A tumor pseudo-progression identification system based on image feature variability

Abstract

The application discloses a tumor false progression identification system based on image feature variability, and relates to the technical field of medical image intelligent analysis; the system comprises a variability feature construction module, a variability consistency calculation module, a disturbance modeling module and a reliability judgment module. The target tumor region is adaptively divided by feature stability driving, and an image feature variability unit set is constructed; a spatial variability consistency graph is constructed, and a consistency component reflecting overall cooperative change and a non-consistency disturbance component reflecting local non-cooperative change are separated; the disturbance modeling module decomposes a time evolution process of multi-phase image features into a trend item and a disturbance item, and extracts a disturbance spectrum feature set representing a transient treatment response behavior; the application can simultaneously depict variability behaviors of tumor image features from spatial and time dimensions, improves the accuracy and stability of false progression identification, and provides reliable auxiliary basis for clinical efficacy evaluation and follow-up decision.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image analysis technology, specifically a tumor pseudoprogression identification system based on image feature variability. Background Technology

[0002] A tumor pseudoprogression identification system based on image feature variability is an intelligent auxiliary judgment system for tumor treatment follow-up image analysis. It systematically models the spatial variation behavior and temporal evolution perturbation characteristics of image features within the tumor region to comprehensively assess whether image changes conform to the typical evolutionary pattern of pseudoprogression. The system constructs image feature variation units to analyze the consistency and inconsistency relationships of image feature changes in different regions, and further characterizes the transient perturbation behavior caused by treatment response during the temporal change of image features. Thus, it can determine the credibility of pseudoprogression even when there is uncertainty in overall image changes. This system can provide more stable, fine-grained, and interpretable auxiliary evidence for clinical efficacy evaluation, treatment plan adjustment, and follow-up decisions, reducing the clinical risks caused by misjudgment of pseudoprogression.

[0003] During tumor radiotherapy, chemotherapy, and immunotherapy, tumor imaging often shows changes such as increased volume, enhanced enhancement, or abnormal signal during follow-up. Some of these imaging changes are not caused by actual tumor progression, but are temporary imaging alterations caused by factors such as inflammatory response, immune cell infiltration, edema, or necrosis repair. Clinically, this phenomenon is usually referred to as "pseudo-progression of tumor."

[0004] Existing methods for evaluating tumor treatment efficacy and differentiating pseudoprogression mostly rely on changes in overall tumor volume, maximum diameter, or simple image intensity statistics as the basis for judgment, lacking a systematic analysis of the spatial heterogeneity and temporal evolution of intratumoral imaging characteristics. In practical applications, when different locations within the tumor region exhibit inconsistent imaging changes, or when transient imaging disturbances occur within a short period of time such as initial enhancement followed by decline, traditional methods often fail to effectively distinguish between true and pseudoprogression, easily leading to premature adjustments to treatment plans or even interruption of effective treatment; specifically:

[0005] Existing tumor pseudoprogression identification systems generally rely on changes in tumor volume at a single time point or overall imaging response indicators, making it difficult to distinguish between pseudoprogression caused by post-treatment inflammatory response, immune cell infiltration, etc., and true tumor progression. Furthermore, they are prone to misjudgment when local imaging changes are inconsistent.

[0006] In the existing process of constructing variant features, there are technical problems such as the use of fixed grids or simple superpixel segmentation, which do not fully consider the stability differences of image features in the spatial and temporal dimensions. This leads to the problem of overly coarse or fine segmentation and unstable features within units in the tumor edge area or areas with strong heterogeneity.

[0007] In existing processes for calculating variation consistency, there are technical problems that rely heavily on simple similarity statistics or local correlation analysis, making it difficult to distinguish between overall coordinated changes and local anomalous changes, and failing to quantitatively characterize the impact of local non-coordinated changes on the overall image evolution pattern. Summary of the Invention

[0008] In view of the above situation and to overcome the defects of the prior art, the present invention provides a tumor pseudoprogression identification system based on image feature variability. The technical solution adopted by the present invention is as follows: The tumor pseudoprogression identification system based on image feature variability provided by the present invention includes a variant feature construction module, a variant consistency calculation module, a perturbation modeling module, and a confidence determination module.

[0009] The variant feature construction module is used to construct image feature variant units. By performing feature stability-driven adaptive partitioning on the target tumor region, an image feature variant unit set is obtained, and the image feature variant unit set is sent to the variant consistency calculation module.

[0010] The mutation consistency calculation module is used for spatial mutation consistency calculation. It constructs a mutation consistency graph with mutation units as nodes and calculates the consistency components and non-consistency disturbance components to obtain spatial mutation consistency reference data. The spatial mutation consistency reference data is then sent to the disturbance modeling module and the credibility determination module.

[0011] The disturbance modeling module is used for time-evolution disturbance modeling. It obtains a disturbance spectrum feature set by decomposing the trend term and the disturbance term, and sends the disturbance spectrum feature set to the credibility determination module.

[0012] The credibility determination module is used to determine the credibility of false progress. It obtains the credibility identification reference data of false progress by fusing spatial variation consistency reference data and perturbation spectrum feature set.

[0013] Furthermore, the image feature variation unit construction is used to structurally represent the local variation behavior of image features within the target tumor region. Specifically, it employs a stability-improved and optimized spatiotemporal multi-feature adaptive variation unit construction method to extract image features including grayscale, texture, edge intensity, and enhancement changes within the target tumor region. Based on the spatial stability and similarity relationship of the image features, the target tumor region is adaptively partitioned with feature stability as the driving force to obtain a set of image feature variation units. This includes the following steps: image data foundation construction, image feature analysis region limitation, multi-type image feature extraction, image feature stability characterization optimization, adaptive partitioning initial structure generation, and image feature variation unit generation.

[0014] The image data infrastructure is specifically constructed by acquiring multi-temporal image data of the target tumor at multiple time points before and after treatment, and performing spatiotemporal registration and intensity normalization on the multi-temporal image data to obtain a spatiotemporally aligned and intensity-consistent multi-temporal image dataset.

[0015] The image feature analysis region limitation specifically involves determining the target tumor region based on the multi-temporal image dataset and generating a corresponding tumor region mask, thereby limiting the spatial range of subsequent image feature extraction and variation analysis.

[0016] The multi-type image feature extraction specifically involves extracting image feature vectors, including grayscale features, texture features, edge intensity features, and enhancement change features, for each temporal image within the tumor region mask, to obtain a multi-temporal image feature field of the tumor region.

[0017] The image feature stability characterization optimization specifically involves calculating the local stability of image features in the spatial neighborhood and the consistency stability in the temporal dimension based on the multi-temporal image feature field, and combining edge intensity information to form an image feature stability distribution, thereby obtaining the image feature stability field of the tumor region.

[0018] The adaptive partitioning initial structure generation specifically involves increasing the partitioning density in regions with low stability and decreasing the partitioning density in regions with high stability, based on the image feature stability field, to generate a stability-guided adaptive partitioning seed set.

[0019] The generation of the image feature variation unit specifically involves using the adaptive partitioning seed set as the initial center, aggregating and partitioning pixels or voxels within the tumor region based on image feature similarity, spatial adjacency, and image feature stability weights, and constraining and verifying the stability within the unit to obtain a set of image feature variation units that satisfy the requirements of stable features within the unit and significant feature differences between units.

