Pneumonia medical assistance system based on improved VMamba model

Abstract

This invention relates to the field of medical assistance technology, specifically to a pneumonia medical assistance system based on an improved VMamba model. The system includes: dividing the lung image to be diagnosed into raster cells based on a raster cell partitioning algorithm; calculating the lesion representation coefficient of each raster cell within each management time period; calculating the lesion representation index of the lesion trend path within the management period based on the raster cells along the lesion trend path; determining whether the raster cells covered by the lesion trend path are lesion regions based on the lesion representation index; inputting the lung field feature map and semantic feature vector into the improved VMamba model after channel-level concatenation, and outputting the pneumonia classification diagnosis result; and linking the pneumonia classification diagnosis result with a pneumonia knowledge graph to generate a visual auxiliary diagnosis report, significantly improving the accuracy and stability of lung medical image processing and achieving full-process automation of intelligent image analysis.

Description

Technical Field

[0001] This invention relates to the field of medical assistance technology, and more specifically to a pneumonia medical assistance system based on an improved VMamba model. Background Technology

[0002] As a common respiratory disease in clinical practice, the efficiency and accuracy of pneumonia diagnosis directly affect the timing of treatment and prognosis of patients. Lung imaging analysis is the core means of pneumonia diagnosis. Traditional manual image interpretation relies on the professional experience of physicians and has problems such as low diagnostic efficiency, large subjective error, and easy to miss small lesions, which can hardly meet the needs of large-scale clinical screening and accurate diagnosis.

[0003] With the application of deep learning technology in medical imaging, various neural network models are gradually being used for pneumonia image diagnosis. Among them, models such as U-Net++ and InceptionV3 have shown certain advantages in lung field segmentation and image classification, respectively. The VMamba model, with its sequence modeling capability based on a state-space model, has potential in capturing long-distance feature dependencies. However, existing technical solutions still have many technical shortcomings: First, the localization of lung lesion areas lacks refined spatiotemporal analysis methods. Traditional segmentation methods do not consider the trend changes of lesions in the time dimension, making it difficult to accurately identify dynamically developing lesion areas. Second, the feature fusion degree between lung field segmentation and pneumonia classification is low. The segmented feature maps and semantic features are not deeply coupled, resulting in insufficient utilization of feature information. Third, when the VMamba model is directly applied to pneumonia image diagnosis, it suffers from background redundant feature interference, low sensitivity in capturing small lesion features, and insufficient cross-layer feature interaction, making it unable to effectively distinguish high-frequency details of lesions from low-frequency background information.

[0004] In summary, there is an urgent need for a pneumonia medical assistance system that can achieve precise localization of lesion areas, deep fusion of multi-dimensional features, accurate identification of tiny lesions, and output of visualized diagnostic reports. This system would address the problems of low diagnostic efficiency, insufficient accuracy, and low feature utilization in existing technologies, and meet the actual needs of accurate clinical diagnosis of pneumonia. Summary of the Invention

[0005] The purpose of this invention is to provide a pneumonia medical assistance system based on an improved VMamba model: to solve the problems of low diagnostic efficiency, insufficient accuracy and low feature utilization in the existing technology, and to meet the actual needs of accurate clinical diagnosis of pneumonia.

[0006] A pneumonia medical support system based on an improved VMamba model, comprising: The lung imaging acquisition module is used to generate management cycles, divide the management cycle into several management periods, and acquire lung images of patients to be diagnosed within each management period. The lesion region determination module is used to divide the lung image to be diagnosed into grid cells based on the grid cell partitioning algorithm to obtain a set of grid cells, where each grid cell corresponds to a part of the lung image to be diagnosed. The module calculates the lesion characterization coefficient of each grid cell in each management period. Based on the lesion characterization coefficient, a lesion trend path is generated in the set of grid cells where the lung image to be diagnosed is located. Based on the grid cells on the lesion trend path, the lesion characterization index of the lesion trend path in the management period is calculated. Based on the lesion characterization index, it is determined whether the grid cells covered by the lesion trend path are lesion regions. The lesion area image processing module is used to input the lung image part to be diagnosed corresponding to the lesion area into the lung field segmentation based on the U-Net++ model, output the lung field feature map, and use the lung field feature map as input to complete the preliminary classification of pneumonia based on the improved InceptionV3 model and extract the semantic feature vector of the lung area. The pneumonia classification and diagnosis module is used to input the lung field feature map and semantic feature vector into the improved VMamba model after channel-level concatenation. It completes the deep fusion and accurate identification of pneumonia image features through frequency domain channel attention fusion, cross-layer cross attention module and salient feature suppression strategy, and outputs pneumonia classification and diagnosis results. The diagnostic report generation module is used to link the pneumonia classification and diagnosis results with the pneumonia knowledge graph to generate a visual auxiliary diagnostic report.

[0007] Furthermore, the process of dividing the lung image to be diagnosed into raster cells based on the raster cell partitioning algorithm to obtain the raster cell set includes the following steps: Superpixel segmentation of lung images to be diagnosed: The SLIC algorithm is used, and the number of superpixels and compactness parameters are set to aggregate the image pixels into several superpixel blocks, each of which is an independent pixel aggregation unit; Regularization correction of superpixel blocks: The minimum bounding rectangle method is used to convert irregular superpixel blocks into rectangular grid units, while preserving the core feature regions of the superpixel blocks; Invalid raster units that do not meet the preset requirements for image quality index are removed. The remaining raster units are assigned a unique number and pixel feature attributes, resulting in N raster units based on superpixels. All raster units are then combined into a raster unit set.

[0008] Furthermore, the process of removing invalid raster cells whose image quality index does not meet the preset requirements includes the following steps: Calculate the image quality index of the raster unit ; Load the image quality index threshold, and determine whether the image quality index of the raster cell exceeds the image quality index threshold. If it does, the raster cell is considered valid; otherwise, the raster cell is considered invalid.

[0009] Furthermore, calculating the lesion characterization coefficient of each grid cell within each management period specifically includes the following process: Extract image features of the lung images to be diagnosed for each raster unit within each management time period. and load lesion image features ; The lesion matching probability of each edge is calculated based on the centrality of each edge. ; Calculate the average value of the lesion matching probability for all edges, and record the average value as the lesion characterization coefficient for each grid cell in each management period.

