature template library; the feature template library is a collection of abdominal plain X-ray images of patients with acute abdomen containing feature annotations; the feature annotations include feature regions such as small free gas in the intestinal wall, gas-fluid levels, intestinal dilation, blurred organ outlines, and differences in tissue density; Processing unit 1001 is used to preprocess abdominal plain X-ray images to obtain standardized images. Preprocessing includes noise suppression, grayscale correction, and standardization. The standardized images are spatially decomposed using pooling to obtain global smooth component feature maps and local detail component feature maps. Similarity is calculated between the local detail component feature maps and a feature template library to generate a weight map. The local detail component feature maps are enhanced based on the weight map to obtain local detail enhancement feature maps. The local detail enhancement feature maps are input into an improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map. The improved Multi-Branch Block includes a first branch, a second branch, and a fusion unit. The fused feature map is input into a lightweight convolutional enhancement module for feature enhancement to obtain a feature enhancement map. The lightweight convolutional enhancement module includes depthwise separable convolution and ConvFFN. The feature enhancement map is post-processed to output auxiliary optimization results. The auxiliary optimization results include the original image, the enhanced image, and the feature annotation map.

[0201] In one possible implementation, the processing unit 1001 is further configured to dynamically adjust the filtering window according to the pixel neighborhood noise density, perform adaptive filtering on the abdominal X-ray image, and suppress noise; segment the abdominal X-ray image into sub-blocks based on the CLAHE algorithm and limit the contrast gain to perform grayscale correction; and perform pixel value normalization processing on the grayscale corrected abdominal X-ray image and adjust the image size to obtain a standardized image.

[0202] In one possible implementation, the processing unit 1001 is further configured to downsample the standardized image using average pooling and max pooling to obtain a global smooth component feature map; upsample the global smooth component feature map based on bilinear interpolation to obtain an upsampled feature map; and subtract the upsampled feature map from the standardized image to obtain local detail component features.

[0203] In one possible implementation, the processing unit 1001 is further configured to: extract local detail feature vectors of the feature-annotated regions in the feature template library; flatten the pixel features of the local detail component feature map to obtain the feature vectors corresponding to each pixel; calculate the cosine similarity between the feature vector of each pixel and the local detail feature vector; and fuse all cosine similarity results to generate a weight map covering the local detail component feature map.

[0204] In one possible implementation, the processing unit 1001 is further configured to perform depthwise separable convolution, local response normalization, and channel adjustment processing on the local detail enhancement feature map based on the first branch to obtain a density difference feature map; perform dilated convolution, edge-aware residual processing, and edge purification on the local detail enhancement feature map based on the second branch to obtain a contour detail feature map; concatenate the density difference feature map, contour detail feature map, and upsampled feature map through a fusion unit; and adjust the weights of the concatenated feature map through a channel attention module, and obtain a fused feature map after convolutional dimensionality reduction processing.

[0205] In one possible implementation, the processing unit 1001 is further configured to perform a 1×1 convolution operation on the fused feature map; perform a 5×5 depthwise separable convolution on the feature map after the 1×1 convolution; input the output of the depthwise separable convolution into ConvFFN, and sequentially perform dimensionality-up convolution, depthwise convolution, and dimensionality-down convolution; the channel expansion ratio of ConvFFN is 1:4; and perform residual concatenation between the dimensionality-down convolutioned feature map and the 1×1 convolutioned feature map to output a feature enhancement map.

[0206] In one possible implementation, the processing unit 1001 is further configured to map the feature enhancement map to the grayscale range of the abdominal plain X-ray image; remove minute artifacts in the grayscale-mapped feature enhancement map using morphological opening operations to obtain an enhanced image; label the small free gas regions of the intestinal wall by adaptive threshold segmentation based on the density difference feature map, label the tissue density difference regions by threshold filtering, and identify and label the stepped gas-liquid planes; and label the abnormal regions of organ contours by edge detection based on the contour detail feature map, and identify and label the dilated regions of the intestinal tract based on the morphological features of the intestinal tract to generate a feature annotation map.

[0207] In one possible implementation, the upsampled feature map has the same pixel size as the normalized image, density difference feature map, and contour detail feature map.