[0020] Furthermore, the spatial variation consistency calculation is used to evaluate the spatial cooperative change relationship among various image feature variation units within the target tumor region. Specifically, it involves constructing a spatial variation consistency map using the set of image feature variation units as nodes, and employing a spatial variation cooperative analysis method based on structural consistency component separation. The spatial variation consistency map is analyzed based on the similarity of different image feature variation units in image feature change direction, change amplitude, and spatial adjacency. The consistency component reflecting the overall cooperative change characteristics and the non-consistent perturbation component reflecting local non-cooperative change characteristics are calculated to obtain spatial variation consistency reference data. This includes the following steps: unit-level variation feature aggregation, spatial variation consistency map construction, spatial cooperative change relationship characterization, consistency perturbation structural component separation, and spatial variation consistency reference data generation.

[0021] The unit-level variation feature aggregation is specifically based on the image feature variation unit set, performing aggregation processing on the multi-temporal image features of each image feature variation unit, extracting unit-level variation feature vectors that reflect the direction and magnitude of image feature changes, and obtaining the unit-level variation feature set of the image feature variation unit.

[0022] The construction of the spatial variation consistency graph specifically involves using the image feature variation units as graph nodes, establishing node connections based on the spatial adjacency relationships between the image feature variation units, and assigning weights to the connections between nodes based on the similarity of the unit-level variation feature vectors in the direction and magnitude of image feature changes, thereby obtaining the spatial variation consistency graph.

[0023] The spatial collaborative change relationship characterization specifically involves performing graph structure analysis on the spatial variation consistency map and constructing a graph Laplacian matrix that reflects the node connection relationship to characterize the overall collaborative change characteristics of image feature variation in the spatial dimension.

[0024] The separation of the consistent perturbation structural components specifically involves performing spectral decomposition on the graph Laplacian matrix to extract consistent components that reflect the overall cooperative change characteristics and non-consistent perturbation components that reflect the local non-cooperative change characteristics.

[0025] The generation of the spatial variation consistency reference data specifically involves calculating the consistency index and the inconsistency perturbation index of image feature variations in the spatial dimension based on the consistency component and the inconsistency perturbation component, thereby obtaining spatial variation consistency reference data used to characterize the spatial variation synergy of the target tumor region.

[0026] Furthermore, the time evolution perturbation modeling is used to characterize the transient perturbation behavior caused by treatment response during the change of target tumor image features over time. Specifically, based on multi-temporal image data, a treatment response transient perturbation spectrum extraction method with trend perturbation dual channels is used to construct a feature sequence of image features changing over time. The feature sequence is then decomposed to decompose the time evolution process of image features into a trend term reflecting long-term change trends and a perturbation term reflecting short-term fluctuations. The perturbation term is analyzed to extract the perturbation intensity, perturbation duration, and its change characteristics to obtain a perturbation spectrum feature set. The process includes the following steps: unit-level time series construction, time axis processing, differentiation between trend term and transient perturbation term, transient perturbation event identification, perturbation spectrum feature extraction, and perturbation spectrum feature set generation.

[0027] The construction of the unit-level time series is specifically based on multi-temporal image data. At the scale of image feature variation units, the features of each temporal image are aggregated to form a unit-level feature sequence that reflects the changes of image features over time. The corresponding timestamp information is recorded to obtain the unit-level image feature time series.

[0028] The time axis processing specifically involves performing time axis normalization on the unit-level image feature time series and using a robust smoothing strategy to denoise the feature series in order to reduce the impact of unequal time intervals and imaging noise on subsequent analysis, thereby obtaining a robust image feature time series.

[0029] The distinction between the trend term and the transient disturbance term is specifically achieved by performing component decomposition on the robust image feature time series. Under the combined effect of trend smoothing constraints and disturbance sparsity constraints, the temporal evolution process of the image features is separated into a trend term reflecting long-term change trends and a disturbance term reflecting short-term transient changes, resulting in a trend term sequence and a disturbance term sequence.

[0030] The transient disturbance event identification specifically involves detecting the disturbance amplitude based on the disturbance term sequence using a robust threshold determination method, locating the time interval of the disturbance occurrence, forming a corresponding set of disturbance events, and obtaining a transient disturbance event description of the image features.

[0031] The perturbation spectrum feature extraction specifically involves calculating time-dimensional feature parameters such as perturbation intensity, perturbation duration, and perturbation change rate for the perturbation term sequence and perturbation event set, respectively, to obtain a perturbation spectrum feature vector characterizing transient perturbation behavior;

[0032] The generation of the perturbation spectrum feature set specifically involves summarizing or statistically analyzing the perturbation spectrum feature vectors corresponding to each image feature variation unit to form a perturbation spectrum feature set used to characterize the temporal evolution perturbation characteristics of the target tumor image features.

[0033] Furthermore, the determination of the credibility of pseudoprogression is used to comprehensively evaluate whether the changes in the target tumor image conform to the characteristics of pseudoprogression. Specifically, it integrates the spatial variation consistency reference data and the perturbation spectrum feature set, and jointly analyzes the consistency change characteristics of the image features in the spatial dimension and the perturbation behavior characteristics in the temporal dimension, so as to determine whether the image changes present a locally inconsistent evolutionary pattern with transient perturbation characteristics, and obtain the pseudoprogression credibility identification reference data.

[0034] The beneficial effects achieved by the present invention using the above solution are as follows:

[0035] (1) In response to the technical problems in existing tumor pseudoprogression identification systems, which generally rely on changes in tumor volume at a single time point or overall imaging response indicators, it is difficult to distinguish between pseudoprogression caused by post-treatment inflammatory response, immune cell infiltration and other factors and real tumor progression, and misjudgment is easy to occur when local imaging changes are inconsistent, this solution creatively adopts a comprehensive pseudoprogression credibility judgment process that combines a variant feature construction module, a variant consistency calculation module and a perturbation modeling module. This realizes collaborative modeling and joint judgment of tumor imaging changes in spatial and temporal dimensions, so that pseudoprogression can be accurately identified in typical scenarios where the overall volume temporarily increases but spatial variation is locally inconsistent and temporal evolution shows a transient perturbation decline, significantly reducing the risk of misjudgment in clinical follow-up and efficacy evaluation.

[0036] (2) In the existing process of constructing variant features, there is a problem that the fixed grid or simple superpixel segmentation method is usually used, which does not fully consider the stability differences of image features in the spatial and temporal dimensions. This leads to the problem that the segmentation is too coarse or too fine in the tumor edge area or areas with strong heterogeneity, and the features within the unit are unstable. This solution creatively adopts a spatiotemporal multi-feature adaptive variant unit construction method with stability improvement and optimization. It realizes the dynamic adjustment of the segmentation density according to the stability distribution of image features, so that more refined variant units are formed in areas with significant inflammatory response and frequent feature fluctuations after treatment, while maintaining a larger unit scale in the core area with relatively stable features. This results in a set of image feature variant units with stable features within the unit and significant differences between units, which improves the reliability of subsequent spatial and temporal analysis.

[0037] (3) In response to the technical problems in the existing process of calculating the consistency of variation, which often uses simple similarity statistics or local correlation analysis, it is difficult to distinguish between overall coordinated changes and local abnormal changes, and it is impossible to quantitatively characterize the impact of local non-coordinated changes on the overall image evolution pattern. This solution creatively adopts a spatial variation coordination analysis method based on the separation of structural consistency components. It realizes the explicit separation of the consistency component reflecting the overall coordinated change and the non-consistent perturbation component reflecting the local non-coordinated change through spatial variation consistency map and graph Laplacian spectrum decomposition. Thus, it can accurately identify the abnormal variation behavior of local areas when the overall image change trend is similar, and provide an effective quantitative basis for the spatial pattern of "significant local inflammatory response and reversible overall evolution" commonly seen in pseudo-progression.