[0010] Furthermore, generating the lesion trajectory path within the raster cell set containing the lung image to be diagnosed, based on the lesion characterization coefficient, specifically includes the following process: Step 1: Obtain the two-dimensional spatial coordinates of each grid cell in the set of grid cells containing the lung image to be diagnosed, based on the medical image spatial analysis system. And the lesion characterization coefficient of each grid cell and The combination forms lesion feature nodes, and the representation of lesion feature nodes is as follows: ; Step 2: Number the lesion feature nodes in the grid cell set, and combine the first and second lesion feature nodes of the grid cell set to form a lesion reference vector. , and the remaining first Lesion feature nodes and the first Each lesion feature node forms a lesion comparison vector. ,in, Calculate separately and Angle between ,angle The calculation formula is as follows: ; Step 3: Put With threshold angle Compare, if the first Individual lesion characteristic nodes Angle greater than the threshold Then it will be arranged in the th order. The lesion feature nodes preceding each lesion feature node are subdivided into groups, and the nodes arranged in the order of the first lesion feature nodes are grouped together. Repeat steps two and three to subdivide the grid cell set for each lesion feature node before the node, and record each subdivided group as a segmented path of the lesion trend. Step 4: Combine all lesion trend segments according to the spatial geographical adjacency of the grid cells to obtain the lesion trend path within the set of grid cells containing the lung image to be diagnosed.

[0011] Furthermore, the calculation of the lesion characterization index of the grid cells within the management period based on the grid cells along the lesion trend path specifically includes the following process: Obtain the lesion characterization coefficient of each grid cell on the lesion trend path in each management period. Establish a rectangular coordinate system with the execution time of the management period as the X-axis and the lesion characterization coefficient corresponding to the management period as the Y-axis. Plot the lesion characterization index curve in the rectangular coordinate system by plotting points. Set the lesion characterization index line and mark the part of the lesion characterization index curve that is above the lesion characterization index line. Calculate the integral value of the marked curve part and record the integral value as the lesion characterization index of the lesion trend path within the management period.

[0012] Furthermore, determining whether the grid cells covering the path of a lesion constitute a lesion region based on the lesion characterization index specifically includes the following process: Load the lesion characterization index threshold, and determine whether the lesion characterization index exceeds the lesion characterization index threshold. If it does, determine that the grid cells covered on the lesion trend path are lesion areas. If not, determine that the grid cells covered on the lesion trend path are not lesion areas.

[0013] Furthermore, the implementation process of frequency domain channel attention fusion is as follows: The feature map after channel-level stitching is transformed from the spatial domain to the frequency domain by discrete cosine transform, separating the high-frequency detailed features of the lesion from the low-frequency redundant features of the background. By combining the channel attention mechanism of the simplified SE-Net architecture, weights are assigned to the feature channels after frequency domain transformation to enhance the response of lesion-related feature channels and suppress interference from irrelevant background channels. The optimized frequency domain features are mapped back to the spatial domain through inverse discrete cosine transform, and then fused with the original features before being input into the SSM module of the VMamba model. Adjust the feature mapping dimension of the VMamba model to match the frequency domain feature dimension, optimize the gating mechanism parameters of the SSM module, and improve the sensitivity of capturing small lesion features.

[0014] Furthermore, the VMamba model is improved by incorporating a frequency domain channel attention mechanism, a cross-layer cross attention module, and a salient feature suppression strategy into the VMamba model. The VMamba model is a sequence modeling framework based on the state space model SSM, which includes a 2D selective scan module SS2D and a visual state space block VSS.

[0015] Compared to existing solutions, the beneficial effects achieved by this invention are: Achieving refined and dynamic localization of lesion areas: The system performs superpixel segmentation and regularization correction of lung images through a grid cell partitioning algorithm, and eliminates invalid grid cells by combining image quality index, ensuring the effectiveness of feature analysis; at the same time, by calculating lesion characterization coefficients for multiple management time periods, generating lesion trend paths, and calculating lesion characterization indices, it achieves continuous spatiotemporal analysis of lesion areas, accurately capturing the development trend of lesions. Compared with traditional static segmentation methods, the accuracy and comprehensiveness of lesion area judgment are greatly improved, effectively avoiding the missed diagnosis of small lesions and dynamically developing lesions.

[0016] Improving the efficiency of lung image feature extraction and fusion: The U-Net++ model was used to complete lung field segmentation, ensuring the segmentation accuracy of lung field feature maps; the InceptionV3 model was improved to achieve preliminary classification of pneumonia and extract semantic feature vectors, and the lung field feature maps and semantic feature vectors were deeply fused by channel-level stitching, realizing the complementarity of spatial and semantic features; the VMamba model was improved by incorporating a frequency domain channel attention fusion mechanism, and the spatial domain and frequency domain feature transformation was achieved through discrete cosine transform, effectively separating high-frequency detailed features of lesions from low-frequency redundant features of the background. The channel attention mechanism of the simplified SE-Net architecture was combined for weight allocation, which enhanced the response of lesion-related features, suppressed background interference, and significantly improved feature utilization.

[0017] Improving the accuracy of pneumonia classification and diagnosis: The improved VMamba model, combined with a cross-layer attention module and a salient feature suppression strategy, achieves deep fusion and accurate identification of pneumonia imaging features, enhancing the sensitivity of capturing small lesion features. In the calculation of lesion characterization coefficients, the accuracy of coefficient calculation is ensured by using a feature matching graph structure and edge centrality calculation, combined with logarithmic transformation to reduce interference from dense nodes. This provides a reliable data foundation for subsequent lesion area judgment and classification diagnosis. Compared with traditional models, the accuracy and generalization ability of pneumonia classification and diagnosis are significantly improved. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a system block diagram of a pneumonia medical assistance system based on an improved VMamba model according to an embodiment of the present invention; Figure 2 This is a flowchart of a pneumonia medical assistance system based on an improved VMamba model according to an embodiment of the present invention; Figure 3 This is a flowchart of another pneumonia medical assistance system based on an improved VMamba model, according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring aspects of this disclosure.

