A method and device for analyzing contrast images of digital diagnosis and treatment equipment

Abstract

The present application provides a method and device for analyzing contrast images of digital diagnosis and treatment equipment, including: predicting the progression of the disease based on quantitative indicators of lesion volume characteristic parameters, density statistical parameters, and the density difference of tissues around the lesion; determining whether the prediction result of the disease progression reaches a preset risk threshold range. If the prediction result is within the preset risk threshold range, it is determined that the disease progression is in a risk state, and a high-risk warning message is generated; according to the judgment result of the disease progression prediction, combined with the fitting data of the lesion volume and density changing over time, the severity of the disease is judged, the lesion progression process is staged and graded, and a quantitative disease diagnosis report is generated.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and particularly to a method and device for analyzing contrast images of digital diagnosis and treatment equipment. Background Art

[0002] When analyzing a sequence of contrast images, affected by different increment granularities, such as pixel level, voxel level, and region level, at the pixel level, how to accurately define the lesion area is a key technical step; first, how to explore image registration technology to achieve spatial alignment of the lesion areas at different time points and eliminate position offsets introduced by factors such as patient body position changes and respiratory movements. Furthermore, how to obtain the change data of the lesion area volume over time based on the fusion of voxel level and the number of pixels is a data processing problem to be solved; at the same time, since the pixel gray value is related to the density of the lesion area, how to obtain the change data of the lesion area density over time is also a major difficulty; in addition, after obtaining quantitative indicators such as the number of pixels, average density, and gray value distribution of the lesion area at each time point, how to establish the correlation between time series data, mine the lesion progression law, and further analyze the morphological characteristics of the lesion area is an important step to depict the morphological complexity and invasion range of the lesion from multiple angles. Finally, how to predict the development of the disease condition and judge the risk of the disease condition to generate a disease diagnosis report is an important objective that this solution needs to handle. Summary of the Invention

[0003] The present invention provides a method for analyzing contrast images of digital diagnosis and treatment equipment, mainly including:

[0004] Obtain a sequence of contrast images collected from a target patient at different time points. For the contrast images at different time points, segment the lesion area respectively from the pixel-level features, voxel-level features, and region-level features to obtain the number of pixels, average density, and gray value distribution of the lesion area at different time points;

[0005] Extract and describe the lesion area features at different time points, generate feature descriptors, perform a ratio test on the feature descriptors, screen out feature matching points, calculate the transformation matrix between the lesion areas through the feature matching points, and obtain the changes between the lesion areas at different time points;

[0006] After alignment registration, obtain the change data of the lesion volume over time according to the number of pixels of the lesion area at each time point, and at the same time calculate the average density of the lesion area according to the pixel gray value within the lesion area to obtain the change data of the lesion density over time;

[0007] By performing data fitting on the data of the change in lesion volume over time and the data of the change in lesion density over time, analyzing the laws of the change in lesion volume and density over time, and obtaining volume dynamic characteristic parameters and density statistical parameters to depict the dynamic change laws during the lesion progression;

[0008] According to the dynamic change laws during the lesion progression, calculate the irregularity of the lesion shape and the smoothness of the boundary, evaluate the morphological complexity and boundary clarity of the lesion, and at the same time compare the density difference between the lesion area and the surrounding normal tissues to obtain a quantitative index of the tissue density difference around the lesion;

[0009] Predict the disease progression based on the lesion volume characteristic parameters, density statistical parameters, and the quantitative index of the tissue density difference around the lesion;

[0010] Judge whether the prediction result of the disease progression reaches the preset risk threshold range. If the prediction result is within the preset risk threshold range, then judge the disease progression as a risk state and generate a high-risk warning message;

[0011] According to the judgment result of the prediction of the disease progression, combined with the fitting data of the change in lesion volume and density over time, judge the severity of the disease, stage and grade the lesion progression process, and generate a quantitative disease diagnosis report.

[0012] The present invention provides a contrast image analysis device for a digital diagnosis and treatment device, mainly including:

[0013] An image processing and feature extraction module for acquiring and analyzing a sequence of contrast images at different time points;

[0014] A lesion area change analysis module for calculating the changes between lesion areas at different time points;

[0015] A lesion dynamic feature analysis module for analyzing the laws of the change in lesion volume and density over time;

[0016] A lesion morphology evaluation module for evaluating the morphological complexity and boundary clarity of the lesion;

[0017] A disease prediction and risk assessment module for predicting the disease progression and generating a high-risk warning message;

[0018] A disease diagnosis and grading module for judging the severity of the disease and generating a quantitative diagnosis report.

[0019] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0020] The present invention discloses a method and device for analyzing contrast images of digital diagnosis and treatment equipment. The method obtains a sequence of contrast images of a patient at different time points, segments lesions at the pixel, voxel, and region levels, extracts features and performs registration. By calculating the data of the changes in the volume and density of the lesions over time, a dynamic change rule model is established, and volume dynamic characteristic parameters and density statistical parameters are obtained. Combining indicators such as the morphological complexity of the lesions, the clarity of the boundaries, and the density difference of the surrounding tissues, and inputting them into a pre-established disease progression prediction model, the prediction of disease progression is realized. The method can also generate high-risk warning information according to the prediction results, stage and grade the process of lesion progression, and generate a quantitative disease diagnosis report. Through multi-dimensional feature analysis and dynamic modeling, the present invention improves the accuracy and reliability of predicting lesion progression, and provides an important reference for clinical diagnosis and treatment decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flowchart of a method for analyzing contrast images of digital diagnosis and treatment equipment according to the present invention.

[0022] Figure 2 It is a schematic diagram of a method for analyzing contrast images of digital diagnosis and treatment equipment according to the present invention.

[0023] Figure 3 It is another schematic diagram of a method for analyzing contrast images of digital diagnosis and treatment equipment according to the present invention.

[0024] Figure 4 It is a schematic structural diagram of a device for analyzing contrast images of digital diagnosis and treatment equipment according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this specification.

[0026] Such as Figures 1-4 , a method for analyzing contrast images of digital diagnosis and treatment equipment in this embodiment may specifically include:

[0027] S101. Obtain a sequence of contrast images collected from a target patient at different time points. For the contrast images at different time points, segment the lesion area respectively from the pixel-level features, voxel-level features, and region-level features to obtain the number of pixels, average density, and gray value distribution of the lesion area at different time points.

[0028] Perform Gaussian smoothing on the images according to the contrast imaging acquisition sequence, and use an edge tracking tool based on pixel gray-scale gradient to obtain the lesion boundary contour map; select the point with the largest gray value as the seed point for the lesion boundary contour map, and perform region filling according to the rule that the difference in pixel density in the eight-neighborhood is less than a preset threshold to obtain the lesion area density distribution map; use the lesion area density distribution map to establish three-dimensional point cloud data with a preset voxel size, remove the noise points in the point cloud data to obtain three-dimensional feature data; extract the centroid coordinates of the lesion area from the three-dimensional feature data to establish a registration reference point, and perform nearest neighbor pairing on the lesion boundary points according to a preset distance threshold to obtain the lesion area volume data and the gray-scale distribution histogram.

[0029] Specifically, the original image sequence of the lesion area is obtained according to the contrast image acquisition interval. Each image in the image sequence is enhanced according to the Gaussian smoothing operator. The enhanced image uses an edge tracking tool based on pixel gray-scale gradient to extract the lesion boundary. The gray-scale histogram of the lesion area is extracted by setting the binarization threshold to 128. The initial contour map of the lesion area is calculated from the gray-scale histogram. The point with the maximum gray-scale value in the center area of the lesion is selected as the seed point from the initial contour map, and the growth radius is set to 8 pixel points. Region filling is performed based on the dilation rule that the difference in pixel density in the 8-neighborhood is less than 10%, and the density distribution map of the lesion area is obtained. The lesion boundary is defined within the range of plus or minus 20% of the density mean value to obtain the corrected two-dimensional segmentation result map. For the corrected two-dimensional segmentation result map, the voxel size is set to 2×2×2 cubic millimeters, and the three-dimensional spatial coordinates are obtained from the corresponding positions of adjacent slice images. Three-dimensional point cloud data is constructed based on the gray-scale values of the voxels. The minimum voxel bounding box is used to screen the point cloud data, and the number of voxels and the average density value of the voxels are extracted to form three-dimensional feature data. For the three-dimensional feature data obtained at different time points, the centroid coordinates of the lesion area are extracted to establish the registration reference points. Based on the nearest neighbor principle, the lesion boundary points are paired, and the maximum distance threshold between the point pairs is set to 3 millimeters. The volume of the lesion area is calculated for the registered data, and the number of pixels, the average density value, and the gray-scale distribution histogram of the lesion area are statistically obtained from the registration results. The contrast image sequence corresponding to the lesion area has density change characteristics at different time points. The image is enhanced by the Gaussian smoothing operator, with the standard deviation set to 1.5 and the kernel window size to 3×3, which can effectively reduce the influence of image noise, maintain the lesion edge features, and improve the accuracy of edge tracking. In the contrast images of liver tumors, there is a density difference between the lesion boundary and the surrounding tissues. The edge tracking tool identifies the boundary position based on the change in pixel gray-scale gradient, and when the gray-scale difference is greater than 30, it is determined as a boundary point. The gray-scale histogram of the lesion area reflects the tissue density distribution characteristics, with the gray-scale value range of 0-255, and the bimodal distribution pattern corresponding to the lesion parenchymal area and the edge transition area. The selection of the seed point affects the result of region growth. The point with the maximum gray-scale value in the center area of the lesion is selected as the starting position to ensure the expansion starts from the lesion parenchymal tissue. In the contrast images of pulmonary nodules, the 8-neighborhood dilation rule is based on the density similarity criterion, and adjacent pixels with a density difference less than 10% are included in the target area to achieve the adaptive segmentation of the lesion area. The lesion boundary correction uses a threshold range of plus or minus 20% of the density mean value to exclude the mis-segmented background tissue and improve the accuracy of the segmentation boundary. In the three-dimensional spatial reconstruction, the voxel size is set to 2×2×2 cubic millimeters to adapt to the slice thickness resolution of CT images and achieve the precise expression of the lesion morphological features. In the contrast sequence of brain tumors, adjacent slice images establish spatial association through the corresponding position coordinates to construct a three-dimensional point cloud data set. The minimum voxel bounding box screens and removes the free noise points, retains the continuous and complete lesion tissue area, and calculates the number of voxels to reflect the change in lesion volume.The registration of lesion regions at different time points is based on the principle of feature point matching. The centroid coordinates are used as reference points to establish a spatial mapping relationship. For example, in the dynamic contrast-enhanced scanning of breast masses, the boundary points of the lesions are paired using the nearest neighbor criterion, and a point pair distance threshold of 3 mm is set to control the matching error, and the corresponding lesion morphology is achieved through registration. After registration, quantitative measurements are performed. The volume change rate of the lesion region is calculated, and the average density value is statistically analyzed to reflect the change in tissue blood supply. The gray distribution histogram describes the heterogeneity characteristics inside the lesion. In the follow-up evaluation of multiple metastatic tumors, a series of contrast images at multiple consecutive time points reflect the dynamic change process of the lesions. By quantitatively analyzing the characteristic parameters of the lesion regions at different time phases, the morphological and density changes of the lesions at adjacent time points are compared, and the treatment effect is evaluated in combination with clinical diagnosis and treatment information. Different types of tumor lesions have specific contrast characteristics. The shape of the density change curve reflects the state of tissue angiogenesis, and the coefficient of variation of the gray distribution characterizes the degree of uniformity of the internal structure of the lesion.