[0038] (4) In the existing time evolution perturbation modeling process, there are technical problems that it is difficult to characterize the short-term transient fluctuations caused by treatment response and cannot distinguish between long-term real progression trends and short-term reversible perturbations by using single trend analysis or simple time smoothing methods. This solution creatively adopts a treatment response transient perturbation spectrum extraction method with trend perturbation dual channels. It decomposes the evolution process of image features over time into long-term trend terms and short-term transient perturbation terms, and performs spectral feature description of perturbation intensity, duration and rate of change. This can accurately capture the transient image perturbation behavior of "enhancement followed by decline" that is common after immunotherapy or radiotherapy, and provide stable and interpretable quantitative support for the determination of the time characteristics of pseudoprogression. Attached Figure Description

[0039] Figure 1 A schematic diagram of the structure of a tumor pseudoprogression identification system based on image feature variability provided by the present invention;

[0040] Figure 2 A flowchart illustrating the steps performed by the system provided for this invention;

[0041] Figure 3 A flowchart illustrating the steps performed by the module for constructing variant features;

[0042] Figure 4 A flowchart illustrating the steps performed by the mutation consistency calculation module;

[0043] Figure 5 A flowchart illustrating the steps performed by the disturbance modeling module.

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0046] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The present invention provides a tumor pseudoprogression identification system based on image feature variability, including a variant feature construction module, a variant consistency calculation module, a perturbation modeling module and a confidence determination module;

[0047] The variant feature construction module is used to construct image feature variant units. By performing feature stability-driven adaptive partitioning on the target tumor region, an image feature variant unit set is obtained, and the image feature variant unit set is sent to the variant consistency calculation module.

[0048] The mutation consistency calculation module is used for spatial mutation consistency calculation. It constructs a mutation consistency graph with mutation units as nodes and calculates the consistency components and non-consistency disturbance components to obtain spatial mutation consistency reference data. The spatial mutation consistency reference data is then sent to the disturbance modeling module and the credibility determination module.

[0049] The disturbance modeling module is used for time-evolution disturbance modeling. It obtains a disturbance spectrum feature set by decomposing the trend term and the disturbance term, and sends the disturbance spectrum feature set to the credibility determination module.

[0050] The credibility determination module is used to determine the credibility of false progress. It obtains the credibility identification reference data of false progress by fusing spatial variation consistency reference data and perturbation spectrum feature set.

[0051] By performing the above operations, this solution addresses the technical problems in existing tumor pseudoprogression identification systems, which generally rely on single-point tumor volume changes or overall imaging response indicators, making it difficult to distinguish between pseudoprogression caused by post-treatment inflammatory responses, immune cell infiltration, etc., and true tumor progression, and are prone to misjudgment when local imaging changes are inconsistent. This solution creatively adopts a comprehensive pseudoprogression credibility determination process that combines a variant feature construction module, a variant consistency calculation module, and a perturbation modeling module. It achieves collaborative modeling and joint determination of tumor imaging changes in both spatial and temporal dimensions, thereby accurately identifying pseudoprogression in typical scenarios where the overall volume temporarily increases but spatial variations are locally inconsistent and temporal evolution shows a transient perturbation decline, significantly reducing the risk of misjudgment in clinical follow-up and efficacy evaluation.

[0052] Example 2, this example is based on the above example, see reference. Figure 1 , Figure 2 and Figure 3 The image feature variation unit construction is used to structurally represent the local variation behavior of image features within the target tumor region. Specifically, it adopts a spatiotemporal multi-feature adaptive variation unit construction method with stability improvement and optimization to extract image features including grayscale, texture, edge intensity, and enhancement changes within the target tumor region. Based on the spatial stability and similarity relationship of the image features, the target tumor region is adaptively partitioned driven by feature stability to obtain a set of image feature variation units. The method includes the following steps: image data foundation construction, image feature analysis region limitation, multi-type image feature extraction, image feature stability characterization optimization, adaptive partitioning initial structure generation, and image feature variation unit generation.

[0053] The image data infrastructure is specifically constructed by acquiring multi-temporal image data of the target tumor at multiple time points before and after treatment, and performing spatiotemporal registration and intensity normalization on the multi-temporal image data to obtain a spatiotemporally aligned and intensity-consistent multi-temporal image dataset.

[0054] In a preferred embodiment, the multi-temporal image data includes at least pre-treatment baseline image data, early post-treatment follow-up image data, and delayed post-treatment follow-up image data; the spatiotemporal registration uses the pre-treatment baseline image data or the image data at the first follow-up time point as a reference image, and registers the image data at other time points to the same spatial coordinate system.

[0055] The spatiotemporal registration can be one or more of the following methods: rigid registration, affine registration, or non-rigid registration; the intensity normalization processing can be normalization processing based on reference tissues surrounding the tumor, standardization processing based on gray mean and standard deviation, or truncation normalization processing based on a preset gray percentile range.

[0056] The image feature analysis region limitation specifically involves determining the target tumor region based on the multi-temporal image dataset and generating a corresponding tumor region mask, thereby limiting the spatial range of subsequent image feature extraction and variation analysis.

[0057] In specific implementation, the tumor region mask can be generated manually, through semi-automatic threshold segmentation, region growth segmentation, or with the assistance of a medical image segmentation model. Preferably, when using model-assisted segmentation, a U-Net or 3DU-Net structure can be used to initially segment the target tumor region, and the final tumor region mask can be obtained by manual review or boundary correction. For multi-temporal image data, it is preferable to generate a basic tumor region mask on a reference temporal phase and map it to other temporal phases through registration transformation to ensure that the analysis regions at different time points have spatial correspondence.

[0058] The multi-type image feature extraction specifically involves extracting image feature vectors, including grayscale features, texture features, edge intensity features, and enhancement change features, for each temporal image within the tumor region mask, to obtain a multi-temporal image feature field of the tumor region.

[0059] Further preferably, the grayscale features include at least one of grayscale mean, grayscale standard deviation, grayscale quantile, or local grayscale contrast; the texture features include at least one of grayscale co-occurrence matrix features, local binary pattern features, wavelet texture features, or gradient direction texture features; the edge intensity features include at least one of gradient magnitude, edge density, edge direction consistency, or boundary sharpness; the enhancement change features include at least one of relative enhancement difference, enhancement peak, enhancement slope, enhancement area change rate, or enhancement difference before and after treatment. The above features are normalized according to a uniform scale and then combined to form a pixel-level image feature vector to avoid bias in subsequent stability calculations and unit aggregation due to dimensional differences between different types of features.

[0060] The image feature stability characterization optimization specifically involves calculating the local stability of image features in the spatial neighborhood and the consistency stability in the temporal dimension based on the multi-temporal image feature field, and combining edge intensity information to form an image feature stability distribution, thereby obtaining the image feature stability field of the tumor region.

[0061] In a preferred embodiment, the image feature stability field is used to characterize the degree of local feature fluctuation at different pixel locations within the tumor region. The lower the stability value, the more obvious the feature variation at that location in the spatial neighborhood or time dimension.