[0022] This embodiment provides a pneumonia medical assistance system based on an improved VMamba model. Figure 1 This is a system block diagram of a pneumonia medical assistance system based on an improved VMamba model according to an embodiment of the present invention, such as... Figure 1 As shown, the system includes: The lung imaging acquisition module is used to generate management cycles, divide the management cycle into several management periods, and acquire lung images of patients to be diagnosed within each management period. The lesion region determination module is used to divide the lung image to be diagnosed into grid cells based on the grid cell partitioning algorithm to obtain a set of grid cells, where each grid cell corresponds to a part of the lung image to be diagnosed. The module calculates the lesion characterization coefficient of each grid cell in each management period. Based on the lesion characterization coefficient, a lesion trend path is generated in the set of grid cells where the lung image to be diagnosed is located. Based on the grid cells on the lesion trend path, the lesion characterization index of the lesion trend path in the management period is calculated. Based on the lesion characterization index, it is determined whether the grid cells covered by the lesion trend path are lesion regions. The lesion area image processing module is used to input the lung image part to be diagnosed corresponding to the lesion area into the lung field segmentation based on the U-Net++ model, output the lung field feature map, and use the lung field feature map as input to complete the preliminary classification of pneumonia based on the improved InceptionV3 model and extract the semantic feature vector of the lung area. The pneumonia classification and diagnosis module is used to input the lung field feature map and semantic feature vector into the improved VMamba model after channel-level concatenation. It completes the deep fusion and accurate identification of pneumonia image features through frequency domain channel attention fusion, cross-layer cross attention module and salient feature suppression strategy, and outputs pneumonia classification and diagnosis results. The diagnostic report generation module is used to link the pneumonia classification and diagnosis results with the pneumonia knowledge graph to generate a visual auxiliary diagnostic report.

[0023] In summary, by using a raster cell partitioning algorithm for superpixel segmentation and regularization correction of lung images, and combining this with image quality index to remove invalid raster cells, and by calculating lesion representation coefficients across multiple management time periods, generating lesion trend paths, and calculating lesion representation indices, the accuracy and comprehensiveness of lesion region identification are significantly improved compared to traditional static segmentation methods, effectively avoiding missed diagnoses of small lesions and dynamically developing lesions. The U-Net++ model is used for lung field segmentation, ensuring the segmentation accuracy of the lung field feature map. An improved InceptionV3 model is used for preliminary pneumonia classification and semantic feature vector extraction. Channel-level stitching is combined to deeply fuse the lung field feature map and semantic feature vector, achieving complementarity between spatial and semantic features. An improved VMamba model incorporates a frequency domain channel attention fusion mechanism, using discrete cosine transform to achieve feature conversion between the spatial and frequency domains, effectively separating high-frequency detail features of lesions from low-frequency redundant features of the background. Combined with the channel attention mechanism of the simplified SE-Net architecture for weight allocation, the response of lesion-related features is strengthened, background interference is suppressed, and feature utilization is significantly improved. The improved VMamba model, combined with a cross-layer attention module and a salient feature suppression strategy, achieves deep fusion and accurate identification of pneumonia imaging features, enhancing the sensitivity of capturing small lesion features. In the calculation of lesion characterization coefficients, the accuracy of coefficient calculation is ensured by using a feature matching graph structure and edge centrality calculation, combined with logarithmic transformation to reduce interference from dense nodes. This provides a reliable data foundation for subsequent lesion area judgment and classification diagnosis. Compared with traditional models, the accuracy and generalization ability of pneumonia classification diagnosis are significantly improved.

[0024] In some embodiments, Figure 2 This is a flowchart of a pneumonia medical assistance system based on an improved VMamba model according to an embodiment of the present invention, as shown below. Figure 2 As shown, the process of dividing the lung image to be diagnosed into raster cells based on the raster cell partitioning algorithm to obtain the raster cell set includes the following steps: Step 1: Perform superpixel segmentation on the lung image to be diagnosed; The SLIC (Simple Linear Iterative Clustering) algorithm is used to set the number of superpixels and compactness parameters to aggregate image pixels into several superpixel blocks, with each superpixel block being an independent pixel aggregation unit. Specifically, parameter initialization sets fixed superpixel parameters for lung grayscale images: Superpixel count: 256 / 512 (adaptively selected based on image size 256×256 or 512×512); Compactness parameter: 20–30 (medical imaging specific range; the higher the value, the closer the superpixel shape is to a rectangle, and the easier it is to rasterize later). Number of iterations: 10 (to ensure a balance between segmentation accuracy and speed).

[0025] Initial cluster centers are generated by uniformly dividing the image into an S×S grid and setting an initial cluster center at the center of each grid. Gradient perturbation is applied to the cluster centers to avoid noise points and strong gradient points at the edges, thus preventing the initial centers from falling in non-lesion areas.

[0026] Pixel similarity is measured using a combination of grayscale distance and spatial distance: Gray-scale distance: the difference in gray-scale values ​​between pixels; Spatial distance: Euclidean distance between pixel coordinates; For each pixel within the lung field region, calculate its similarity to the adjacent cluster centers, and assign the pixel to the most similar superpixel block.

[0027] Iterative optimization and center update repeatedly update the superpixel cluster center (mean vector) until the cluster center no longer moves or the number of iterations is reached; finally, the lung image is segmented into several irregular superpixel blocks with edges that fit the lesion. Each superpixel block is composed of pixels with similar grayscale, texture and spatial location.

[0028] Step 2: Regularize and correct the superpixel blocks; The minimum bounding rectangle method is used to convert irregular superpixel blocks into rectangular grid units, preserving the core feature regions of the superpixel blocks: extracting superpixel contours, performing edge detection on each superpixel block, and obtaining its bounding contour coordinate set. The minimum bounding rectangle is calculated for each superpixel contour, yielding the coordinates of the top-left corner, bottom-right corner, width *w*, and height *h* of the rectangle. Regularization mapping uses the boundary of the minimum bounding rectangle as the grid unit boundary, mapping irregular superpixel blocks into standard rectangular grid units. During the mapping process, all original pixels within the superpixel are preserved, without altering the lesion texture or grayscale features; only the grid shape is standardized.