[0030] By extracting the region-level features of the contrast images at different time points, calculating the feature quantization values of the lesion regions, comparing the preset feature quantization threshold with the feature quantization values of the lesion regions, the target region is preliminarily segmented; the selection conditions of the preset seed points are set, and several seed points are selected from the target region according to the selection conditions, and the segmentation region is expanded starting from the seed points until the preset expansion condition is reached to achieve the final segmentation.

[0031] For the series of contrast images, a Gaussian smoothing kernel function is used for noise reduction processing, and the candidate region is obtained by projection calculation according to the gray value distribution matrix of the denoised image; according to the local gray value mean and variance of the candidate region, the position of the seed point is obtained through the preset uniform region determination condition, and the region expansion is performed from the seed point using the gray difference threshold to obtain the expanded segmentation region; for the expanded segmentation region, the edge contour line is extracted, the feature points are obtained by calculating the edge gradient data within the detection window, and the region segmentation is performed according to the curvature of the feature points until the area converges to obtain the final segmentation region; the shape feature parameters are extracted from the final segmentation region, and the difference in the regional morphological features is judged by the feature point pairing distance threshold. If the difference in the shape feature parameters is within the preset range and the feature point pairing rate exceeds the preset threshold, it is determined as the corresponding region.

[0032] Specifically, the original image data is obtained from the contrast image sequence at different time points. For the original image data, a Gaussian smoothing kernel function with a window size of 3×3 and a standard deviation of 1.5 is used for noise reduction. The regional gray value distribution matrix is extracted from the denoised image. The characteristic quantization value of the lesion area is calculated through integral projection in the horizontal and vertical directions. If the quantization value is within the range of plus or minus 15% of the preset threshold, this area is taken as the candidate area, and the effectiveness of the area is judged by comparing with the preset quantization threshold to obtain the preliminary segmentation area. The local gray value mean and variance are calculated within the preliminary segmentation area. It is set that the gray standard deviation within a 5×5 pixel block is less than 10 as the judgment condition for the uniform area. The position with the largest gray value is selected from the uniform area as the starting seed point. It is set that the gray difference between adjacent pixels within the 8-neighborhood of the seed point is less than 20% as the expansion growth threshold, and a gray-value-based regional expansion tool is used to expand outward from the seed point position to generate the expanded segmentation area. For the expanded segmentation area, the regional edge contour line is extracted. The detection window size is set to 7 pixels, and the first-order differences of the gray values within the window in the vertical and horizontal directions are calculated to obtain the edge gradient data. The edge points with a gradient value greater than 30 are selected as the feature points, and an edge linked list is constructed based on the 8-neighborhood connectivity of the feature points. The edge point with the largest local curvature is obtained from the linked list as the new seed point, and the segmentation continues according to the expansion growth threshold until the change rate of the segmentation area between two adjacent times is less than 5%. Three shape feature parameters, namely area, perimeter, and roundness, are extracted from the final segmentation area. The morphological features of the target areas at different time points are compared. Based on the nearest neighbor criterion, the edge feature points are paired. The maximum distance threshold between the feature point pairs is set to 3 pixels. If the difference in the shape feature parameters between the two areas is within 10% and the feature point pairing rate is greater than 80%, it is determined as the corresponding area, and the automatic registration of the target areas at different time points is completed. During the acquisition process of medical contrast images, the lesion area shows a specific gray change law at different time points. The random noise interference is reduced by the Gaussian smoothing kernel function. In the dynamic enhanced scan of the liver, a 3×3 window is used in combination with a standard deviation parameter of 1.5 to achieve image smoothing while maintaining the edge features. During the quantification of the lesion area features, the integral projection in the horizontal and vertical directions reflects the regional gray distribution characteristics. The gray distribution of normal liver parenchymal tissue is uniform, and the lesion area shows local gray abnormality. When the deviation of the quantization value exceeds 15%, it indicates the presence of a lesion. The local regional gray statistical features correspond to the tissue density distribution. In pancreatic space-occupying lesions, a gray standard deviation within a 5×5 pixel block less than 10 indicates uniform tissue structure. Selecting the position with the largest gray value as the seed point ensures the expansion starting from the center of the lesion. Tumors rich in blood vessels show obvious enhancement in the arterial phase. The gray difference between adjacent pixels less than 20% corresponds to the tissue density continuity, and the segmentation area is gradually expanded based on the gray similarity principle. For the extraction of the lesion edge features, a 7-pixel-sized detection window is used. In the dynamic enhancement of pulmonary nodules, an edge gradient value exceeding 30 indicates the tissue interface, and a complete edge contour is constructed through 8-neighborhood connectivity analysis.The edge curvature reflects the change in the contour shape. The points with the maximum local curvature often correspond to the morphological features of the lesion. Using these points as new seed points for further segmentation can improve the accuracy of boundary localization. If the change rate of the segmented area is less than 5% after two consecutive segmentations, it indicates that the segmented area tends to be stable. Comparing the morphological features of the lesion area at different time points, for example, in the mammography sequence of breast masses, the area represents the size of the lesion, the perimeter reflects the complexity of the boundary, and the roundness describes the regularity of the shape. In the nearest neighbor feature point pairing, a distance threshold of 3 pixels is suitable for slight deformations of the lesion, a tolerance range of 10% for the shape feature parameters is considered for the density dynamic changes during the imaging process, and a feature point pairing rate of 80% ensures the accuracy of the corresponding relationship of the lesion area. In the tracking and evaluation of multiple metastatic tumors, the dynamic changes of the lesions are recorded by angiographic images at multiple consecutive time points. By quantitatively analyzing the characteristic parameters of the lesion area at different time phases, the laws of the morphological and density changes of the lesions are demonstrated. Different types of tumor lesions have specific angiographic manifestations. The shape of the density change curve corresponds to the characteristics of tissue blood supply, the changes in the regional morphological parameters reflect the invasive growth characteristics of the lesions, and the results of shape feature registration support the evaluation of the lesion evolution process. In the segmentation of heterogeneous tumor regions, the combination of edge gradient and curvature features is used to improve the accuracy of the segmentation boundary, and the multiple seed point expansion strategy is suitable for complex lesion shapes.

[0033] S102. Extract and describe the characteristics of the lesion area at different time points, generate feature descriptors, perform a ratio test on the feature descriptors, screen out the feature matching points, calculate the transformation matrix between the lesion areas through the feature matching points, and obtain the changes between the lesion areas at different time points.

[0034] Use the Gaussian blur tool to smooth the original image of the lesion area, extract the edge contour line from the smoothed image through the edge detection operator, calculate the gradient histogram according to the edge contour line to obtain the feature descriptor; calculate the Euclidean distance value between the corresponding positions at different time points according to the feature descriptor, and obtain the initial matching point set through the comparison result of the Euclidean distance value and the preset distance threshold; randomly select matching point pairs from the initial matching point set to calculate the affine transformation matrix, and count the number of inliers that satisfy the affine transformation matrix to obtain the optimal transformation parameters; perform spatial mapping on the lesion area image according to the optimal transformation parameters, use the bilinear interpolation tool to resample the transformed image, and judge the registration result by calculating the mutual information value of the images before and after resampling.