[0062] For any pixel p within the tumor region, its image feature stability value S(p) is calculated jointly by spatial local stability, temporal consistency stability, and edge constraint factor, and is specifically expressed as follows:

[0063] ;

[0064] In the formula, It is the image feature stability value of pixel p. It is a spatial stability weighting coefficient. It is the spatial neighborhood of pixel p. It is the image feature vector of pixel q in the neighborhood. It is a spatial scale adjustment parameter. It is a time stability weighting coefficient. It is the image feature vector of pixel p at time point t. It is a set of time points corresponding to multi-temporal image data. It is a time-scale adjustment parameter. These are the edge constraint weight coefficients. It is the edge intensity value at pixel p. It is the edge constraint scale adjustment parameter. , , satisfy Preferably, Take a value between 0.35 and 0.45. Take a value between 0.35 and 0.45. Take a value between 0.10 and 0.25;

[0065] The adaptive partitioning initial structure generation specifically involves increasing the partitioning density in regions with low stability and decreasing the partitioning density in regions with high stability, based on the image feature stability field, to generate a stability-guided adaptive partitioning seed set.

[0066] More preferably, in this step, the image feature stability field is normalized so that the value of S(p) is between 0 and 1; when the stability value of a pixel position is lower than the first stability threshold, the region where the position is located is identified as a high-variability region, and a smaller seed spacing is used to generate a segmentation seed; when the stability value of a pixel position is higher than the second stability threshold, the region where the position is located is identified as a stable region, and a larger seed spacing is used to generate a segmentation seed; when the stability value of a pixel position is between the first stability threshold and the second stability threshold, a default seed spacing is used to generate a segmentation seed; preferably, the first stability threshold is 0.35 to 0.45, and the second stability threshold is 0.65 to 0.75; the smaller seed spacing is 0.5 to 0.7 times the default seed spacing, and the larger seed spacing is 1.3 to 1.8 times the default seed spacing; thereby, the post-treatment inflammatory response, necrotic edges, and uneven enhancement regions obtain higher local segmentation resolution, while the tumor core region with relatively stable image features maintains a lower segmentation density;

[0067] The image feature variation unit generation specifically involves using the adaptive partitioning seed set as the initial center, aggregating and partitioning pixels within the tumor region based on image feature similarity, spatial adjacency, and image feature stability weights, and constraining and verifying the stability within the unit to obtain an image feature variation unit set that satisfies the condition of stable features within the unit and significant feature differences between units.

[0068] In a preferred embodiment, when generating image feature mutation units, pixel p is assigned to the mutation unit corresponding to the k-th adaptive partitioning seed, and an aggregation cost function is constructed for determination to obtain a set of image feature mutation units. The calculation formula of the aggregation cost function is as follows:

[0069] ;

[0070] In the formula, It is the aggregation cost of pixel p belonging to the kth mutation unit. It is the feature similarity weight coefficient. It is the image feature vector of pixel p. It is the feature center of the k-th mutation unit. It is a spatial compactness weighting coefficient. These are the spatial coordinates of pixel p. It is the spatial center of the k-th mutation unit. It is the square of the scale parameter of the k-th variant unit. It is the stability and consistency weighting coefficient. It is the image feature stability value of pixel p. It is the average stability value within the kth mutation unit;

[0071] Preferably, , , All are non-negative weight parameters, and satisfy the following conditions: ,in, A value of 0.45 to 0.60 is used to ensure that image features within the same variant unit have high similarity; A value of 0.20 to 0.35 is used to limit excessive discretization of the cell space. A value of 0.15 to 0.30 is used to constrain pixel stability levels to be similar within the same variation unit;

[0072] For imaging data with indistinct tumor boundaries or significant post-treatment inflammatory response, the level of contrast can be appropriately increased. The value of is chosen to enhance the ability to fine-grainedly divide local unstable regions;

[0073] Assign pixel p to make The smallest mutation unit, and updated iteratively. , and Continue until the unit partitioning results converge;

[0074] The iterative update includes: recalculating the feature center, spatial center, and average stability value of each mutation unit based on the current pixel allocation result, and recalculating the aggregation cost function based on the updated center parameters; when the pixel allocation change ratio between two consecutive iterations is less than a preset change threshold, or the number of iterations reaches a preset maximum number of iterations, the unit partitioning result is determined to converge; preferably, the preset change threshold is 1% to 3%, and the preset maximum number of iterations is 10 to 30.

[0075] After the initial aggregation and segmentation are completed, the stability constraint within each image feature variation unit is checked and the difference between units is checked. If the variance of the image features within a variation unit is higher than the preset variance threshold within the unit, or the dispersion of the stability value within the variation unit is higher than the preset stability dispersion threshold, then the variation unit is further subdivided; if the feature center distance between two adjacent variation units is lower than the preset difference between units threshold, and the average stability difference between the two is less than the preset stability difference threshold, then the two adjacent variation units are merged.

[0076] Preferably, the intra-unit variance threshold is determined based on the 75th quantile of the variance of image features within the tumor region, the stability discrete threshold is between 0.10 and 0.20, and the inter-unit difference threshold is determined based on the 25th quantile of the distance between the feature centers of all adjacent units.

[0077] By performing the above operations, this solution addresses the technical problem that existing variant feature construction methods often employ fixed grids or simple superpixel segmentation, which fail to fully consider the stability differences of image features in the spatial and temporal dimensions. This leads to overly coarse or fine segmentation and unstable features within units in tumor margin regions or highly heterogeneous areas. This solution creatively adopts a stability-improved and optimized spatiotemporal multi-feature adaptive variant unit construction method. It achieves dynamic adjustment of segmentation density based on the stability distribution of image features, resulting in finer variant units in regions with significant post-treatment inflammatory responses and frequent feature fluctuations, while maintaining a larger unit scale in the relatively stable core regions. This yields a set of image feature variant units with stable features within units and significant differences between units, improving the reliability of subsequent spatial and temporal analyses.

[0078] Example 3, this example is based on the above examples, see reference. Figure 1 , Figure 2 and Figure 4The spatial variation consistency calculation is used to evaluate the spatial cooperative change relationship between various image feature variation units within the target tumor region. Specifically, it constructs a spatial variation consistency map using the set of image feature variation units as nodes, and adopts a spatial variation cooperative analysis method based on structural consistency component separation. According to the similarity of different image feature variation units in image feature change direction, change amplitude and spatial adjacency relationship, the spatial variation consistency map is analyzed to calculate the consistency component reflecting the overall cooperative change characteristics and the non-consistency perturbation component reflecting the local non-cooperative change characteristics, and obtains spatial variation consistency reference data. The calculation includes the following steps: unit-level variation feature aggregation, spatial variation consistency map construction, spatial cooperative change relationship characterization, consistency perturbation structural component separation and spatial variation consistency reference data generation.

[0079] The unit-level variation feature aggregation is specifically based on the image feature variation unit set, performing aggregation processing on the multi-temporal image features of each image feature variation unit, extracting unit-level variation feature vectors that reflect the direction and magnitude of image feature changes, and obtaining the unit-level variation feature set of the image feature variation unit.

[0080] In a preferred embodiment, for the i-th image feature variation unit, the image feature vectors of each pixel in the variation unit are first aggregated at the unit level to obtain the unit-level image feature vector of the variation unit at the t-th time point. The unit-level aggregation can adopt mean aggregation, median aggregation, or weighted aggregation based on stability weights. Preferably, when there are local noise or unstable pixels at the boundary in the unit, a weighted aggregation method based on the stability value of image features is adopted to reduce the interference of unstable pixels on the unit-level features.