[0029] Step 3: Remove invalid raster units whose image quality index does not meet the preset requirements, assign unique numbers and pixel feature attributes to the remaining raster units, and finally obtain N raster units based on superpixels. Combine all the raster units into a raster unit set.

[0030] Furthermore, the process of removing invalid raster cells whose image quality index does not meet the preset requirements includes the following steps: Calculate the image quality index of the raster unit : ; ; in, For pixels within a grid unit The grayscale values ​​are mapped to the corresponding feature values ​​in the fuzzy feature plane. , and These are the preset grayscale value thresholds. ; The number of rows within a grid cell. This represents the number of columns within a grid cell; for example, , and The thresholds are 60, 120, and 180, respectively. These thresholds are suitable for most lung images without obvious exposure abnormalities, taking into account the grayscale differentiation of the lung field, lesions, and background, making them the preferred choice for basic application scenarios. An image quality index threshold is applied to determine whether the image quality index of a raster cell exceeds the threshold. If it does, the raster cell is considered valid; otherwise, it is considered invalid. The image quality index threshold can be set to 0.65, which is suitable for routinely acquired lung images in clinical settings, without significant overexposure, underexposure, or motion blur. It serves as the basic threshold for determining the validity of raster cells and can eliminate approximately 35% of raster cells containing invalid information, balancing feature preservation and noise filtering.

[0031] In some embodiments, calculating the lesion characterization coefficient of each grid cell within each management time period specifically includes the following process: Extract image features of the lung images to be diagnosed for each raster unit within each management time period. and load lesion image features The process of acquiring lesion image features is as follows: Pneumonia images of the same type as the lung images to be diagnosed are selected, covering common pneumonia types such as bacterial pneumonia, viral pneumonia, and mycoplasma pneumonia. These images also include pneumonia images of different lesion sizes, locations, and disease stages, with a sample size of no less than 1000 cases to ensure comprehensive feature coverage. Radiologists with associate chief physician titles or above accurately annotate the lesion regions of the sample images, clearly identifying the core and edge regions of the lesion in each image. Background noise, artifacts, and normal lung tissue areas are removed from the images to form gold standard images with lesion region annotations. Following the aforementioned raster unit division algorithm, all annotated gold standard images undergo superpixel segmentation (SLIC algorithm), regularization correction (minimum bounding rectangle method), and invalid raster removal (image quality index judgment) to obtain lesion raster units of the same specifications as the images to be diagnosed, ensuring... Compared with the imaging features to be diagnosed The unit-dimensional matching process first uses grayscale features to initially distinguish lesions from the background (e.g., removing invalid grids with grayscale anomalies and dividing grayscale threshold intervals to differentiate between normal and diseased tissues). Then, it uses texture features to accurately identify and match lesions (e.g., texture aggregation of superpixel blocks and texture feature matching of graph structures). Finally, the two types of features are fused into a unified image feature. and .

[0032] based on and Generate feature matching graph structure ,in, This is a set of nodes representing the feature information of the lung region to be diagnosed in lung imaging features. , For the number of nodes, The edge set representing the feature information of the lung lesion region in the lesion image features. The number of sides is Among them, two adjacent nodes Constructing edges ; The centrality of an edge in a feature matching graph structure is defined as the average of the node centrality measures of two adjacent nodes, as shown in the following formula: ;in, For the edge The centrality of a node is used to represent the influence of importance between two connected nodes. , Represents the measure of node centrality; The lesion matching probability of each edge is calculated based on the centrality of each edge. : set up To mitigate the impact of dense nodes, after normalization, the lesion matching probability of each edge is calculated using a formula. : ; in, yes The maximum value, yes The average value, This is used to limit the maximum probability; to ensure matching accuracy, 0.85 is acceptable.

[0033] Calculate the average value of the lesion matching probability for all edges, and record the average value as the lesion characterization coefficient for each grid cell in each management period.

[0034] In some embodiments, generating a lesion trajectory path within the set of raster cells containing the lung image to be diagnosed, based on the lesion characterization coefficient, specifically includes the following process: Step 1: Obtain the two-dimensional spatial coordinates of each grid cell in the set of grid cells containing the lung image to be diagnosed, based on the medical image spatial analysis system. And the lesion characterization coefficient of each grid cell and The combination forms lesion feature nodes, and the representation of lesion feature nodes is as follows: Among them, the medical image spatial analysis system can be 3DSlicer, which is designed specifically for the visualization, segmentation, registration, and spatial analysis of 3D or 2D medical images and is fully compatible with lung CT / DR images.

[0035] Step 2: Number the lesion feature nodes in the grid cell set, and combine the first and second lesion feature nodes of the grid cell set to form a lesion reference vector. , and the remaining first Lesion feature nodes and the first Each lesion feature node forms a lesion comparison vector. ,in, Calculate separately and Angle between ,angle The calculation formula is as follows: ; Step 3: Put With threshold angle Compare, if the first Individual lesion characteristic nodes Angle greater than the threshold Then it will be arranged in the th order. The lesion feature nodes preceding each lesion feature node are subdivided into groups, and the nodes arranged in the order of the first lesion feature nodes are grouped together. Repeat steps two and three to further subdivide the grid cell set for each lesion feature node preceding it, and record each subdivided group as a segmented path of the lesion trend; where, the threshold angle A threshold of 30° is suitable for common focal pneumonias such as bacterial pneumonia and mycoplasma pneumonia. The spatial spread or development trend of such lesions is relatively gentle, and the vector direction mutation is mostly concentrated above 30°. This threshold can accurately distinguish between "continuous trend" and "turning trend", taking into account the accuracy and completeness of segmented path division. It is the basic threshold for generating lesion trend path.

[0036] Step 4: Combine all lesion trend segments according to the spatial geographical adjacency of the grid cells to obtain the lesion trend path within the set of grid cells containing the lung image to be diagnosed.