[0035] Specifically, according to the original images of the lesion area at different time points, the Gaussian blur tool with a window size of 3×3 and a standard deviation of 1.5 was used to smooth the images. The edge contour of the lesion area was extracted by the Sobel edge detection operator. The grayscale gradient amplitude and direction in the 8-neighborhood of the edge point were calculated. The gradient histogram in the 16×16 pixel window was extracted from each edge point. The gradient distribution was statistically calculated according to the 8-directional intervals. The gradient histogram value was normalized by dividing the histogram sum to obtain a 128-dimensional feature descriptor. For the normalized feature descriptors, the Euclidean distance between the feature descriptors at the corresponding positions at different time points was calculated, and the distance threshold was set to 0.6. For each feature point, the two candidate matching points with the smallest distance were selected. If the ratio of the minimum distance to the second smallest distance was less than 0.8, the matching point pair was retained. All feature points were traversed and matched by the point pair matching tool based on the distance threshold to generate an initial matching point set. Three pairs of matching points were randomly selected from the initial matching point set, and the corresponding affine transformation matrix was calculated. The number of inliers that satisfied the transformation matrix was counted. The maximum number of iterations was set to 100, and the inlier threshold was 3 pixels. The random sampling process was repeated until the number of inliers was greater than 60% or the maximum number of iterations was reached. The transformation matrix with the largest number of inliers was selected as the optimal solution, and the rotation angle, translation vector, and scaling factor were calculated from the transformation matrix. The images of the lesion area at different time points were spatially mapped according to the optimal transformation matrix, and the bilinear interpolation tool with a 16×16 pixel window was used to resample the transformed images. The mutual information values ​​of the images before and after resampling were calculated. If the mutual information value was greater than the preset threshold of 0.8, the registration was determined to be valid. The overlap coefficient of the lesion area was calculated from the registered image, and the area ratio and boundary distance of the registered area were obtained. In medical image processing, the lesion area shows density and morphological changes at different time points, and accurate feature extraction is crucial for registration. Gaussian blurring uses a 3×3 window with a 1.5 standard deviation parameter. In the dynamic enhanced scanning of the liver, pixel-level smoothing is used to reduce random noise interference and maintain the edge structural characteristics of the lesion. The sobel operator is directional sensitive to edge detection. In the CT image of lung nodules, the horizontal and vertical gradient operators are used to jointly extract the edge contour, and the 8-neighborhood gradient statistics reflect the local structural characteristics. The 16×16 pixel window contains sufficient local feature information, the 8-directional gradient histogram describes the edge distribution law, and the normalization process eliminates the influence of image brightness and contrast changes. In the feature matching process, the Euclidean distance measures the similarity between feature descriptors. In the follow-up of pancreatic space-occupying lesions, the distance threshold of 0.6 distinguishes similar and non-similar feature points. The minimum distance ratio test excludes fuzzy matching, and the ratio threshold of 0.8 strikes a balance between ensuring matching accuracy and matching quantity. As shown in the dynamic enhanced sequence of breast masses, the distance-based point pair matching gives priority to corresponding points with high local feature similarity to form a stable initial matching set.The affine transformation matrix is solved by constructing an equation system with 3 pairs of matching points. In the tracking and evaluation of multiple metastatic tumors, 100 iterations of sampling improve the probability of searching for the optimal solution. The inlier threshold of 3 pixels allows for slight deformation. The inlier ratio requirement of 60% ensures the statistical reliability of the transformation matrix. The rotation angle and translation vector describe the change in the spatial pose of the lesion, and the scaling factor reflects the volume change. In the evaluation of the registration result, bilinear interpolation in a 16×16 pixel window maintains the continuity of the image gray level. In the enhanced scan sequence of gliomas, a mutual information value above 0.8 indicates a reliable registration result. The overlap coefficient of the lesion area quantifies the registration accuracy, the area ratio reflects the volume change, and the boundary distance describes the degree of contour matching. Different types of tumor lesions show specific morphological changes during the treatment process, and the registration result supports the analysis of the lesion evolution process. In the extraction of heterogeneous tumor region features, multi-scale gradient features improve the expression ability of descriptors, and the combination of local feature matching and global transformation constraints ensures the accuracy of the registration result. In dynamic contrast-enhanced imaging, the gray level change in the lesion area reflects the characteristics of tissue blood supply, and the change in morphological parameters corresponds to the characteristics of infiltrative growth. The image sequence at consecutive time points records the dynamic process of the lesion. By quantitatively analyzing the change in the characteristic parameters of the lesion area and combining with clinical follow-up information, the treatment effect is evaluated. The construction of feature descriptors and the matching strategy adapt to the morphological characteristics of different types of lesions, and the verification of the registration result ensures the reliability of the lesion evolution analysis.

[0036] S103. After performing alignment registration, according to the number of pixels in the lesion area at each time point, the change data of the lesion volume over time is obtained. At the same time, according to the pixel gray level values in the lesion area, the average density of the lesion area is calculated to obtain the change data of the lesion density over time.

[0037] The boundary contour line of the lesion is obtained by using the region growing tool based on seed points. The lesion volume value is obtained by converting according to the number of connected region pixels within the contour line and the image resolution. For the image region within the lesion boundary contour line, histogram equalization tool is used for enhancement processing. The gray level value matrix is extracted from the enhanced image and the regional average density value is obtained through projection integral operation. According to the lesion volume value and the regional average density value, time series data is constructed. The abnormal points exceeding the preset value in the time series data are corrected by linear interpolation to obtain a normalized volume density change sequence. For the normalized volume density change sequence, the change rate is obtained by calculating the difference between adjacent time points. The mean and standard deviation of the change rate are used to construct the lesion progression feature parameter set, and the feature parameter set is normalized to obtain the standardized lesion progression feature.

[0038] Specifically, according to the lesion area images at different time points after registration, the lesion boundary contour line is extracted through a region growing tool based on seed points. The seed point position is set at the location with the maximum regional gray value, and the growth threshold is that the gray value difference of neighboring pixels is less than 15%. If two pixel points are 8-neighborhood connected and the gray value difference is less than the threshold, they are determined to belong to the same region. Calculate the number of connected region pixels within the contour line, convert the actual volume of the lesion according to the image resolution of 0.5 mm per pixel, and record the volume change of the lesion over time in the format of timestamp plus volume value. For the pixel points within the lesion area, the contrast-limited histogram equalization tool is used to enhance the image. The contrast limit threshold is set to 0.02, and the block size is 8×8 pixels. The gray value matrix within the lesion area range is extracted from the enhanced image, and the regional average density value is calculated through horizontal and vertical direction projection integration. A time series data table is constructed by combining time series information. The volume and density values in the time series data are processed by 5-point median filtering. The standard deviation of the data within an 11-data-point sliding window is calculated. If a single-point value exceeds 2 times the standard deviation, it is marked as an outlier. The linear interpolation of the previous and subsequent time points is used to correct the outlier data, and the data is normalized with a normalization coefficient of 0.05 to obtain a standardized volume density change sequence. The difference between adjacent time points is extracted from the standardized sequence as the change rate, and the mean and standard deviation of the change rate are calculated to construct a feature dataset including volume change rate, density change rate, volume standard deviation, and density standard deviation. The data is normalized according to the numerical distribution range of the feature data to obtain the normalized lesion progression feature parameters. In medical image analysis, the volume and density changes in the lesion area reflect the disease progression status. Accurately extracting regional features is crucial for evaluating the treatment effect. Selecting the seed point at the location with the maximum regional gray value ensures growth starting from the lesion parenchymal tissue. In the dynamic enhanced scan of the liver, the 15% gray value difference threshold adapts to the gradual change characteristics of tissue density. Using 8-neighborhood connectivity determination considers the spatial position relationship of pixels and realizes the integrity segmentation of the region. The 0.5 mm per pixel resolution setting is applicable to conventional CT scans and supports the quantitative calculation of the lesion volume. The contrast-limited histogram equalization shows good results in the dynamic enhancement of breast masses. The 0.02 contrast limit threshold avoids over-enhancement artifacts, and the 8×8 pixel block size optimizes the gray distribution within the local area. The horizontal and vertical direction projection integration reflects the regional density distribution characteristics. For example, in the evaluation of pancreatic space-occupying lesions, the shape of the integration curve corresponds to the tissue density uniformity. The time series data table records the dynamic changes during the follow-up of the lesion and supports the analysis of the disease progression. Discrete noise points are eliminated through 5-point median filtering. In the tracking and evaluation of multiple metastatic tumors, the 11-point sliding window statistics capture the data change trend. The 2-fold standard deviation threshold identifies abnormal fluctuations, and linear interpolation supplements the data to maintain the sequence continuity. The 0.05 normalization coefficient maps parameters with different dimensions to a unified numerical interval, facilitating the comprehensive analysis of feature parameters.The calculation of the difference between adjacent time points reflects the progression rate of the lesion. For example, in the follow-up evaluation of gliomas, the volume change rate corresponds to the tumor growth state, and the density change rate reflects the change in tissue characteristics. The mean and standard deviation of the change rate describe the stability of the lesion progression. Normalization of the characteristic data eliminates the difference in parameter scales and constructs a quantitative index system for lesion progression. In the analysis of lesion temporal characteristics, volume and density parameters complement each other to jointly depict the law of disease progression. For example, in the monitoring of the radiotherapy process of lung cancer, volume changes reflect tumor burden, and density changes indicate necrosis and liquefaction. Follow-up data at multiple consecutive time points record the treatment response process, and the efficacy is evaluated by quantitatively analyzing the changes in lesion characteristic parameters. Different types of tumor lesions present specific progression patterns, and the volume-density change curve describes the histopathological changes. Standardized characteristic parameters support the prediction of lesion progression and provide data support for clinical treatment decisions. In the analysis of heterogeneous tumors, the combination of local region feature extraction and global parameter statistics accurately depicts the lesion progression characteristics. The continuous processing and outlier correction of follow-up data ensure the reliability of characteristic parameters, and normalization processing realizes the unified expression of multi-dimensional features. In the construction of lesion progression characteristic parameters, static morphological features and dynamic change indicators complement each other to form a complete quantitative evaluation system.