[0081] The formula for calculating the unit-level image feature vector is as follows:

[0082] ;

[0083] In the formula, It is the unit-level image feature vector of the i-th image feature variation unit at time point t. It is the set of pixels corresponding to the i-th image feature variation unit. It is the image feature stability value of pixel p. It is the image feature vector of pixel p at time point t;

[0084] After obtaining the unit-level image feature vectors at each time point, using the pre-treatment baseline time point as the reference time point, the feature changes at the post-treatment follow-up time points relative to the baseline time point are calculated, and the changes at different follow-up time points are combined to form a unit-level variation feature vector. Preferably, when the multi-temporal image data includes the pre-treatment baseline time point, the early post-treatment follow-up time point, and the delayed post-treatment follow-up time point, the formula for calculating the unit-level variation feature vector of the i-th image feature variation unit is as follows:

[0085] ;

[0086] In the formula, It is the unit-level variation feature vector of the i-th image feature variation unit. This is the baseline time point before treatment. It is the unit-level image feature vector corresponding to the baseline time point before treatment. This is the early follow-up time point after treatment. It is the unit-level image feature vector corresponding to the early follow-up time point after treatment. It is a delayed follow-up time point after treatment. It is the unit-level image feature vector corresponding to the delayed follow-up time point after treatment. This vector simultaneously represents the early change magnitude, the delayed change magnitude, and the decline or continuous change trend during the follow-up period, providing a basis for subsequent judgment of the consistency of change direction and similarity of change magnitude among different variant units.

[0087] The construction of the spatial variation consistency graph specifically involves using the image feature variation units as graph nodes, establishing node connections based on the spatial adjacency relationships between the image feature variation units, and assigning weights to the connections between nodes based on the similarity of the unit-level variation feature vectors in the direction and magnitude of image feature changes, thereby obtaining the spatial variation consistency graph.

[0088] The spatial adjacency relationship is determined based on the boundary contact relationship and spatial distance relationship between image feature variant units. Specifically, if two image feature variant units share a boundary within the tumor region mask, or if the spatial center distance between two image feature variant units is less than a preset adjacency distance threshold, then it is determined that the two have a spatial adjacency relationship; otherwise, it is determined that the two do not have a spatial adjacency relationship.

[0089] Preferably, the preset adjacency distance threshold can be taken as 1.2 to 1.8 times the average equivalent diameter of all variant units. This is achieved by simultaneously considering boundary contact relationships and spatial center distance relationships.

[0090] Preferably, the spatial variation consistency graph is denoted as G=(V,E,W), where V is the set of image feature variation unit nodes, E is the set of spatially adjacent edges, and W is the edge weight matrix of the spatial variation consistency graph; for the i-th variation unit and the j-th variation unit that have a spatial adjacency relationship, the formula for calculating their edge weights is:

[0091] ;

[0092] In the formula, It is the edge weight between the i-th image feature variation unit and the j-th variation unit, which have a spatial adjacency relationship. This is an identifier indicating whether the i-th image feature variation unit and the j-th variation unit are spatially adjacent; it takes a value of 1 if they are adjacent, and a value of 0 otherwise. It is the unit-level variation feature vector of the i-th image feature variation unit. It is the unit-level mutation feature vector of the j-th mutation unit. The whole is used to characterize the difference in the magnitude of change. The whole is used to characterize the consistency of the direction of change. It is the square of the scaling parameter of the variation magnitude;

[0093] Wherein, the larger the value of the edge weight, the more spatially adjacent the i-th variant unit and the j-th variant unit are, and the similar the magnitude and direction of their image feature changes; the smaller the value of the edge weight, the more spatially adjacent they may be, but the larger the difference in their magnitude or the different directions of their changes. Preferably, the variation magnitude scale parameter is determined based on the median or mean of the feature difference distance between all adjacent variant units to accommodate the differences in the magnitude of image feature changes between different cases. Furthermore, to avoid isolated noise edges affecting the graph structure analysis, edge weights less than a preset edge weight threshold can be set to zero. The preset edge weight threshold is preferably between 0.05 and 0.15.

[0094] The spatial collaborative change relationship characterization specifically involves performing graph structure analysis on the spatial variation consistency map and constructing a graph Laplacian matrix that reflects the node connection relationship to characterize the overall collaborative change characteristics of image feature variation in the spatial dimension.

[0095] The separation of the consistent perturbation structural components specifically involves performing spectral decomposition on the graph Laplacian matrix to extract consistent components that reflect the overall cooperative change characteristics and non-consistent perturbation components that reflect the local non-cooperative change characteristics.

[0096] In a preferred embodiment, the degree matrix D is constructed based on the edge weight matrix W. g , where the degree matrix D gThe i-th diagonal element represents the sum of all edge weights connected to the i-th mutation unit. Then, a normalized graph Laplacian matrix is ​​constructed to characterize the smoothing relationship between adjacent mutation units in the spatial variation consistency graph. The calculation formula is:

[0097] ;

[0098] In the formula, L is the normalized graph Laplacian matrix, I is the identity matrix, W is the edge weight matrix of the spatial variation consistency graph, and D... g It is a degree matrix. By constructing a normalized graph Laplacian matrix, the influence of differences in the number of connections between different nodes on the subsequent spectral decomposition results can be reduced, making the representation of spatial cooperative change relationships more stable.

[0099] In a preferred embodiment, the normalized graph Laplacian matrix L is subjected to spectral decomposition to obtain the graph basis vectors and corresponding eigenvalues, calculated using the following formula:

[0100] ;

[0101] In the formula, It is composed of eigenvectors The resulting spectral basis matrix It is composed of eigenvalues The resulting eigenvalue matrix consists of N, which is the number of image feature variation units, and diag, which is the matrix diagonal function.

[0102] Arrange the unit-level variation feature vectors of all image feature variation units in node order to construct a unit-level variation feature matrix U, where the i-th row of U corresponds to the unit-level variation feature vector of the i-th variation unit. Based on the spectral basis matrix Φ, project the unit-level variation feature matrix U onto the low-frequency spectral subspace to obtain a consistency component U reflecting the overall coordinated change. c The residual between the original variant feature matrix and the consistent component is used as the non-consistent perturbation component U. d The calculation formula is:

[0103] ;

[0104] ;

[0105] In the formula, It is a consistent component. These are non-uniform disturbance components. It is a matrix composed of the first r low-frequency spectrum basis vectors, where r represents the low-frequency spectrum truncation order. It is the transpose of a matrix composed of the first r low-frequency spectrum basis vectors. Since the low-frequency spectrum subspace corresponds to a smooth variation pattern on the spatial variation consistency map, therefore... Used to represent the overall coordinated change presented by multiple adjacent variant units; and It is used to represent local residual changes that cannot be explained by low-frequency cooperative modes, and thus can characterize local non-cooperative changes or anomalous disturbances;

[0106] More preferably, the low-frequency spectrum truncation order r is determined based on the spectrum energy accumulation ratio or a preset ratio;

[0107] Preferably, the cumulative proportion of the atlas energy can be determined based on the projection energy of the unit-level variation feature matrix U onto each atlas basis vector. Specifically, after projecting U onto each atlas basis vector, the square of the corresponding projection norm is calculated, and the smallest integer with a cumulative projection energy proportion of 70% to 85% is selected as r. When the number of image feature variation units is small, r can also be set to 10% to 25% of the number of nodes N. In this way, the main overall coordinated change patterns within the target tumor region are preserved, while avoiding excessive absorption of local non-cooperative perturbations into the consistency component.