[0208] In one possible implementation, the processing unit 1001 is also used to include the feature annotation map into the feature template library.

[0209] The processing unit 1001 can be a processor or a controller, and the communication unit 1002 can be a communication interface, transceiver, transceiver circuit, transceiver device, etc. The term "communication interface" is a general term and may include one or more interfaces. The storage unit 1003 can be a memory. When the CRNet-based acute abdominal pain pre-detection and optimization device 100 is a chip, the processing unit 1001 can be a processor or a controller, and the communication unit 1002 can be an input interface and / or an output interface, pins, or circuits, etc. The storage unit 1003 can be a storage unit within the chip (e.g., a register, cache, etc.) or a storage unit located outside the chip (e.g., read-only memory (ROM), random access memory (RAM, etc.).

[0210] The communication unit can also be referred to as a transceiver unit. The antenna and control circuit with transceiver functions in the CRNet-based acute abdomen pre-detection auxiliary optimization device 100 can be considered as the communication unit 1002 of the CRNet-based acute abdomen pre-detection auxiliary optimization device 100, and the processor with processing functions can be considered as the processing unit 1001 of the CRNet-based acute abdomen pre-detection auxiliary optimization device 100. Optionally, the device in the communication unit 1002 used to implement the receiving function can be considered as a communication unit, which is used to execute the receiving steps in the embodiments of this application. The communication unit can be a receiver, a receiver circuit, etc. The device in the communication unit 1002 used to implement the transmitting function can be considered as a transmitting unit, which is used to execute the transmitting steps in the embodiments of this application. The transmitting unit can be a transmitter, a transmitter, a transmitting circuit, etc.

[0211] Figure 10 If the integrated units in the process are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. Storage media for storing computer software products include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0212] Figure 10 The units in the process can also be called modules; for example, a processing unit can be called a processing module.

[0213] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0214] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0215] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A pre-detection and optimization method for acute abdomen based on CRNet, characterized in that, include: Acquire abdominal plain X-ray images and feature template library; The feature template library is a collection of plain X-ray images of the abdomen of patients with acute abdomen containing feature annotations; the feature annotations include characteristic regions such as tiny free gas in the intestinal wall, gas-fluid levels, intestinal dilation, blurred organ outlines, and differences in tissue density; The abdominal plain X-ray image is preprocessed to obtain a standardized image; the preprocessing includes noise suppression, grayscale correction and standardization. The standardized image is spatially decomposed by pooling to obtain a global smooth component feature map and a local detail component feature map. A weighted map is generated by calculating the similarity between the local detail component feature map and the feature template library. The local detail component feature map is enhanced based on the weight map to obtain a local detail enhanced feature map; The local detail enhancement feature map is input into the improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map; the improved Multi-Branch Block includes a first branch, a second branch, and a fusion unit; The fused feature map is input into a lightweight convolutional enhancement module for feature enhancement to obtain an enhanced feature map; the lightweight convolutional enhancement module includes depthwise separable convolution and ConvFFN; The feature enhancement map is post-processed to output auxiliary optimization results; the auxiliary optimization results include the original image, the enhanced image, and the feature annotation map.

2. The method according to claim 1, characterized in that, The preprocessing of the abdominal plain X-ray image to obtain a standardized image includes: The filtering window is dynamically adjusted based on the pixel neighborhood noise density to perform adaptive filtering on the abdominal plain X-ray image and suppress noise. Based on the CLAHE algorithm, abdominal plain X-ray images are segmented into sub-blocks and contrast gain is limited for grayscale correction. The pixel values ​​of the grayscale-corrected abdominal plain X-ray images were normalized and the image size was adjusted to obtain a standardized image.

3. The method according to claim 1, characterized in that, The step of performing spatial domain decomposition on the normalized image through pooling to obtain global smooth component feature maps and local detail component feature maps includes: The standardized image is downsampled using average pooling and max pooling to obtain the global smooth component feature map; The global smooth component feature map is upsampled based on bilinear interpolation to obtain the upsampled feature map. The local detail component features are obtained by subtracting the upsampled feature map from the standardized image.