[0039] According to the number of pixels in the lesion area at each time point, combined with the voxel-level features of the contrast image, where the voxel-level features include voxel gray value, voxel spatial position, and voxel shape and size, determine the number of pixels and voxel volume at each time point, calculate the total volume of the lesion, record the total volume of the lesion at each time point, and obtain the change data of the lesion volume over time.

[0040] The contrast image data is denoised using a Gaussian smoothing kernel function, the smoothed image is binarized using a fixed threshold, and the voxel set of the initial lesion area is obtained through a region growing algorithm; the space of the initial lesion area is divided using a three-dimensional grid, the voxels are classified according to the Euclidean distance from the voxel to the center point of the grid, and the connectivity is judged by the distance and gray value difference between adjacent grid cells to obtain the three-dimensional grid connectivity map of the lesion area; the boundary point set is extracted from the three-dimensional grid connectivity map, the boundary box is constructed according to the maximum and minimum coordinate values of the boundary points in three directions, and the maximum diameter and volume of the lesion are obtained by calculating the diagonal length of the boundary box and counting the total number of voxels in the connected domain; a time series data table is constructed for the lesion volume, and the change data of the lesion volume over time is obtained through linear interpolation.

[0041] Specifically, according to the contrast image data, the Gaussian smoothing kernel function is used to denoise the image. The kernel window size is set to 3×3 pixels and the standard deviation is 1.5. The gray value range of the voxels is extracted from the smoothed image and a gray histogram is established. The voxels are binarized based on a fixed threshold of 128. The seed point position is set at the location with the maximum regional gray value, and region growing is performed according to the criterion that the gray value difference of neighboring voxels is less than 10%, to obtain the voxel set of the initial lesion region. For the initial lesion region, a three-dimensional grid with a size of 2×2×2 cubic millimeters is set to divide the space, and the absolute coordinate values of the center points of each grid are recorded. Voxels are classified based on the Euclidean distance from the voxel center point to the grid center point. If the distance between the center points of adjacent grid cells is less than 3 voxels and the gray value difference is within 15%, it is determined that the two grid cells belong to the same connected domain, and a three-dimensional grid connectivity graph of the lesion region is constructed. The set of boundary points of the lesion region is extracted from the three-dimensional grid connectivity graph, and a three-dimensional bounding box is constructed by searching for the maximum and minimum coordinate values of the boundary points in the x, y, and z directions. The length of the diagonal of the bounding box is calculated as the maximum diameter of the lesion. Combining with the voxel resolution of 0.5 millimeters per pixel, the actual spatial size is converted, and the total number of voxels in the connected domain is counted and multiplied by the volume of a single voxel of 0.125 cubic millimeters to obtain the total volume of the lesion. For the lesion volume values obtained at each time point, a time series data table is constructed in the format of timestamp plus volume value. The volume data sequence is smoothed by 5-point median filtering. By calculating the volume change rate between adjacent time points, the change rate threshold is set to 20%, and the abnormal points exceeding the threshold are marked. Linear interpolation of 4 valid data points before and after the abnormal point is used for data correction to obtain the curve of the lesion volume changing with time. In the process of medical image analysis, voxel feature extraction is the basis for three-dimensional reconstruction of lesions. For example, when Gaussian smoothing is applied to dynamic contrast-enhanced liver scans, a 3×3 window with a standard deviation parameter of 1.5 is used, which not only retains the edge structure features of the lesion but also effectively suppresses random noise. The gray value distribution reflects the tissue density characteristics. For example, in the contrast image of pulmonary nodules, a binarization threshold of 128 is used to distinguish the lesion region from the background tissue. The seed point is selected at the location with the maximum gray value to ensure growth starting from the lesion parenchymal tissue, and a 10% gray value difference threshold adapts to the gradual change characteristics of tissue density. In the three-dimensional grid division, a grid size of 2×2×2 cubic millimeters matches the commonly used CT scan resolution in clinical practice. For example, in the analysis of pancreatic space-occupying lesions, the spatial coordinates record the voxel position distribution, and the Euclidean distance calculation reflects the spatial relationship between voxels. The connectivity determination threshold of 3 voxels takes into account the influence of scan slice thickness, and a 15% gray value tolerance adapts to the tissue density inhomogeneity. The grid connectivity graph describes the spatial morphological features of the lesion. The extraction of lesion boundary features is expressed by a three-dimensional bounding box. For example, in the dynamic contrast-enhanced sequence of breast masses, the maximum and minimum coordinate values in the x, y, and z directions determine the spatial range. A resolution setting of 0.5 millimeters per pixel corresponds to high-resolution thin-slice scanning, and a voxel volume of 0.125 cubic millimeters supports accurate measurement.Voxel counting combined with actual size conversion accurately reflects the size change of the lesion. In the processing of time-series data, 5-point median filtering eliminates discrete noise. In the tracking and evaluation of multiple metastatic tumors, a 20% volume change rate threshold identifies significant changes. Linear interpolation of the 4 data points before and after maintains the smoothness of the curve, and the volume change curve depicts the dynamic evolution process of the lesion. The three-dimensional reconstruction of the lesion involves multiple processing steps, and different types of tumor lesions have specific imaging manifestations. For example, in the follow-up evaluation of gliomas, voxel feature analysis supports the precise localization of the lesion, and connected component construction ensures regional integrity. In the processing of areas with uneven tissue density, grid expression reduces the influence of local noise, and bounding box calculation adapts to irregular shape features. Follow-up data at consecutive time points records the changes in the lesion during the treatment process, and the efficacy is evaluated by quantitatively analyzing the changes in volume parameters. For example, in the monitoring of lung cancer chemotherapy, the shape of the volume curve reflects the treatment response, and outlier correction eliminates the influence of acquisition errors. In the analysis of different imaging sequences, voxel feature extraction and three-dimensional reconstruction parameters are adjusted according to the scanning conditions to ensure the accuracy of the measurement results. In the volume measurement of heterogeneous tumors, multi-level feature extraction and spatial relationship analysis are combined to improve the accuracy of regional division.

[0042] S104. By performing data fitting on the data of the change in lesion volume over time and the data of the change in lesion density over time, analyze the laws of the change in lesion volume and density over time, and obtain volume dynamic characteristic parameters and density statistical parameters to depict the dynamic change laws during the progression of the lesion.

[0043] Use a cubic spline interpolation tool to smooth the time-series data of the lesion volume and density. Calculate the derivative values according to the smoothed data sequence, and mark the positions of the characteristic points through the sign changes of the derivative values; perform second-order polynomial function fitting on the lesion volume data according to the positions of the characteristic points, and extract the quadratic term coefficient and the linear term coefficient from the fitting function to obtain volume dynamic characteristic parameters; for the lesion density data, use a preset threshold to divide the density data into high-value segments, middle segments, and low-value segments, and obtain density change characteristic parameters through the statistics of the data segments; construct a feature vector according to the volume dynamic characteristic parameters and the density change characteristic parameters, and use the min-max normalization method to normalize the feature vector to obtain the lesion progression feature set.