[0108] The generation of the spatial variation consistency reference data specifically involves calculating the consistency index and inconsistency perturbation index of image feature variations in the spatial dimension based on the consistency component and the inconsistency perturbation component, thereby obtaining spatial variation consistency reference data used to characterize the spatial variation synergy of the target tumor region.

[0109] In a preferred embodiment, the spatial variation consistency reference data includes a spatial consistency index Q. c Inconsistent disturbance index Q d The spatial consistency index is used to characterize the relative strength of the overall coordinated change components relative to the original variation characteristics, and the non-coordinated perturbation index is used to characterize the relative strength of the local non-coordinated change components relative to the original variation characteristics. The specific calculation formula is as follows:

[0110] ;

[0111] ;

[0112] In the formula, It is a spatial consistency indicator. It is a non-consistent disturbance indicator. It is the Frobenius norm. It is a small constant to prevent the denominator from being zero, and is preferably 10. −6 Up to 10 −4 ;when higher and When the level is low, it indicates that the imaging changes within the target tumor region mainly show overall coordinated changes; when When the elevation and perturbation units are concentrated in local areas, it indicates the presence of local non-cooperative variation behavior within the target tumor region;

[0113] Furthermore, the distribution data of perturbation units can be determined based on the amplitude of the non-uniform perturbation component corresponding to each variant unit. When the amplitude of the non-uniform perturbation of a certain variant unit is higher than the 75th percentile of the perturbation amplitudes of all variant units, the variant unit is marked as a local perturbation unit. By outputting the location distribution, quantity ratio, and perturbation amplitude statistics of local perturbation units, a spatial dimension reference is provided for subsequent determination of the credibility of false progress.

[0114] By performing the above operations, this solution addresses the technical problem that existing variation consistency calculations often rely on simple similarity statistics or local correlation analysis, making it difficult to distinguish between overall coordinated changes and local anomalous changes, and failing to quantitatively characterize the impact of local non-coordinated changes on the overall image evolution pattern. This solution creatively employs a spatial variation coordination analysis method based on the separation of structural consistency components. Through spatial variation consistency maps and graphical Laplacian spectral decomposition, it explicitly separates the consistency components reflecting overall coordinated changes from the non-consistent perturbation components reflecting local non-coordinated changes. This allows for accurate identification of anomalous variation behavior in local areas even when the overall image change trend is similar, providing effective quantitative evidence for the spatial pattern of "significant local inflammatory response and reversible overall evolution" commonly found in pseudo-progression.

[0115] Example 4, this example is based on the above examples, see below. Figure 1 and Figure 5 The aforementioned time evolution perturbation modeling is used to characterize the transient perturbation behavior caused by treatment response during the change of target tumor image features over time. Specifically, based on multi-temporal image data, a treatment response transient perturbation spectrum extraction method with trend perturbation dual channels is used to construct a feature sequence of image features changing over time. The feature sequence is then decomposed to decompose the time evolution process of image features into a trend term reflecting long-term change trends and a perturbation term reflecting short-term fluctuations. The perturbation term is analyzed to extract the perturbation intensity, duration, and change characteristics to obtain a perturbation spectrum feature set. The process includes the following steps: unit-level time series construction, time axis processing, differentiation between trend term and transient perturbation term, transient perturbation event identification, perturbation spectrum feature extraction, and perturbation spectrum feature set generation.

[0116] The construction of the unit-level time series is specifically based on multi-temporal image data. At the scale of image feature variation units, the features of each temporal image are aggregated to form a unit-level feature sequence that reflects the changes of image features over time. The corresponding timestamp information is recorded to obtain the unit-level image feature time series.

[0117] The time axis processing specifically involves performing time axis normalization on the unit-level image feature time series and using a robust smoothing strategy to denoise the feature series in order to reduce the impact of unequal time intervals and imaging noise on subsequent analysis, thereby obtaining a robust image feature time series.

[0118] The distinction between the trend term and the transient disturbance term is specifically achieved by performing component decomposition on the robust image feature time series. Under the combined effect of trend smoothing constraints and disturbance sparsity constraints, the temporal evolution process of the image features is separated into a trend term reflecting long-term change trends and a disturbance term reflecting short-term transient changes, resulting in a trend term sequence and a disturbance term sequence.

[0119] In a preferred embodiment, the robust image feature time series of the i-th image feature variation unit is decomposed into a long-term trend term sequence and a transient disturbance term sequence, and solved by constructing an optimization objective. The calculation formula is as follows:

[0120] ;

[0121] In the formula, It is a long-term trend sequence. It is a sequence of transient perturbation terms. It is the robust image feature time series of the i-th image feature variation unit. This is the trend smoothing constraint weighting coefficient, with a preferred value range of 0.1 to 1.0. It is the second difference of the trend term sequence, used to constrain the smoothing of long-term trend changes. This is the perturbation sparse constraint weight coefficient, with a preferred value range of 0.05 to 0.5. It is the L1 norm of the transient perturbation term, used to constrain the sparsity of the transient perturbation term. This is the perturbation continuity constraint weighting coefficient, with a preferred value range of 0.05 to 0.3. It is the first-order difference of the transient disturbance term, used to constrain the local continuity of the disturbance between adjacent time points;

[0122] It should be noted that when the unit-level image feature time series is a multi-dimensional feature series, the trend perturbation decomposition of each feature dimension can be performed separately and then statistically summarized. Alternatively, a unit-level comprehensive response sequence can be formed first based on preset feature weights or principal component analysis, and then the trend perturbation decomposition of the unit-level comprehensive response sequence can be performed.

[0123] The transient disturbance event identification specifically involves detecting the disturbance amplitude based on the disturbance term sequence using a robust threshold determination method, locating the time interval of the disturbance occurrence, forming a corresponding set of disturbance events, and obtaining a transient disturbance event description of the image features.

[0124] The robustness threshold is preferably determined based on the median absolute deviation of the perturbation term sequence;

[0125] The perturbation spectrum feature extraction specifically involves calculating time-dimensional feature parameters such as perturbation intensity, perturbation duration, and perturbation change rate for the perturbation term sequence and perturbation event set, respectively, to obtain a perturbation spectrum feature vector characterizing transient perturbation behavior;

[0126] In a preferred embodiment, for the m-th transient perturbation event identified in the i-th image feature variation unit, a perturbation spectrum feature vector is constructed, calculated using the following formula:

[0127] ;

[0128] In the formula, It is the perturbation spectrum feature vector corresponding to the m-th transient perturbation event for the i-th image feature variation unit. It is the peak intensity of the disturbance. It is the duration of the disturbance. It is the rate of rise of the disturbance. It is the rate of return of the disturbance. The disturbance fallback ratio is determined based on the decrease between the peak amplitude of the disturbance event and the lowest amplitude after the peak. It is used to characterize whether the image features fall back after the disturbance peak appears. Preferably, when the disturbance fallback ratio is greater than 0.30 to 0.50 and the duration of the disturbance is within a preset follow-up time window, the disturbance event is marked as a disturbance event with transient fallback characteristics.

[0129] The generation of the perturbation spectrum feature set specifically involves summarizing or statistically analyzing the perturbation spectrum feature vectors corresponding to each image feature variation unit to form a perturbation spectrum feature set used to characterize the temporal evolution perturbation characteristics of the target tumor image features.