[0037] In some embodiments, calculating the lesion characterization index of the grid cells within the management period based on the grid cells along the lesion trend path specifically includes the following process: To obtain the lesion representation coefficient of each grid cell on the lesion trend path in each management period, a rectangular coordinate system is established with the execution time of the management period as the X-axis and the lesion representation coefficient corresponding to the management period as the Y-axis. The lesion representation index curve is plotted in the rectangular coordinate system by plotting points. A lesion representation index line is set, and the part of the lesion representation index curve above the lesion representation index line is marked. The integral value of the marked curve part is calculated, and the integral value is recorded as the lesion representation index of the lesion trend path within the management period. The process of setting the lesion representation index line is as follows: Step 1: Select the grid cell dataset of normal lung tissue samples. From the constructed unified lung image dataset of 21,000+ images, pure normal lung tissue image samples (without pneumonia, nodules, inflammation, etc., a total of 3,000 images) labeled by radiologists are selected. According to the grid cell partitioning algorithm in the technical solution (SLIC superpixel segmentation + minimum bounding rectangle + image quality index removal), the effective grid cell set corresponding to normal lung tissue (a total of about 120,000 effective grid cells) is obtained.

[0038] Step 2: Calculate the baseline characterization coefficients of normal lung tissue grid cells; For the above-mentioned normal lung tissue grid cells, the standard management cycle of clinical diagnosis was simulated (divided into 3-5 management periods, consistent with the diagnostic observation cycle of pneumonia lesions). The lesion characterization coefficient of each grid cell in each period was calculated (following the process of feature extraction → edge centrality calculation → matching probability mean). Then, the characterization coefficients of all normal grid cells and all periods were statistically analyzed. The arithmetic mean of the characterization coefficients of all samples and the standard deviation of the characterization coefficients of all samples were calculated. The sum of the standard deviation and the arithmetic mean was recorded as the constant corresponding to the lesion characterization index line.

[0039] In some embodiments, determining whether a grid cell covering the path of a lesion is a lesion region based on a lesion characterization index specifically includes the following process: Load the lesion characterization index threshold, and determine whether the lesion characterization index exceeds the lesion characterization index threshold. If it does, determine that the grid cells covered on the lesion trend path are lesion areas. If not, determine that the grid cells covered on the lesion trend path are not lesion areas. The lesion characterization index is the integral value of the portion of the lesion characterization index curve above the index line within the management period. It reflects the cumulative intensity of the lesion characterization coefficient of the grid unit exceeding the baseline level within the management period. The larger the value, the more significant the lesion characteristics of the grid unit and the more obvious the development trend. There is no fixed upper limit to its value, and the lower limit is 0 (when no part of the curve is above the index line). In clinical practice, due to the characteristics of lung imaging and the length of the management period, the value is mostly concentrated in the range of 0 to 50 (unit: coefficient time period, which is a composite unit of characterization coefficient and management period). Preferably, the lesion characterization index threshold of 15 (unit: coefficient·time period) is suitable for routine analysis scenarios such as focal pneumonia such as bacterial or mycoplasma pneumonia, and the management period is divided into 3 to 5 time periods (such as collecting images once every 3 days). The characterization coefficient of such lesions shows a steady upward or fluctuating trend. When the cumulative integral value exceeds 15, it can be clearly identified as a lesion area. It takes into account the accuracy of the judgment and the clinical universality, and is the default basic threshold of the system.

[0040] In some embodiments, the lung image portion corresponding to the lesion area is input into a U-Net++ model to complete lung field segmentation, outputting a lung field feature map. Using the lung field feature map as input, a preliminary classification of pneumonia is completed based on an improved InceptionV3 model, and semantic feature vectors of the lung region are extracted. Unified dataset construction.

[0041] Data integration and deduplication: Based on 5288 images, including 1626 cases of COVID-19, 1800 cases of ordinary pneumonia, and 1802 normal controls, the dataset integrates the advantages of the Montgomery County lung field segmentation dataset (138 images with professional lung field segmentation masks) and the pneumonia chest X-ray 4-classification dataset (16933 training images and 4232 test images). Duplicate, blurry, and invalid samples are removed through image feature comparison, resulting in a unified dataset with 21000+ valid samples. At the same time, 1500+ low-quality images from primary care facilities (including blurry, high-noise, and low-contrast samples) are deliberately included to balance data scale, annotation quality, and scene adaptability.

[0042] Format and size standardization: All images are uniformly converted to 256×256 pixel 8-bit grayscale images, and pixel values ​​are normalized by dividing by 255 to eliminate distribution differences caused by different devices and formats; data augmentation operations such as elastic transformation, random flipping, and Gaussian noise addition are performed on the main dataset with a probability of 0.5.

[0043] Annotation completion and standardization: The method of "pre-trained U-Net++ semi-automatic annotation + radiologist manual review" is adopted to supplement all images with lung field segmentation mask (lung_mask) and lesion segmentation mask (lesion_mask). Small lesions with a diameter of 3-5mm are cross-annotated by more than two professional physicians to ensure annotation accuracy and meet the dual requirements of U-Net++ segmentation task and Vmamba fusion optimization.

[0044] Unified labeling system: The original multi-source labels are uniformly mapped to four standard labels—Normal, COVID, Lung_Opacity, and Viral_Pneumonia. Among them, "tuberculosis" in the MontgomeryCounty dataset is classified into Viral_Pneumonia according to image features, and "common pneumonia" in the main dataset is subdivided into Lung_Opacity and Viral_Pneumonia according to lesion texture, which meets the multi-classification requirements of InceptionV3 classification and Vmamba fusion optimization.

[0045] Stratified sampling partitioning: A stratified sampling method with a ratio of 8:1:1 is used to divide the unified dataset into a training set (16800+ images), a validation set (2100+ images), and a test set (2100+ images). This ensures that the proportion of four-category label samples and the proportion of low-quality images at the grassroots level in each subset are consistent with the overall dataset. This provides a unified training and validation benchmark for U-Net++, InceptionV3, and Vmamba, and ensures the fairness and reproducibility of the three-segment link performance comparison.