[0044] Specifically, according to the temporal data of the lesion volume and density, a cubic spline interpolation tool is used to smooth the data sequence. The spline node interval is set to 3 time points, and the boundary derivative value is set to 0. The first derivative value is calculated point by point for the interpolation curve. If the signs of the derivative values of two adjacent points are opposite and the change amplitude exceeds 20%, it is marked as an inflection point. If the derivative value is greater than zero and the next derivative value is less than zero, it is marked as a maximum point. If the derivative value is less than zero and the next derivative value is greater than zero, it is marked as a minimum point. The curve slope between two adjacent feature points is calculated to obtain the set of feature parameters. For the volume change data, a second-order polynomial function is used to fit the temporal data points. The coefficients of the polynomial function are calculated, and the quadratic term coefficient is extracted as the acceleration parameter, and the linear term coefficient is extracted as the growth rate parameter. According to the fitting error between the function and the data points, the threshold is set to 10%. The 95% confidence interval of the parameters is calculated through 1000 times of random resampling, and the fluctuation range of the volume dynamic characteristics is obtained from the upper and lower limits of the confidence interval. For the density change data sequence, 2 fixed thresholds are set, and 20% above and below the data mean is taken as the segmentation standard. The density data is divided into three intervals: high-value segment, middle segment, and low-value segment. The proportion of data points, mean, and standard deviation in each interval are statistically analyzed, and the density change rate between adjacent time points is calculated. The maximum value, minimum value, and median of the density change rate are extracted as statistical feature parameters. Combining the volume feature parameters and density statistical parameters, a feature vector of lesion progression is constructed, including parameters such as volume acceleration, growth rate, density segmentation ratio, and density change rate. The feature values are mapped to the 0-1 interval through the min-max normalization method to obtain the set of standardized feature parameters, and the progression process characteristics of the lesion are recorded in chronological order. In the temporal analysis of medical images, the changes in lesion volume and density reflect the disease progression law. The cubic spline interpolation uses a node interval of 3 time points to balance smoothness and detail preservation. For example, in the dynamic contrast-enhanced scan of liver tumors, the boundary derivative value is set to 0 to avoid oscillation at both ends of the curve. When the change in the sign of the derivative value exceeds 20%, it is marked as an inflection point. For example, in the follow-up evaluation of pulmonary nodules, it corresponds to the critical moment of treatment response. For example, a sudden drop in density indicates necrosis, and a sudden decrease in volume reflects the effectiveness of treatment. The second-order polynomial fitting reflects the growth dynamics characteristics of the lesion. For example, in the follow-up analysis of breast masses, the quadratic term coefficient reflects the growth acceleration. A positive value indicates an accelerated progression, and a negative value indicates an inhibited growth. The 10% fitting error threshold balances the model complexity and fitting accuracy. The 95% confidence interval is constructed through 1000 times of random resampling to evaluate the reliability of parameter estimation. The volume change reflects the overall burden of the lesion, and the shape of the growth curve corresponds to different progression stages. The density segmentation analysis uses 20% above and below the data mean as the threshold. For example, in the follow-up of pancreatic space-occupying lesions, the high-value segment corresponds to solid tissue, and the low-value segment indicates cystic degeneration and necrosis. The proportion of interval data reflects the tissue composition, and the density change rate describes the tissue transformation speed. The maximum value, minimum value, and median constitute the density dynamic characteristics, depicting the internal structure evolution process of the lesion.The eigenvector fuses multi-dimensional parameter information. For example, in the monitoring of glioma progression, volume acceleration characterizes the growth dynamics, and the density segmentation ratio reflects the degree of heterogeneity. The influence of dimensions is eliminated through min-max standardization, and the features are mapped to the 0-1 interval for comprehensive analysis. The time-series features record and describe the progression trajectory of the lesion, supporting prognostic evaluation. In the analysis of dynamic contrast-enhanced sequences, volume and density features corroborate each other and jointly reflect the essence of the lesion. For example, in the follow-up of lung cancer chemotherapy, a decrease in volume accompanied by a decrease in density indicates effective treatment, and an increase in density inhomogeneity implies the risk of drug resistance. The change process of the lesion is tracked at consecutive time points, and the treatment effect is evaluated through quantitative feature analysis. Different types of tumor lesions present specific progression patterns, and the combination of feature parameters reflects the characteristics of biological behavior. In the analysis of lesion progression features, static morphological parameters and dynamic change indicators complement each other to construct a complete quantitative evaluation system. During the data processing, the smoothing and fitting parameters are adjusted according to the time resolution to ensure the accuracy of feature extraction. In the analysis of heterogeneous lesions, multi-level feature extraction and parameter fusion are combined to comprehensively depict the progression law. Standardization processing enables feature comparison between different cases, supporting clinical cohort studies.

[0045] S105. According to the dynamic change law during the progression of the lesion, calculate the irregularity of the lesion shape and the smoothness of the boundary, evaluate the morphological complexity and boundary clarity of the lesion, and at the same time compare the density difference between the lesion area and the surrounding normal tissue to obtain a quantitative index of the tissue density difference around the lesion.

[0046] Use the sobel operator to calculate the gradient of the gray-scale image of the lesion area, obtain the set of boundary points according to a preset threshold, and generate a closed contour line through the set of boundary points; obtain a reference point sequence for the closed contour line, calculate a sequence of discrete curvature values according to the reference point sequence, and count the proportion of the number of points exceeding the curvature threshold in the sequence of discrete curvature values to obtain the boundary curvature; construct an annular region according to the closed contour line, where the annular region is obtained by expanding the lesion boundary outward by a certain number of pixels, sample radial sampling points along the boundary from the annular region, and construct a radial sampling line through the radial sampling points; fit a density change curve according to the density sequence on the radial sampling line, calculate a boundary clarity index and a density difference parameter through the density change curve, and obtain a lesion invasion degree index according to the boundary clarity index and the density difference parameter.

[0047] Specifically, based on the grayscale image of the lesion area, the Sobel operator is used to calculate the image gradient. The boundary point set is extracted by taking 50% of the maximum gradient value as the threshold. The boundary points are connected in a clockwise direction to form a closed contour line. The Graham scan tool is used to calculate the minimum circumscribed polygon, and the number of pixel points within the contour line is counted to obtain the actual area. The ratio of the actual area to the area of the minimum circumscribed polygon is calculated to obtain the shape irregularity index. For the boundary contour line, a reference point is taken every 5 pixels along the contour line. A circular window with a radius of 7 pixels is taken centered on the reference point, and the discrete curvature value of the boundary points within the window is calculated. The curvature threshold is set to 0.3, and the proportion of the number of points exceeding the threshold to the total number of reference points is counted to obtain the boundary curvature. The 5-point mean filter is used to smooth the coordinate sequence of the boundary points, and the ratio of the boundary lengths before and after smoothing is calculated to obtain the boundary smoothness parameter. The boundary of the lesion area is extended outward by 10 pixels to construct an annular area. The Sobel operator is used to calculate the gradient map within the annular area. A sampling point is taken every 3 pixels along the boundary, and 5 pixels are extended in both the inner and outer directions from the sampling point to form a radial sampling line. The density value sequence on the radial sampling line is extracted, and the mean and standard deviation of the sequence are calculated to obtain the density distribution characteristics. Based on the density sequence on the radial sampling line, the least squares method is used to fit the density change curve, and the slope of the curve at the midpoint of the sampling line is calculated as the density gradient value. The gradient threshold is set to 20% of the density mean, and the proportion of the sampling lines greater than the threshold is counted to obtain the boundary clarity index. The mean value of the density difference between the inner and outer sides is calculated to obtain the tissue density difference parameter. The boundary clarity index and the density difference parameter are multiplied to obtain a quantitative index of the lesion invasion degree. In medical image analysis, the morphological features and boundary characteristics of lesions reflect their biological behaviors. The Sobel operator has direction sensitivity in edge detection. For example, in the dynamic enhanced scan of the liver, the 50% maximum gradient threshold can both retain the main boundary and suppress noise interference. The boundary points are connected clockwise to form a closed contour, and the minimum circumscribed polygon constructed by the Graham scan tool reflects the shape appearance characteristics, and the area ratio quantifies the shape irregularity. In the boundary morphology analysis, the sampling interval of every 5 pixels balances the computational amount and accuracy requirements. In the CT image of pulmonary nodules, the curvature calculation window with a radius of 7 pixels adapts to the local changes of the boundary. The curvature threshold of 0.3 distinguishes smooth boundaries and sharp turns, and the curvature statistics reflect the complexity of the boundary. The 5-point mean filter eliminates boundary noise. For example, in the ultrasound image of breast masses, the length ratio before and after smoothing reflects the irregularity of the boundary. The annular area analysis extends the range by 10 pixels. In the evaluation of pancreatic space-occupying lesions, it includes the boundary transition zone of the lesion. The radial sampling with a 3-pixel interval ensures complete boundary coverage, and 5 pixels are extended in both the inner and outer directions to obtain the tissue density change. The density sequence statistics of the sampling line reflect both local characteristics and overall distribution, and reveal the characteristics of boundary invasion in lymph node examinations. The fitting of the density change curve describes the tissue transition characteristics. For example, in the enhanced scan of glioblastoma, the midpoint slope reflects the sharpness of the tissue interface.The gradient threshold of the 20% density mean differentiates clear and fuzzy boundaries and indicates the depth of invasion in colorectal cancer staging. The clarity of the boundary and the density difference comprehensively reflect the degree of invasion and support the staging diagnosis. In the analysis of the lesion morphology and boundary, the irregularity and the density difference support each other. For example, in the multi-parametric imaging of prostate cancer, an irregular boundary is often accompanied by a change in the density gradient. In the analysis of the continuous lesion area, the morphological parameters reflect the overall characteristics, and the density characteristics reflect the local changes. In the time-sequential follow-up, the change in the boundary characteristics reflects the progression of the lesion and supports the treatment monitoring. In the analysis of heterogeneous lesions, the multi-level feature extraction corroborates each other. In the evaluation of soft tissue tumors, the morphological complexity and the boundary fuzziness jointly indicate the degree of malignancy. In the quantification of boundary characteristics, the geometric parameters and the density parameters complement each other to construct a complete quantitative evaluation system. In the analysis of different imaging sequences, the parameter settings are adjusted according to the image characteristics to ensure the accuracy of feature extraction. In the evaluation of the invasiveness of the lesion, the morphological features and the density features are jointly applied to provide an objective basis.

[0048] S106. Predict the progression of the disease based on the quantitative indicators of the lesion volume feature parameters, density statistical parameters, and the density difference of the tissue around the lesion.

[0049] Extract the volume feature parameters and density statistical parameters from the time-sequential images of the lesion, perform min-max normalization processing to obtain the normalized feature data, calculate the correlation coefficient matrix based on the normalized feature data, and obtain the feature subset from the correlation coefficient matrix; calculate the feature mean and standard deviation in different disease progression categories based on the feature subset, obtain the features with a discrimination degree higher than the threshold through the ratio of the between-class scatter and the within-class scatter, and sort them in descending order of the information gain rate to obtain the feature vector; use the feature vector to perform cross-validation on the training data, calculate the validation index through the support vector machine classifier, and select the optimal parameter combination to train the classification model according to the validation index; receive the new feature data, perform normalization processing on the new feature data, calculate the distance from the feature point to the classification boundary based on the classification model, and determine the disease progression category and confidence level through the classification boundary distance.