[0130] In a preferred embodiment, the perturbation spectrum feature set includes perturbation intensity statistics, perturbation duration statistics, perturbation fallback ratio statistics, and transient fallback perturbation unit ratio. The transient fallback perturbation unit ratio represents the proportion of image feature variant units with transient fallback characteristics to the total number of image feature variant units. It is used to reflect the degree of short-term reversible fluctuation caused by treatment response from the overall level of the target tumor. Thus, the perturbation spectrum feature set includes not only the perturbation behavior of individual variant units, but also the temporal perturbation distribution characteristics of the entire target tumor region.

[0131] By performing the above operations, this solution addresses the technical problem in existing time-evolution perturbation modeling processes, where the use of single-trend analysis or simple time smoothing methods is often insufficient to characterize short-term transient fluctuations caused by treatment response and to distinguish between long-term true progression trends and short-term reversible perturbations. This solution creatively employs a dual-channel trend perturbation method for extracting the treatment response transient perturbation spectrum. This method decomposes the evolution of image features over time into long-term trend terms and short-term transient perturbation terms, and provides spectral characterization of perturbation intensity, duration, and rate of change. This allows for accurate capture of the "enhancement followed by decline" transient image perturbation behavior commonly seen after immunotherapy or radiotherapy, providing stable and interpretable quantitative support for determining the temporal characteristics of pseudoprogression.

[0132] Example 5, this example is based on the above examples, see below. Figure 1 and Figure 2 The aforementioned false progression confidence determination is used to comprehensively evaluate whether the changes in the target tumor image conform to the characteristics of false progression. Specifically, it integrates the spatial variation consistency reference data and the perturbation spectrum feature set, and jointly analyzes the consistency change characteristics of the image features in the spatial dimension and the perturbation behavior characteristics in the temporal dimension, so as to determine whether the image changes present a locally inconsistent evolutionary pattern with transient perturbation characteristics, and obtain false progression confidence identification reference data.

[0133] In this embodiment, the determination of the credibility of pseudoprogression is based on the obtained spatial and temporal analysis results, and a comprehensive assessment is made on whether the changes in the target tumor image conform to the characteristics of pseudoprogression.

[0134] Specifically, it receives spatial variation consistency reference data obtained from the spatial variation consistency calculation step and a set of perturbation spectrum features obtained from the temporal evolution perturbation modeling step, and on this basis, it performs joint analysis on the consistency change features of image features in the spatial dimension and the perturbation behavior features in the temporal dimension.

[0135] In a preferred embodiment, the false progress confidence identification reference data includes a false progress confidence score, which is calculated jointly based on a spatial inconsistency disturbance index, a temporal transient disturbance index, and a disturbance fallback confirmation index. The specific calculation formula is as follows:

[0136] ;

[0137] In the formula, It is the confidence score for false progress; Sigmoid is the Sigmoid function. It is a bias term. The non-consistent perturbation fusion weight coefficients It is a non-uniformity perturbation index calculated from the spatial variation consistency. These are the weighting coefficients for time-transient perturbation fusion. It is a time-transient perturbation intensity index obtained from the perturbation spectrum feature set. It is the disturbance fallback confirmation fusion weight coefficient. It is a disturbance fallback confirmation index obtained from the disturbance spectrum feature set. It is the spatial consistency fusion weight coefficient. It is a spatial consistency indicator, because pseudo-progression usually manifests as localized non-uniform changes and short-term reversible perturbations, therefore , and As a positive indicator; and because overall, consistent changes more closely resemble the actual progress in imaging patterns, therefore It is used as a negative indicator in the calculation;

[0138] Among them, the time transient disturbance intensity index The results can be obtained by normalizing the peak intensity of each transient disturbance event, preferably by taking the mean, upper quartile, or maximum value of the peak intensity of all disturbance events.

[0139] The disturbance fallback confirmation indicator It can be determined based on the proportion of disturbance events with transient fallback characteristics, the proportion of transient fallback disturbance units, or the statistical value of the disturbance fallback proportion;

[0140] In a preferred embodiment, the fusion weighting coefficients can be determined based on training with historical labeled case data, or they can be calibrated after setting initial values ​​based on expert experience; preferably, in the absence of large-scale labeled data... and Take a value between 0.25 and 0.35. Take a value between 0.20 and 0.30. The value is set to 0.15 to 0.25 to allow spatial inconsistency, temporal transient perturbation, and perturbation fallback characteristics to participate in the credibility determination, while preserving the inhibitory effect of overall continuous consistency changes on the determination of false progress.

[0141] Meanwhile, the bias term Calibration can be performed based on the prior ratio of false progress to true progress in the training samples, or adjusted based on the distribution of high, medium, and low confidence intervals on the validation set; during initialization, [the following can be done]: Set to -0.2 to 0.2, and perform secondary calibration using case data;

[0142] Preferably, when When, output a result indicating high confidence level for false progress; when When, output the confidence level of moderate pseudo-progression; when At that time, a low false progression confidence level result is output. The threshold can be adjusted according to different tumor types, treatment methods, or follow-up periods;

[0143] When spatial variability consistency reference data shows that the imaging changes within the target tumor region exhibit localized non-uniformity, and the perturbation spectrum feature set indicates that the imaging features, during their temporal changes, have perturbation intensity exceeding a preset threshold, perturbation duration within a preset time window, and exhibit transient perturbation characteristics of initial enhancement followed by decline, , and The corresponding increase leads to a higher confidence score for false progress. The tumor tends to increase; conversely, when the imaging changes within the target tumor region show a consistent and continuous overall change, and no significant transient decline disturbance event is detected, The corresponding increase leads to a higher confidence score for false progress. Suppressed;

[0144] Using the above method, the local non-uniformity features in the spatial dimension and the transient decline features in the temporal dimension can be transformed into continuous confidence scores, and the false progress judgment results of high confidence, medium confidence or low confidence can be output according to the score range.

[0145] Finally, based on the above joint analysis results, reference data for identifying the credibility of pseudoprogression was generated to characterize the credibility of target tumor imaging changes to pseudoprogression characteristics, serving as a reference for subsequent clinical evaluation or follow-up decisions.

[0146] It should be noted that the aforementioned reference data for identifying the reliability of pseudoprogression is used to provide auxiliary reference for clinical efficacy evaluation, follow-up examination arrangements, or further examinations, and does not directly replace the final diagnostic conclusion made by physicians based on pathological examination, clinical symptoms, laboratory indicators, and multidisciplinary consultation results.

[0147] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process or method.