[0046] The U-Net++ model was built using the PyTorch framework, with 30 training epochs and a batch size of 4. The Adam optimizer (learning rate 0.0003) was selected, along with the ReduceLROnPlateau learning rate scheduler (patience=3, factor=0.5). BCE_Dice_Loss (α=0.5) was used as the loss function. During training, the validation set loss was monitored, and the learning rate was reduced by 50% if the loss did not improve for three consecutive epochs. After training, the model performance was validated on the test set, and the Dice coefficient was 0.967, IoU was 0.938, and Recall was 0.960. The segmentation accuracy of complex edge regions such as the lung apex and hilum was good, and high-quality lung field feature maps were output.

[0047] In some embodiments, an improved InceptionV3 model is constructed by inserting an SA spatial attention module and residual blocks based on a pre-trained InceptionV3 model, and supplementing small lesion samples with a generative adversarial network (GAN). The detailed technical solution for the improved InceptionV3 model is as follows: (I) Specific Structure of the SA Spatial Attention Module This invention inserts an SA spatial attention module after the three convolutional modules and before the pooling layer in the InceptionV3 model. The module adopts a three-level structure of channel compression, spatial modeling, and weight activation, specifically as follows: Channel compression: The input feature map is processed in parallel using global average pooling and global max pooling to compress the C×H×W feature map into a C×1×1 feature vector, capturing the average and extreme value information of the features respectively; where C represents the number of channels of the feature map, and the height H and width W of the feature map. Spatial modeling: The two compressed feature vectors are concatenated into a 2C×1×1 array. The number of channels is reduced to C / 16 by a 1×1 convolution kernel. Then, spatial attention modeling is completed by a 3×3 depthwise separable convolution, and a spatial attention feature map of C / 16×H×W is output. Weighted activation: The number of channels is restored to C by a 1×1 convolution kernel, and a spatial attention weight map between 0 and 1 is generated by the Sigmoid activation function. The weight map is then multiplied pixel by pixel with the original input feature map to enhance the spatial feature response of the lesion region.

[0048] (ii) Number and connection method of residual blocks One residual block is inserted between each of the four Inception modules in the InceptionV3 model, for a total of four residual blocks. All residual blocks adopt a bottleneck residual structure, with the following parameters: The first layer is a 1×1 convolution: the number of convolution kernels is 1 / 4 of the number of input channels, the stride is 1, the padding is 0, and the activation function is ReLU; The second 3×3 convolutional layer has the same number of kernels as the first layer, a stride of 1, padding of 1, and the activation function is ReLU. The third 1×1 convolutional layer has the same number of kernels as the number of input channels, a stride of 1, padding of 0, and no activation function. Residual connection: Identical mapping is used. If the input and output feature map sizes are inconsistent, dimensionality matching is achieved through 1×1 convolution. Finally, the residual branch and the main branch feature map are added together and then activated by ReLU.

[0049] (III) Training strategies for supplementing GAN with small lesion samples Conditional Generative Adversarial Network (CGAN) was used to generate small pneumonia lesion samples with a diameter of 3-5 mm to expand the dataset. The specific training strategy was as follows: Network structure: The generator adopts a deconvolution structure, taking random noise and lesion category labels as input, and outputting a 256×256 image of small lung lesions; the discriminator adopts a convolution structure, taking real or generated lung images and lesion category labels as input, and outputting the image realism and category discrimination results. Loss function: The adversarial loss + L1 pixel loss + lesion feature loss are jointly trained. The adversarial loss adopts WGAN-GP loss, the L1 pixel loss constrains the pixel difference between the generated image and the real image, and the lesion feature loss extracts features through pre-trained InceptionV3 to constrain the feature consistency of the generated lesions. Training parameters: batch size = 16, learning rate = 0.0002, optimizer is Adam, training epochs are 200, and a small lesion sample is generated every 10 epochs. Finally, 2000 lung image samples containing bacterial, viral and COVID-19 small lesions are added to the dataset.

[0050] Using the lung field feature map output by the lesion area image processing module as input, the PyTorch training framework was adopted, with 30 training rounds and a batch size of 8. The SGD optimizer (learning rate 0.001, momentum 0.9) was selected, and the cross-entropy loss function was used. After training, the model performance was validated on the test set. The basic recognition accuracy of the four-class classification of pneumonia reached 88.7%, and 512-dimensional semantic feature vectors were extracted, providing a benchmark for subsequent VMamba model fusion optimization.

[0051] In some embodiments, the construction of an improved VMamba model and accurate identification of pneumonia. VMamba model adaptation: The original VMamba model (Base scale) was built using the PyTorch framework. The feature extraction logic was adjusted for the characteristics of pneumonia images. The lung field feature map and the semantic feature vector were concatenated at the channel level. Adaptive noise reduction processing of median filtering + Gaussian filtering was performed on low-quality images at the grassroots level. All feature values ​​were scaled to the range of [-1,1]. The VMamba model is a sequence modeling framework based on the state space model SSM, which includes the 2D selective scanning module SS2D and the visual state space block VSS (Visual State Space).

[0052] Three major improvement mechanisms are embedded: Frequency domain channel attention fusion: DCT transform, SE-Net channel attention, and IDCT inverse transform modules are embedded between the feature input stage and the SSM (State Space Model) module. The VMamba feature mapping dimension is adjusted to 256 dimensions, and the SSM module gating mechanism parameters are optimized. Cross-layer attention module: A dual-branch attention structure is embedded between the SSM modules of layers 2 and 4 of the VMamba model. The Q / K / V dimensions are set to 128, and residual connections and layer normalization are added. Significant feature suppression strategy: Add a random patch occlusion module, with patch size randomly switching between 16×16 / 32×32, and gradually increase the occlusion probability from 0.3 to 0.5. Introduce L2 regularization as a significant feature penalty term (weight 0.001) in the loss function. Model training: The training epochs were set to 300, and the AdamW optimizer was used (learning rate 0.0001, weight decay 0.01), with a dynamic batch size strategy (batch size=8 for the first 15 epochs, batch size=16 after 15 epochs), and equal sampling was performed according to image quality level. Performance Validation and Diagnostic Results Output: The performance of the improved VMamba model was validated on the test set. The accuracy of the three-class classification of pneumonia reached 98.9%, the sensitivity of COVID identification was 100%, the accuracy of pneumonia identification was 99.44%, the identification rate of lesions with a diameter of 3-5mm was improved by 12%, the false negative rate was reduced by 6%, the processing time of a single image was 2.8 seconds, and the pneumonia classification diagnosis results were output.