[0050] Specifically, based on the volume feature parameters and density statistical parameters extracted from the time-series images of the lesions, the minimum-maximum normalization tool is used to normalize the feature values, with the normalization interval set to 0 to 1. For the standardized feature data, a correlation coefficient matrix between pairwise features is calculated. The correlation coefficient threshold is set to 0.8. One of the feature pairs with a correlation coefficient higher than the threshold is retained, and the other is removed to obtain a feature subset with redundant features eliminated. For each feature in the feature subset, the mean and standard deviation of the feature in different disease progression categories are calculated respectively. The discrimination threshold is set to 0.5, and the ratio of the between-class dispersion to the within-class dispersion of the feature value distribution is statistically calculated. Features with a discrimination exceeding the threshold are selected from the calculation results, and the features are sorted in descending order according to the information gain ratio, and the features in the top 75% of the ranking are retained to construct a feature vector. The 5-fold cross-validation method is used to divide the dataset, and a support vector machine classifier is constructed for each fold of data. The Gaussian kernel function parameter is set to 0.1, and the penalty parameter is set to 10. The classification accuracy, sensitivity, and specificity are calculated on the validation set, and the parameter combination with the optimal validation index is selected as the final model parameter, and the classification model is retrained through the entire training set. The input new feature data is standardized, and the trained classification model is used to calculate the distance from the feature point to the classification boundary. The boundary threshold is set to plus or minus 0.5. If the distance is greater than 0.5, it is determined that the progression is accelerating; if the distance is less than -0.5, it is determined that the progression is slowing down; if the distance is between -0.5 and 0.5, it is determined that the progression is stable. The disease progression category and confidence level are obtained from the determination results. The confidence level values for each disease progression category are obtained and converted into corresponding prediction probability values through methods such as the softmax function. In medical image feature analysis, data normalization is crucial for eliminating the dimensional difference. For example, in the follow-up evaluation of liver tumors, the volume change rate and density distribution parameters are comparable after being normalized from 0 to 1. The correlation coefficient matrix reflects the correlation strength between features. For example, in the dynamic enhancement of nodules, the volume growth rate and density change rate often show a strong correlation. The correlation threshold of 0.8 balances the feature independence and information integrity. The feature selection process focuses on the class discrimination ability. For example, in the classification of pancreatic space-occupying lesions, there are significant differences in the density distribution features between benign and malignant lesions. The discrimination threshold of 0.5 screens out features with discriminative value, and the discrimination calculation considers the within-class aggregation and between-class separability. The 75% feature retention ratio balances information preservation and dimensionality reduction and simplification. In the prediction of lesion progression, morphological features and density features complement each other. The 5-fold cross-validation ensures the stability of the model. For example, in the prognosis evaluation of breast masses, the data in different folds reflect the consistency of feature expression. The Gaussian kernel function parameter of 0.1 adapts to the non-linear feature mapping, and the penalty parameter of 10 balances the model generalization ability. The validation index comprehensively evaluates the classification performance. The accuracy reflects the overall correct judgment rate, the sensitivity reflects the progression recognition ability, and the specificity characterizes the accuracy of the stability judgment.The setting of the distance threshold for the classification boundary reflects the prediction confidence. In the monitoring of glioma progression, the positive and negative thresholds of 0.5 clearly distinguish between definite progression and the state to be observed. The distance calculation quantifies the classification reliability, and the larger the absolute value, the more certain the prediction. The determination of the progression category combines the characteristic changes at multiple time points and dynamically reflects the trend of disease evolution. In the extraction of lesion features, the time-series data reflects the dynamic change law. In the follow-up after colorectal cancer surgery, the shape of the volume growth curve corresponds to different prognostic types. During the feature selection process, the statistical characteristics and clinical relevance support each other to construct a stable and reliable prediction index system. During the classifier training, the parameter optimization and performance evaluation promote each other to improve the prediction accuracy. In the multi-modal feature fusion analysis, the features in different dimensions reflect different attributes of the lesion. For example, in the prediction of lymph node metastasis, the morphological changes and density changes jointly indicate the prognostic risk. The data preprocessing and feature selection complement each other to ensure the quality of the input features. The model training and validation complement each other to establish a reliable prediction tool. In the prediction of disease progression, the quantitative features and classification judgments support each other to provide a basis for clinical decision-making. In the feature processing flow related before and after, the parameter settings of each link are based on clinical practice experience to construct a complete prediction system. In the prediction of different types of lesions, the feature extraction and model training parameters are adjusted according to the specific application scenario to ensure the reliability of the prediction results.

[0051] S107. Determine whether the prediction result of the disease progression reaches the preset risk threshold range. If the prediction result is within the preset risk threshold range, then determine that the disease progression is in a risk state and generate a high-risk warning message.

[0052] The standardized prediction value is obtained by using the min-max normalization tool according to the prediction probability value, and the standardized prediction value is within the range of zero to one; the upper limit value and the lower limit value of the risk threshold are set for the standardized prediction value, and the risk status flag is obtained by calculating the position relationship between the standardized prediction value and the risk threshold interval; the confidence interval is calculated for the standardized prediction value by using the interval estimation tool based on the standard deviation, and the overlap ratio value is obtained by counting the overlap range between the confidence interval and the risk threshold interval; the risk level is judged according to the overlap ratio value. If the overlap ratio value exceeds the preset overlap threshold, the risk level is increased, otherwise it is decreased. The risk score is constructed by the risk level and the overlap ratio value; the warning level identifier is set according to the risk score. The warning level identifier includes a high-risk red identifier, a medium-risk yellow identifier, and a low-risk green identifier. The risk warning message is generated by the warning level identifier and the timestamp.

[0053] Specifically, for the predicted probability values obtained from the prediction of disease progression, the minimum-maximum normalization tool is used to map the predicted values to the interval from 0 to 1. The upper limit of the normalized risk threshold is set to 0.8 and the lower limit is set to 0.6. Calculate the position relationship between the current predicted value and the risk threshold interval. If the predicted value falls within the threshold interval, record the risk status flag, and at the same time calculate the relative position percentage of the predicted value within this interval. For the confidence interval of the predicted value, an interval estimation tool based on the standard deviation is used to calculate the 95% confidence interval, and the proportion of the overlapping range between the confidence interval and the risk threshold interval is statistically analyzed. The overlapping ratio threshold is set to 0.3. If the overlapping ratio exceeds the threshold, the risk level is increased; otherwise, the risk level is decreased. A risk score is constructed from the risk level and the overlapping ratio. The risk score is divided into three levels. A score above 90 is set as the high-risk level, a score between 70 and 90 is set as the medium-risk level, and a score below 70 is set as the low-risk level. According to the calculation result of the risk score, the corresponding risk level is selected, and red, yellow, and green warning indicators are set according to the risk level to generate a risk flag containing the warning level and the timestamp. The warning information is constructed in a key-value pair format, and six keyword fields including the warning title, predicted value, risk score, risk level, warning time, and confidence interval are set. A standard format warning prompt is generated through string concatenation, and different prompt level flags are set according to the warning level to complete the generation of the risk warning information. In the medical warning system, the standardization of risk assessment is crucial. For example, in the prediction of the risk of liver cancer recurrence, the prediction probability is mapped through normalization from 0 to 1, which is convenient for setting the threshold. The dual-threshold design of 0.8 and 0.6 constructs a buffer interval to avoid frequent alarms caused by prediction fluctuations. The relative position of the predicted value within the threshold interval reflects the degree of risk. For example, a predicted value of 0.75 is located at 75% of the interval, indicating a relatively high degree of risk. The confidence interval analysis reflects the reliability of the prediction. In the follow-up assessment of pulmonary nodules, the probability that the 95% confidence interval contains the true risk is high. The overlapping ratio threshold of 0.3 balances sensitivity and specificity. For example, in the prediction of the progression of breast masses, a larger overlapping ratio indicates the uncertainty of the prediction result and requires increased vigilance. The risk score comprehensively considers the predicted value and uncertainty to form a more robust warning basis. The three-level risk classification corresponds to the needs of clinical practice. In the monitoring of pancreatic cystic tumors, a high-risk score above 90 indicates the need for timely intervention. The red, yellow, and green warning indicators intuitively reflect the risk level. A medium-risk status between 70 and 90 requires enhanced follow-up. The risk level is closely related to clinical management, and different levels trigger different warning responses. The standardization of warning information improves the efficiency of message transmission. For example, in postoperative monitoring, the key-value pair format contains complete warning elements. The predicted value, risk score, and confidence interval constitute a quantitative description, and the timestamp records the sequence of warning occurrences. The multi-dimensional information supports clinical decision-making, and different level markers guide intervention measures. In the warning of disease progression, prediction accuracy and warning timeliness complement each other. For example, in the follow-up assessment of glioblastoma, the dynamic comparison between the prediction probability and the risk threshold reflects the trend of disease change.Confidence interval analysis reveals the reliability of predictions, and risk scores quantify the urgency of warnings. Standardizing warning information facilitates clinical applications, and the standard format is conducive to information transmission and processing. In the risk assessment system, multi-level warning mechanisms support each other. For example, in the monitoring of lymphoma treatment efficacy, a predicted value exceeding 0.8 directly triggers a high-risk warning, and a value in the range of 0.6 to 0.8 is judged based on the confidence level. The overlap between the 95% confidence interval and the risk threshold indicates a boundary situation, and a comprehensive assessment needs to be combined with clinical manifestations. The three-level risk classification corresponds to different follow-up strategies, and standardizing warning information promotes multidisciplinary collaboration. In the monitoring of heterogeneous lesions, the risk assessment criteria need to be adjusted dynamically. There are differences in the risk characteristics of different types of tumors, and the warning parameter settings should be optimized according to specific disease types. The standardized warning mechanism ensures the accuracy of information transmission and supports clinical risk management. During the warning process, quantitative indicators and risk levels complement each other to build a comprehensive warning system.