[0148] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0149] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A tumor pseudoprogression identification system based on image feature variability, characterized in that: It includes a mutation feature construction module, a mutation consistency calculation module, a perturbation modeling module, and a credibility determination module; The variant feature construction module is used to construct image feature variant units. By performing feature stability-driven adaptive partitioning on the target tumor region, an image feature variant unit set is obtained, and the image feature variant unit set is sent to the variant consistency calculation module. The image feature variation unit construction is used to structurally represent the local variation behavior of image features within the target tumor region. Specifically, it adopts a spatiotemporal multi-feature adaptive variation unit construction method with stability improvement and optimization to extract image features including grayscale, texture, edge intensity, and enhancement changes within the target tumor region. Based on the spatial stability and similarity relationship of the image features, the target tumor region is adaptively partitioned driven by feature stability to obtain a set of image feature variation units. The method includes the following steps: image data foundation construction, image feature analysis region limitation, multi-type image feature extraction, image feature stability characterization optimization, adaptive partitioning initial structure generation, and image feature variation unit generation. The mutation consistency calculation module is used for spatial mutation consistency calculation. It constructs a mutation consistency graph with mutation units as nodes and calculates the consistency components and non-consistency disturbance components to obtain spatial mutation consistency reference data. The spatial mutation consistency reference data is then sent to the disturbance modeling module and the credibility determination module. The spatial variation consistency calculation is used to evaluate the spatial cooperative change relationship among various imaging feature variation units within the target tumor region. Specifically, it involves constructing a spatial variation consistency map using the set of imaging feature variation units as nodes, and employing a spatial variation cooperative analysis method based on structural consistency component separation. The spatial variation consistency map is analyzed based on the similarity of different imaging feature variation units in the direction, magnitude, and spatial adjacency of image feature changes. The calculation of the consistency component reflecting the overall cooperative change characteristics and the non-consistent perturbation component reflecting local non-cooperative change characteristics yields spatial variation consistency reference data. The process includes the following steps: unit-level variation feature aggregation, spatial variation consistency map construction, spatial cooperative change relationship characterization, consistency perturbation structural component separation, and spatial variation consistency reference data generation. The disturbance modeling module is used for time-evolution disturbance modeling. It obtains a disturbance spectrum feature set by decomposing the trend term and the disturbance term, and sends the disturbance spectrum feature set to the credibility determination module. The aforementioned time-evolution perturbation modeling is used to characterize the transient perturbation behavior caused by treatment response during the change of target tumor image features over time. Specifically, based on multi-temporal image data, a treatment response transient perturbation spectrum extraction method with trend perturbation dual channels is used to construct a feature sequence of image features changing over time. The feature sequence is then decomposed to decompose the time evolution process of image features into a trend term reflecting long-term change trends and a perturbation term reflecting short-term fluctuations. The perturbation term is analyzed to extract the perturbation intensity, duration, and change characteristics to obtain a perturbation spectrum feature set. The process includes the following steps: unit-level time series construction, time axis processing, differentiation between trend term and transient perturbation term, transient perturbation event identification, perturbation spectrum feature extraction, and perturbation spectrum feature set generation. The credibility determination module is used to determine the credibility of false progress. It obtains the credibility identification reference data of false progress by fusing spatial variation consistency reference data and perturbation spectrum feature set.

2. The tumor pseudoprogression identification system based on image feature variability according to claim 1, characterized in that: The image data infrastructure is specifically constructed by acquiring multi-temporal image data of the target tumor at multiple time points before and after treatment, and performing spatiotemporal registration and intensity normalization on the multi-temporal image data to obtain a spatiotemporally aligned and intensity-consistent multi-temporal image dataset. The image feature analysis region limitation specifically involves determining the target tumor region based on the multi-temporal image dataset and generating a corresponding tumor region mask, thereby limiting the spatial range of subsequent image feature extraction and variation analysis. The multi-type image feature extraction specifically involves extracting image feature vectors, including grayscale features, texture features, edge intensity features, and enhancement change features, for each temporal image within the tumor region mask, to obtain a multi-temporal image feature field of the tumor region. The image feature stability characterization optimization specifically involves calculating the local stability of image features in the spatial neighborhood and the consistency stability in the temporal dimension based on the multi-temporal image feature field, and combining edge intensity information to form an image feature stability distribution, thereby obtaining the image feature stability field of the tumor region. The adaptive partitioning initial structure generation specifically involves increasing the partitioning density in regions with low stability and decreasing the partitioning density in regions with high stability, based on the image feature stability field, to generate a stability-guided adaptive partitioning seed set. The generation of the image feature variation unit specifically involves using the adaptive partitioning seed set as the initial center, aggregating and partitioning pixels within the tumor region based on image feature similarity, spatial adjacency, and image feature stability weights, and constraining and verifying the stability within the unit to obtain a set of image feature variation units that satisfy the requirements of stable features within the unit and significant feature differences between units.

3. The tumor pseudoprogression identification system based on image feature variability according to claim 2, characterized in that: The unit-level variation feature aggregation is specifically based on the image feature variation unit set, performing aggregation processing on the multi-temporal image features of each image feature variation unit, extracting unit-level variation feature vectors that reflect the direction and magnitude of image feature changes, and obtaining the unit-level variation feature set of the image feature variation unit. The construction of the spatial variation consistency graph specifically involves using the image feature variation units as graph nodes, establishing node connections based on the spatial adjacency relationships between the image feature variation units, and assigning weights to the connections between nodes based on the similarity of the unit-level variation feature vectors in the direction and magnitude of image feature changes, thereby obtaining the spatial variation consistency graph. The spatial collaborative change relationship characterization specifically involves performing graph structure analysis on the spatial variation consistency map and constructing a graph Laplacian matrix that reflects the node connection relationship to characterize the overall collaborative change characteristics of image feature variation in the spatial dimension. The separation of the consistent perturbation structural components specifically involves performing spectral decomposition on the graph Laplacian matrix to extract consistent components that reflect the overall cooperative change characteristics and non-consistent perturbation components that reflect the local non-cooperative change characteristics. The generation of the spatial variation consistency reference data specifically involves calculating the consistency index and the inconsistency perturbation index of image feature variations in the spatial dimension based on the consistency component and the inconsistency perturbation component, thereby obtaining spatial variation consistency reference data used to characterize the spatial variation synergy of the target tumor region.

4. The tumor pseudoprogression identification system based on image feature variability according to claim 3, characterized in that: The construction of the unit-level time series is specifically based on multi-temporal image data. At the scale of image feature variation units, the features of each temporal image are aggregated to form a unit-level feature sequence that reflects the changes of image features over time. The corresponding timestamp information is recorded to obtain the unit-level image feature time series. The time axis processing specifically involves performing time axis normalization on the unit-level image feature time series and using a robust smoothing strategy to denoise the feature series in order to reduce the impact of unequal time intervals and imaging noise on subsequent analysis, thereby obtaining a robust image feature time series. The distinction between the trend term and the transient disturbance term is specifically achieved by performing component decomposition on the robust image feature time series. Under the combined effect of trend smoothing constraints and disturbance sparsity constraints, the temporal evolution process of the image features is separated into a trend term reflecting long-term change trends and a disturbance term reflecting short-term transient changes, resulting in a trend term sequence and a disturbance term sequence. The transient disturbance event identification specifically involves detecting the disturbance amplitude based on the disturbance term sequence using a robust threshold determination method, locating the time interval of the disturbance occurrence, forming a corresponding set of disturbance events, and obtaining a transient disturbance event description of the image features. The perturbation spectrum feature extraction specifically involves calculating time-dimensional feature parameters such as perturbation intensity, perturbation duration, and perturbation change rate for the perturbation term sequence and perturbation event set, respectively, to obtain a perturbation spectrum feature vector characterizing transient perturbation behavior; The generation of the perturbation spectrum feature set specifically involves summarizing or statistically analyzing the perturbation spectrum feature vectors corresponding to each image feature variation unit to form a perturbation spectrum feature set used to characterize the temporal evolution perturbation characteristics of the target tumor image features.

5. A tumor pseudoprogression identification system based on image feature variability according to claim 4, characterized in that: The pseudoprogression confidence determination is used to comprehensively evaluate whether the changes in the target tumor image conform to the characteristics of pseudoprogression. Specifically, it integrates the spatial variation consistency reference data and the perturbation spectrum feature set, and jointly analyzes the consistency change characteristics of the image features in the spatial dimension and the perturbation behavior characteristics in the temporal dimension, so as to determine whether the image changes present a locally inconsistent evolutionary pattern with transient perturbation characteristics, and obtain pseudoprogression confidence identification reference data.

Images