[0053] In some embodiments, Figure 3 This is a flowchart of another pneumonia medical assistance system based on an improved VMamba model according to an embodiment of the present invention, as shown below. Figure 3 As shown, the implementation process of frequency domain channel attention fusion includes: Step 1: The feature map after channel-level stitching is transformed from the spatial domain to the frequency domain through discrete cosine transform to separate the high-frequency detailed features of the lesion from the low-frequency redundant features of the background. Step 2: Combining the channel attention mechanism of the simplified SE-Net architecture, weights are assigned to the feature channels after frequency domain transformation to enhance the response of lesion-related feature channels and suppress interference from irrelevant background channels; Step 3: The optimized frequency domain features are mapped back to the spatial domain through inverse discrete cosine transform, and then fused with the original features and input into the SSM module of the VMamba model. Step 4: Adjust the feature mapping dimension of the VMamba model to match the frequency domain feature dimension, optimize the gating mechanism parameters of the SSM module, and improve the sensitivity of capturing small lesion features.

[0054] Among them, the detailed technical solution of the cross-layer cross-attention module (a) Specific design of bi-branch attention The cross-layer attention module is divided into shallow feature branches and deep feature branches. The two branches adopt a symmetrical structure and are customized for the feature dimensions of the VMamba framework, specifically as follows: Shallow feature branch: The 256×64×64 high-resolution feature map output from the second layer SSM module of VMamba is input, and the number of channels is reduced to 128 through a 1×1 convolution, which is used as the attention value. ) and key ( ), capturing detailed features such as the edges and contours of lesions; Deep Feature Branch: The 512×16×16 deep semantic feature map output from the 4th layer SSM module of VMamba is input, and the number of channels is reduced to 128 through 1×1 convolution. After upsampling to 64×64, it is used as the attention query. ), capturing semantic features such as lesion type and lesion extent; Cross-interaction: shallow branching With deep branches Perform matrix multiplication, generate a cross-layer attention weight map through softmax activation, and then combine the weight map with the shallow branch's... Multiplying these together yields an attention feature map that integrates features from both deep and shallow layers.

[0055] (ii) Calculation method of attention weight The cross-layer attention weights are calculated using scaled dot product attention, and the specific formula is as follows: ; in: The query matrix for deep feature branches is 128×4096. The key matrix of the shallow feature branches is 128×4096. The value matrix of the shallow feature branches is (128×4096). Take 128 as / Dimensions.

[0056] (III) Specific locations of residual connectivity and layer normalization Layer normalization (LN): respectively in / / After feature mapping and attention weight calculation, a normalization layer is added, with the normalization dimension being the channel dimension, to avoid gradient explosion during training. Residual connection: The output feature map of the cross-layer attention module is added to the original input feature map of the shallow feature branch by residual addition, and then input into the next layer SSM module after passing through the GELU activation function, so as to ensure the continuity of feature propagation.

[0057] Detailed technical solution of salient feature suppression strategy (a) Rules for selecting areas covered by patches An adaptive occlusion rule based on lung field masking is adopted, which performs random patch occlusion only within the lung field area, avoiding irrelevant background areas such as the chest wall, ribs, and heart. Specifically: Lung field mask constraint: Based on the lung field mask obtained by U-Net++ segmentation, only generate occlusion patches in lung field regions with a mask value of 1 to ensure that the occlusion regions are all valid lung tissue; Lesion region weight: The labeled lesion region is assigned a 0.7 occlusion probability, and the normal region in the lung field is assigned a 0.3 occlusion probability. Large lesion regions are heavily occluded, forcing the model to pay attention to the surrounding small lesions. Patch size and location: The patch size is randomly switched between 16×16 and 32×32, and the location is randomly generated within the lung field mask to ensure that it does not exceed the lung field boundary, and the number of patches masked at one time does not exceed 30% of the lung field area.

[0058] The specific calculation method for the L2 regularization penalty term: In the VMamba model's loss function, a feature channel-level L2 regularization penalty term is introduced to constrain the model's over-response to a single salient feature channel. The specific calculation method is as follows: The overall form of the loss function is: ; For cross-entropy loss, The penalty item weight is 0.001. This is an L2 regularization penalty term; L2 regularization calculation: For the feature map output by the last layer VSS module of VMamba, calculate the sum of squared eigenvalues ​​of each channel, and then calculate the average of all channels. The formula is: ; The eigenvalue of the i-th channel at position (x,y).

[0059] In some embodiments, the workflow of the diagnostic report generation module is as follows: Generate a visual auxiliary diagnostic report by linking the pneumonia knowledge graph. A pneumonia knowledge graph was built based on Neo4j, containing 44,000 entities and 300,000 relations. The py2neo library was used to realize the association and retrieval of model diagnosis results with the knowledge graph. Grad-CAM visualization technology is used to generate lesion identification heatmaps, which show the lesion areas of interest to the model. Integrate pneumonia classification and diagnosis results, lesion heat maps, and knowledge graph retrieval of treatment processes or commonly used drugs and examination suggestions to generate standardized and visualized auxiliary diagnostic reports; The front-end interface is built using Vue, enabling functions such as image uploading, lung field segmentation, pneumonia identification, and report viewing or export. The back-end interface is built using Spring Boot and Flask, supporting dual-mode deployment on a server (Alibaba Cloud ECS) and local (CPU i5-10400 + 8G memory). It was tested in two primary healthcare institutions, and the operation was convenient for doctors, with a diagnostic result consistency rate of over 95% with that of professional physicians.

[0060] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or 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, the 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., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0061] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software 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.