[0054] S108. Based on the judgment result predicted according to the disease progression, combined with the fitting data of the lesion volume and density changing over time, judge the severity of the disease, stage and grade the process of lesion progression, and generate a quantitative disease diagnosis report.

[0055] Use a quadratic polynomial fitting tool to perform piecewise fitting on the curves of lesion volume and density changing over time, calculate the volume change rate and density change rate in each interval according to the curves; judge the lesion progression level according to the volume change rate and density change rate, obtain a severity score for the progression level, and the severity score consists of a volume progression score, a density change score and a prediction probability score; determine the disease staging status according to the change trend of the progression level at consecutive time points. If the progression level rises at consecutive time points, it is determined to be the progressive stage, and if the progression level is stable at consecutive time points, it is determined to be the stable stage; generate the content of the diagnosis report for the disease staging status and severity score, and the content of the diagnosis report includes the volume density change rate, the severity score and the disease staging result.

[0056] Specifically, according to the prediction judgment results of disease progression and the data of the change in lesion volume density, a quadratic polynomial fitting tool is used to perform piecewise fitting on the change curves of volume and density over time. The fitting interval length is set to 3 months, and the change rate within each interval is calculated. For the volume change rate, a rapid progression threshold of 20% per month and a moderate progression threshold of 10% per month are set. For the density change rate, a significant change threshold of 15% per month and a mild change threshold of 5% per month are set. The progression grades of volume and density are obtained from the change rate values. For the volume density progression grade and the prediction probability value, a severity scoring system is constructed. The volume progression score interval is set to 0 - 40 points, the density change score interval is set to 0 - 30 points, and the prediction probability score interval is set to 0 - 30 points. The total score is calculated according to the scores corresponding to the progression grades. A total score of 90 - 100 points is classified as extremely severe, 75 - 89 points is classified as severe, 60 - 74 points is classified as moderate, and below 60 points is classified as mild. The severity of the disease is determined from the scoring results. According to the change in the progression grade at consecutive time points, a continuous increase in the progression grade at 3 consecutive time points is determined as the progressive stage, a stable progression grade at 3 consecutive time points is determined as the stable stage, and a continuous decrease in the progression grade at 3 consecutive time points is determined as the remission stage. Combining the severity score, the disease is staged to generate a standardized diagnosis result including the progression speed, severity, and staging phase. The content of the diagnostic report is constructed in a structured format. The patient information, examination time, lesion location, size, and shape are set as basic information. The volume change rate, density change amplitude, severity score, and progression stage are set as diagnostic key points. The risk prediction probability and warning level are set as prompt information. Corresponding descriptive statements are generated through numerical index mapping, and the disease progression process is recorded in chronological order to complete the generation of the diagnostic report. In medical image quantitative analysis, the dynamic change characteristics of lesions reflect the disease progression state. Quadratic polynomial fitting can capture the accelerating or decelerating trend of lesions. For example, in the follow-up evaluation of liver cancer, a 3-month fitting interval balances sensitivity and stability. The rapid progression threshold of a 20% monthly increase in volume corresponds to a shortening of the tumor doubling time, indicating a change in biological behavior in nodule monitoring. The significant change threshold of a 15% decrease in density reflects tissue necrosis in pancreatic cystic lesions. The severity scoring system synthesizes multi-dimensional features. For example, in the dynamic contrast enhancement of breast masses, the volume progression accounts for 40% of the weight, reflecting the change in tumor burden, and the density change and prediction probability each account for 30% of the weight to balance morphology and risk prediction. Above 90 points for extremely severe corresponds to a high-risk state of rapid progression, indicating poor prognosis in the evaluation of glioblastoma. The severe grading of 75 - 89 points requires close follow-up, and the moderate grading of 60 - 74 points indicates that the disease is controllable. The staging judgment is based on observations at consecutive time points. For example, in the postoperative monitoring of colorectal cancer, an accelerated progression at 3 consecutive points indicates an increased risk of recurrence. The characteristics of the progressive stage include a rapid increase in volume and a significant change in density, corresponding to treatment resistance in the evaluation of lymphoma chemotherapy. The stable stage indicates that the disease is under control, and the remission stage reflects the effectiveness of treatment.The severity score and staging judgment corroborate each other to construct a complete disease assessment system. Standardizing the diagnostic report improves the efficiency of information transmission. For example, in the follow-up of soft tissue tumors, the basic information records the spatio-temporal information of the lesion, and the diagnostic key points quantify the progression characteristics. A volume change rate of 20% per month is translated into a description of rapid progression, and a density change of 15% per month corresponds to a significant change prompt. A severity score of 85 generates a description of severe progression, and a risk prediction probability of 0.8 triggers a high-risk warning. During the disease assessment process, quantitative indicators and qualitative descriptions complement each other. For example, in prostate cancer radiomics, volume and density changes reflect histological changes, and the progression speed corresponds to the malignancy degree. Observing consecutive time points reveals the law of disease evolution, and multi-dimensional scoring ensures the reliability of judgment. Standardized reports promote multi-disciplinary collaboration and support clinical decision-making. In the diagnosis of different diseases, the assessment criteria need to be optimized and adjusted. Tumors with a high degree of malignancy adopt more sensitive progression thresholds, and slowly growing lesions select longer observation intervals. The weight of the severity score can be adjusted according to the characteristics of the disease to ensure the clinical relevance of the assessment results. During the report generation process, the selection of descriptive statements strictly corresponds to quantitative indicators to ensure the accuracy of expression. The setting of warning information is based on the needs of clinical intervention to construct a standardized follow-up management system.

[0057] The above embodiments are only one of the preferred embodiments of the present invention and should not be used to limit the protection scope of the present invention. Any meaningless modifications or polishing made on the main design concept and spirit of the present invention, as long as the technical problems solved are still consistent with the present invention, should be included in the protection scope of the present invention.

Claims

1. A method for analyzing contrast images of digital diagnosis and treatment equipment, characterized in that, The method includes: acquiring a sequence of contrast images collected at different time points of a target patient, and for the contrast images at different time points, segmenting the lesion region from pixel-level features, voxel-level features, and region-level features respectively to obtain the number of pixels, average density, and gray value distribution of the lesion region at different time points; extracting and describing the lesion region features at different time points to generate feature descriptors, performing a ratio test on the feature descriptors to screen out feature matching points, calculating the transformation matrix between the lesion regions through the feature matching points to obtain the changes between the lesion regions at different time points; after alignment registration, obtaining the change data of the lesion volume over time according to the number of pixels in the lesion region at each time point, and at the same time calculating the average density of the lesion region according to the pixel gray values in the lesion region to obtain the change data of the lesion density over time; by fitting the change data of the lesion volume over time and the change data of the lesion density over time, analyzing the laws of the change of the lesion volume and density over time to obtain volume dynamic characteristic parameters and density statistical parameters for depicting the dynamic change laws during the lesion progression process; according to the dynamic change laws during the lesion progression process, calculating the irregularity of the lesion shape and the smoothness of the boundary to evaluate the morphological complexity and boundary clarity of the lesion, and at the same time comparing the density difference between the lesion region and the surrounding normal tissues to obtain a quantitative index of the tissue density difference around the lesion; predicting the disease progression based on the lesion volume characteristic parameters, density statistical parameters, and the quantitative index of the tissue density difference around the lesion; judging whether the prediction result of the disease progression reaches the preset risk threshold range, if the prediction result is within the preset risk threshold range, then judging the disease progression as a risk state and generating a high-risk warning message; according to the judgment result of the prediction of the disease progression, combining the fitting data of the lesion volume and density changing over time, judging the severity of the disease, staging and grading the lesion progression process, and generating a quantitative disease diagnosis report.

2. The method according to claim 1, characterized in that, The acquisition of a sequence of contrast images collected at different time points of a target patient, and for the contrast images at different time points, segmenting the lesion region from pixel-level features, voxel-level features, and region-level features respectively to obtain the number of pixels, average density, and gray value distribution of the lesion region at different time points, includes: performing Gaussian smoothing processing on the images according to the contrast image acquisition sequence, and using an edge tracking tool based on pixel gray value gradient to obtain the lesion boundary contour map; selecting the point with the maximum gray value as the seed point for the lesion boundary contour map, and performing region filling according to the rule that the density difference of eight-neighbor pixels is less than a preset threshold to obtain the lesion region density distribution map; using the lesion region density distribution map to establish three-dimensional point cloud data with a preset voxel size, removing the noise points in the point cloud data to obtain three-dimensional feature data; extracting the centroid coordinates of the lesion region from the three-dimensional feature data to establish a registration reference point, and performing nearest neighbor pairing on the lesion boundary points according to a preset distance threshold to obtain the lesion region volume data and the gray distribution histogram. It further includes: calculating the feature quantization value of the lesion area by extracting the regional-level features of the contrast images at different time points, comparing the preset feature quantization threshold with the feature quantization value of the lesion area, and preliminarily segmenting to obtain the target area; setting the selection conditions for the preset seed points, selecting several seed points from the target area according to the selection conditions, and expanding the segmentation area starting from the seed points until the preset expansion condition is reached to achieve the final segmentation, specifically including: performing noise reduction processing on the contrast image sequence using a Gaussian smoothing kernel function, and obtaining the candidate area through projection calculation according to the gray value distribution matrix of the denoised image; obtaining the seed point positions according to the local gray value mean and variance of the candidate area through the preset uniform area determination condition, and expanding the area from the seed points using the gray difference threshold to obtain the expanded segmentation area; extracting the edge contour line for the expanded segmentation area, obtaining the feature points by calculating the edge gradient data within the detection window, and performing area segmentation according to the curvature of the feature points until the area converges to obtain the final segmentation area; extracting the shape feature parameters from the final segmentation area, judging the difference in the regional morphological features by the feature point pairing distance threshold, and if the difference in the shape feature parameters is within the preset range and the feature point pairing rate exceeds the preset threshold, determining it as the corresponding area.