4. The method according to claim 1, characterized in that, The step of calculating the similarity between the local detail component feature map and the feature template library to generate a weight map includes: Extract local detail feature vectors from the feature-annotated regions in the feature template library; The pixel features of the local detail component feature map are flattened to obtain the feature vector corresponding to each pixel. Calculate the cosine similarity between the feature vector of each pixel and the local detail feature vector; All cosine similarity results are fused to generate a weighted map that covers the local detail component feature map.

5. The method according to claim 3, characterized in that, The step of inputting the local detail enhancement feature map into the improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map includes: Based on the first branch, the local detail enhancement feature map is processed by depthwise separable convolution, local response normalization and channel adjustment to obtain a density difference feature map. Based on the second branch, the local detail enhancement feature map is subjected to dilated convolution, edge-aware residual processing and edge purification to obtain the contour detail feature map. The density difference feature map, the contour detail feature map, and the upsampled feature map are stitched together by a fusion unit; The weights of the concatenated feature maps are adjusted using a channel attention module, and the fused feature maps are obtained after convolutional dimensionality reduction.

6. The method according to claim 1, characterized in that, The step of inputting the fused feature map into the lightweight convolutional enhancement module for feature enhancement to obtain the enhanced feature map includes: Perform a 1×1 convolution operation on the fused feature map; Perform a 5×5 depthwise separable convolution on the feature map after 1×1 convolution. The output of the depthwise separable convolution is input into the ConvFFN, and then the dimensionality-up convolution, depthwise convolution, and dimensionality-down convolution are performed sequentially; the channel expansion ratio of the ConvFFN is 1:

4. The feature map after dimensionality reduction convolution is residually concatenated with the feature map after 1×1 convolution to output the feature enhancement map.

7. The method according to claim 5, characterized in that, The post-processing of the feature enhancement map to output auxiliary optimization results includes: The feature enhancement map is mapped to the grayscale range of the abdominal plain X-ray image; The enhanced image is obtained by removing minute artifacts in the feature enhancement map after grayscale mapping using morphological opening operations; Based on the density difference feature map, the region of tiny free gas in the intestinal wall is labeled by adaptive threshold segmentation, the region of tissue density difference is labeled by threshold filtering, and the stepped gas-liquid plane is identified and labeled; based on the contour detail feature map, the region of abnormal organ contour is labeled by edge detection, and the region of intestinal dilation is identified and labeled based on the morphological features of the intestinal tube, thus generating the feature label map.

8. The method according to claim 5, characterized in that, The upsampled feature map has the same pixel size as the normalized image, the density difference feature map, and the contour detail feature map.

9. The method according to claim 1, characterized in that, After post-processing the feature enhancement map and outputting the auxiliary optimization result, the method further includes: incorporating the feature annotation map into the feature template library.

10. A pre-detection and optimization device for acute abdomen based on CRNet, characterized in that, The device includes: a communication unit and a processing unit; The communication unit is used to acquire abdominal plain X-ray images and a feature template library; the feature template library is a collection of abdominal plain X-ray images of patients with acute abdomen containing feature annotations; the feature annotations include feature regions such as small free gas in the intestinal wall, gas-fluid levels, intestinal dilation, blurred organ outlines, and differences in tissue density; The processing unit is used to preprocess the abdominal plain X-ray image to obtain a standardized image; the preprocessing includes noise suppression, grayscale correction, and standardization; the standardized image is spatially decomposed by pooling to obtain a global smooth component feature map and a local detail component feature map; the similarity between the local detail component feature map and the feature template library is calculated to generate a weight map; the local detail component feature map is enhanced based on the weight map to obtain a local detail enhancement feature map; the local detail enhancement feature map is input into an improved Multi-Branch Block for multi-scale feature fusion to obtain a fused feature map; the improved Multi-Branch Block includes a first branch, a second branch, and a fusion unit; the fused feature map is input into a lightweight convolutional enhancement module for feature enhancement to obtain a feature enhancement map; the lightweight convolutional enhancement module includes depthwise separable convolution and ConvFFN; the feature enhancement map is post-processed to output auxiliary optimization results; the auxiliary optimization results include the original image, the enhanced image, and the feature annotation map.

Images