[0062] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0063] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0065] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A pneumonia medical assistance system based on an improved VMamba model, characterized in that, The system includes: The lung imaging acquisition module is used to generate management cycles, divide the management cycle into several management periods, and acquire lung images of patients to be diagnosed within each management period. The lesion region determination module is used to divide the lung image to be diagnosed into grid cells based on the grid cell partitioning algorithm to obtain a set of grid cells, where each grid cell corresponds to a part of the lung image to be diagnosed. The module calculates the lesion characterization coefficient of each grid cell in each management period. Based on the lesion characterization coefficient, a lesion trend path is generated in the set of grid cells where the lung image to be diagnosed is located. Based on the grid cells on the lesion trend path, the lesion characterization index of the lesion trend path in the management period is calculated. Based on the lesion characterization index, it is determined whether the grid cells covered by the lesion trend path are lesion regions. The lesion area image processing module is used to input the lung image part to be diagnosed corresponding to the lesion area into the lung field segmentation based on the U-Net++ model, output the lung field feature map, and use the lung field feature map as input to complete the preliminary classification of pneumonia based on the improved InceptionV3 model and extract the semantic feature vector of the lung area. The pneumonia classification and diagnosis module is used to input the lung field feature map and semantic feature vector into the improved VMamba model after channel-level concatenation. It completes the deep fusion and accurate identification of pneumonia image features through frequency domain channel attention fusion, cross-layer cross attention module and salient feature suppression strategy, and outputs pneumonia classification and diagnosis results. The diagnostic report generation module is used to link the pneumonia classification and diagnosis results with the pneumonia knowledge graph to generate a visual auxiliary diagnostic report.

2. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, The process of dividing the lung image to be diagnosed into raster cells based on the raster cell partitioning algorithm to obtain the raster cell set includes the following steps: Superpixel segmentation of lung images to be diagnosed: The SLIC algorithm is used, and the number of superpixels and compactness parameters are set to aggregate the image pixels into several superpixel blocks, each of which is an independent pixel aggregation unit; Regularization correction of superpixel blocks: The minimum bounding rectangle method is used to convert irregular superpixel blocks into rectangular grid units, while preserving the core feature regions of the superpixel blocks; Invalid raster units that do not meet the preset requirements for image quality index are removed. The remaining raster units are assigned a unique number and pixel feature attributes, resulting in N raster units based on superpixels. All raster units are then combined into a raster unit set.

3. The pneumonia medical assistance system based on the improved VMamba model according to claim 2, characterized in that, The process of removing invalid raster cells whose image quality index does not meet the preset requirements includes the following steps: Calculate the image quality index of the raster unit ; Load the image quality index threshold, and determine whether the image quality index of the raster cell exceeds the image quality index threshold. If it does, the raster cell is considered valid; otherwise, the raster cell is considered invalid.

4. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, The calculation of the lesion characterization coefficient for each grid cell in each management period includes the following steps: Extract image features of the lung images to be diagnosed for each raster unit within each management time period. and load lesion image features ; The lesion matching probability of each edge is calculated based on the centrality of each edge. ; Calculate the average value of the lesion matching probability for all edges, and record the average value as the lesion characterization coefficient for each grid cell in each management period.

5. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, Generating a lesion trajectory path within the raster cell set containing the lung image to be diagnosed based on the lesion characterization coefficient specifically includes the following process: Step 1: Obtain the two-dimensional spatial coordinates of each grid cell in the set of grid cells containing the lung image to be diagnosed, based on the medical image spatial analysis system. And the lesion characterization coefficient of each grid cell and The combination forms lesion feature nodes, and the representation of lesion feature nodes is as follows: ; Step 2: Number the lesion feature nodes in the grid cell set, and combine the first and second lesion feature nodes of the grid cell set to form a lesion reference vector. , and the remaining first Lesion feature nodes and the first Each lesion feature node forms a lesion comparison vector. ,in, Calculate separately and Angle between ,angle The calculation formula is as follows: ; Step 3: Put With threshold angle Compare, if the first Individual lesion characteristic nodes Angle greater than the threshold Then it will be arranged in the th order. The lesion feature nodes preceding each lesion feature node are subdivided into groups, and the nodes arranged in the order of the first lesion feature nodes are grouped together. Repeat steps two and three to subdivide the grid cell set for each lesion feature node before the node, and record each subdivided group as a segmented path of the lesion trend. Step 4: Combine all lesion trend segments according to the spatial geographical adjacency of the grid cells to obtain the lesion trend path within the set of grid cells containing the lung image to be diagnosed.

6. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, The calculation of the lesion characterization index of the grid cells within the management period based on the lesion trend path specifically includes the following process: Obtain the lesion characterization coefficient of each grid cell on the lesion trend path in each management period. Establish a rectangular coordinate system with the execution time of the management period as the X-axis and the lesion characterization coefficient corresponding to the management period as the Y-axis. Plot the lesion characterization index curve in the rectangular coordinate system by plotting points. Set the lesion characterization index line and mark the part of the lesion characterization index curve that is above the lesion characterization index line. Calculate the integral value of the marked curve part and record the integral value as the lesion characterization index of the lesion trend path within the management period.

7. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, Determining whether a lesion's path is covered by a raster cell based on its lesion characterization index includes the following process: Load the lesion characterization index threshold, and determine whether the lesion characterization index exceeds the lesion characterization index threshold. If it does, determine that the grid cells covered on the lesion trend path are lesion areas. If not, determine that the grid cells covered on the lesion trend path are not lesion areas.

8. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, The implementation process of frequency domain channel attention fusion is as follows: The feature map after channel-level stitching is transformed from the spatial domain to the frequency domain by discrete cosine transform, separating the high-frequency detailed features of the lesion from the low-frequency redundant features of the background. By combining the channel attention mechanism of the simplified SE-Net architecture, weights are assigned to the feature channels after frequency domain transformation to enhance the response of lesion-related feature channels and suppress interference from irrelevant background channels. The optimized frequency domain features are mapped back to the spatial domain through inverse discrete cosine transform, and then fused with the original features before being input into the SSM module of the VMamba model. Adjust the feature mapping dimension of the VMamba model to match the frequency domain feature dimension, optimize the gating mechanism parameters of the SSM module, and improve the sensitivity of capturing small lesion features.

9. The pneumonia medical assistance system based on the improved VMamba model according to claim 1, characterized in that, The improved VMamba model incorporates a frequency domain channel attention mechanism, a cross-layer cross attention module, and a salient feature suppression strategy. The VMamba model is a sequence modeling framework based on the state space model SSM, which includes a 2D selective scan module SS2D and a visual state space block VSS.

Images