3. The method according to claim 1, wherein Extracting and describing the lesion area features at different time points, generating the feature descriptors, performing a ratio test on the feature descriptors, screening out the feature matching points, and calculating the transformation matrix between the lesion areas through the feature matching points to obtain the changes between the lesion areas at different time points, including: performing smoothing processing on the original image of the lesion area using a Gaussian blur tool, extracting the edge contour line from the smoothed image through an edge detection operator, and calculating the gradient histogram according to the edge contour line to obtain the feature descriptors; calculating the Euclidean distance values between the corresponding positions at different time points according to the feature descriptors, and obtaining the initial matching point set through the comparison result of the Euclidean distance value and the preset distance threshold; randomly selecting matching point pairs from the initial matching point set to calculate the affine transformation matrix, and counting the number of inliers that satisfy the affine transformation matrix to obtain the optimal transformation parameters; performing spatial mapping on the lesion area image according to the optimal transformation parameters, resampling the transformed image using a bilinear interpolation tool, and judging the registration result by calculating the mutual information value between the resampled image and the original image.

4. The method according to claim 1, wherein After the alignment and registration, according to the number of pixels in the lesion area at each time point, the change data of the lesion volume over time is obtained. At the same time, according to the pixel gray values in the lesion area, the average density of the lesion area is calculated, and the change data of the lesion density over time is obtained, including: obtaining the lesion boundary contour line through the region growing tool based on seed points, and converting the number of connected region pixels within the contour line and the image resolution to obtain the lesion volume value; for the image area within the lesion boundary contour line, using the histogram equalization tool for enhancement processing, extracting the gray value matrix from the enhanced image and obtaining the regional average density value through projection integral operation; constructing time series data based on the lesion volume value and the regional average density value, linearly interpolating and correcting the abnormal points exceeding the preset value in the time series data to obtain the normalized volume density change sequence; for the normalized volume density change sequence, obtaining the change rate by calculating the difference between adjacent time points, constructing the lesion progression feature parameter set using the mean and standard deviation of the change rate, and normalizing the feature parameter set to obtain the standardized lesion progression feature. It also includes: according to the number of pixels in the lesion area at each time point, combined with the voxel-level features of the contrast image, the voxel-level features including voxel gray value, voxel spatial position, and voxel shape and size, determining the number of pixels and voxel volume at each time point, calculating the total volume of the lesion, recording the total volume of the lesion at each time point, and obtaining the change data of the lesion volume over time, specifically including: performing noise reduction processing on the contrast image data using the Gaussian smoothing kernel function, performing binary segmentation on the smoothed image using a fixed threshold, and obtaining the voxel set of the initial lesion area through the region growing algorithm; dividing the space of the initial lesion area using a three-dimensional grid, classifying voxels according to the Euclidean distance from the voxel to the center point of the grid, and judging the connectivity by the distance and gray value difference between adjacent grid cells to obtain the three-dimensional grid connectivity map of the lesion area; extracting the boundary point set from the three-dimensional grid connectivity map, constructing a bounding box according to the maximum and minimum coordinate values of the boundary points in three directions, and obtaining the maximum diameter and volume of the lesion by calculating the diagonal length of the bounding box and counting the total number of voxels in the connected domain; constructing a time series data table for the lesion volume, and obtaining the change data of the lesion volume over time through linear interpolation.

5. The method according to claim 1, wherein By performing data fitting on the data of the change in lesion volume over time and the change in lesion density over time, analyzing the laws of the change in lesion volume and density over time, and obtaining volume dynamic characteristic parameters and density statistical parameters to depict the dynamic change law during the progression of the lesion, including: using a cubic spline interpolation tool to smooth the time-series data of the lesion volume and density, calculating the derivative values according to the smoothed data sequence, and marking the positions of the characteristic points through the sign changes of the derivative values; performing second-order polynomial function fitting on the lesion volume data according to the positions of the characteristic points, and extracting the quadratic term coefficient and the linear term coefficient from the fitting function to obtain volume dynamic characteristic parameters; for the lesion density data, dividing the density data into a high-value segment, a middle segment, and a low-value segment using a preset threshold, and obtaining density change characteristic parameters through the statistics of the data segments; constructing a feature vector according to the volume dynamic characteristic parameters and the density change characteristic parameters, and performing normalization processing on the feature vector using the min-max normalization method to obtain a lesion progression feature set.

6. The method according to claim 1, wherein According to the dynamic change law during the progression of the lesion, calculating the irregularity of the lesion shape and the smoothness of the boundary, evaluating the morphological complexity and boundary clarity of the lesion, and at the same time comparing the density difference between the lesion area and the surrounding normal tissue to obtain a quantitative index of the density difference of the surrounding tissue of the lesion, including: using a sobel operator to calculate the gradient of the grayscale image of the lesion area, obtaining a set of boundary points according to a preset threshold, and generating a closed contour line through the set of boundary points; obtaining a reference point sequence for the closed contour line, calculating a sequence of discrete curvature values according to the reference point sequence, and statistically calculating the proportion of the number of points in the sequence of discrete curvature values that exceed the curvature threshold to obtain the boundary curvature; constructing an annular region according to the closed contour line, where the annular region is obtained by expanding the lesion boundary outward by a certain number of pixels, sampling radial sampling points along the boundary from the annular region, and constructing a radial sampling line through the radial sampling points; fitting a density change curve according to the density sequence on the radial sampling line, calculating a boundary clarity index and a density difference parameter through the density change curve, and obtaining a lesion invasion degree index according to the boundary clarity index and the density difference parameter.

7. The method according to claim 1, wherein The quantitative indicators based on the lesion volume characteristic parameters, density statistical parameters, and the density difference of the surrounding tissues of the lesion are used to predict the disease progression, including: extracting the volume characteristic parameters and density statistical parameters from the temporal images of the lesion, obtaining the normalized feature data by using the min-max normalization process, calculating the correlation coefficient matrix according to the normalized feature data, and obtaining the feature subset from the correlation coefficient matrix; calculating the feature mean and standard deviation in different disease progression categories according to the feature subset, obtaining the features with a discrimination degree higher than the threshold through the ratio of the between-class scatter and the within-class scatter, and sorting them in descending order of the information gain ratio to obtain the feature vector; using the feature vector to perform cross-validation on the training data, calculating the validation index through the support vector machine classifier, and selecting the optimal parameter combination to train the classification model according to the validation index; receiving the new feature data, performing standardization processing on the new feature data, calculating the distance from the feature point to the classification boundary according to the classification model, and determining the disease progression category and confidence level through the classification boundary distance.

8. The method according to claim 1, characterized in that Judging whether the prediction result of the disease progression reaches the preset risk threshold range. If the prediction result is within the preset risk threshold range, it is judged that the disease progression is in a risk state and a high-risk warning message is generated, including: obtaining the standardized prediction value by using the min-max normalization tool according to the prediction probability value, and the standardized prediction value is within the range of zero to one; setting the upper limit value and the lower limit value of the risk threshold for the standardized prediction value, and obtaining the risk state mark by calculating the position relationship between the standardized prediction value and the risk threshold interval; calculating the confidence interval for the standardized prediction value by using the interval estimation tool based on the standard deviation, and obtaining the overlap ratio value by counting the overlap range between the confidence interval and the risk threshold interval; judging the risk level according to the overlap ratio value. If the overlap ratio value exceeds the preset overlap threshold, the risk level is increased, otherwise it is decreased, and a risk score is constructed through the risk level and the overlap ratio value. Setting the warning level identifier according to the risk score. The warning level identifier includes a high-risk red identifier, a medium-risk yellow identifier, and a low-risk green identifier, and generating a risk warning message through the warning level identifier and the timestamp.

9. The method according to claim 1, wherein The judgment result predicted according to the disease progression, combined with the fitting data of the lesion volume and density changing over time, is used to judge the severity of the disease, stage and grade the lesion progression process, and generate a quantitative disease diagnosis report, including: using a quadratic polynomial fitting tool to perform piecewise fitting on the curves of the lesion volume and density changing over time, and calculating the volume change rate and density change rate within each interval according to the curves; judging the lesion progression level according to the volume change rate and density change rate, and obtaining a severity score for the progression level, where the severity score is composed of a volume progression score, a density change score and a prediction probability score; judging the disease staging status according to the change trend of the progression level at consecutive time points. If the progression level rises at several consecutive time points, it is judged as the progressive stage. If the progression level is stable at several consecutive time points, it is judged as the stable stage; generating the diagnosis report content for the disease staging status and severity score, where the diagnosis report content includes the volume density change rate, the severity score and the disease staging result.

10. A contrast image analysis device for a digital diagnosis and treatment device, characterized in that, Implement the contrast image analysis method according to any one of claims 1-9. The device includes: an image processing and feature extraction module for acquiring and analyzing a sequence of contrast images at different time points; a lesion area change analysis module for calculating the changes between lesion areas at different time points; a lesion dynamic feature analysis module for analyzing the law of the lesion volume and density changing over time; a lesion morphology evaluation module for evaluating the morphological complexity and boundary clarity of the lesion; a disease prediction and risk assessment module for predicting the disease progression and generating a high-risk warning message; and a disease diagnosis and grading module for judging the severity of the disease and generating a quantitative diagnosis report.

Images