Tumor development analysis method and system based on image recognition

Abstract

The invention relates to the technical field of medical image processing, in particular to a tumor development analysis method and system based on image recognition. The method comprises the following steps: acquiring a multi-temporal tumor image data set, analyzing the multi-temporal tumor image data set, and determining spatial probabilistic boundary representation information; according to the spatial probabilistic boundary representation information, performing multi-dimensional feature fusion analysis on the tumor boundary, and determining tumor development situation feature information; performing analogue simulation on the tumor development situation according to the tumor development situation feature information, and determining tumor development risk information; and analyzing the tumor development risk information, and determining and outputting dynamic follow-up visit suggestion information. According to the method and the device, the spatial probabilistic boundary representation information is constructed, so that the misjudgment on the tumor boundary caused by binarization decision due to the forced closed contour tendency of an image recognition algorithm is avoided, the boundary illusion generated when the tumor boundary is analyzed in the prior art is effectively eliminated, and the evaluation accuracy of the tumor development situation is improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method and system for tumor development analysis based on image recognition. Background Technology

[0002] In modern cancer treatment, analyzing tumor development is a crucial technical means to ensure the accuracy and timeliness of the treatment process. While traditional biopsy can provide direct evidence of tumor status, it carries the risk of cancer cell metastasis and can easily cause patient panic and subsequent complications. Non-invasive tumor development analysis technology based on image recognition is a key technology to reduce the risks in the tumor status analysis process.

[0003] However, existing tumor development analysis technologies based on image recognition are prone to creating illusions about tumor boundaries due to the inherent tendency of image recognition technology to force closed contours during the analysis of tumor development trends. This can lead to misjudgments of tumor development trends and result in low reference value of the analysis results. Summary of the Invention

[0004] This application provides a tumor development analysis method and system based on image recognition to solve the above-mentioned technical problems.

[0005] In a first aspect, this application provides a tumor development analysis method based on image recognition, the method comprising: Acquire a multi-temporal tumor image dataset, analyze the multi-temporal tumor image dataset, and determine the spatial probabilistic boundary representation information; Based on the spatial probabilistic boundary representation information, a multi-dimensional feature fusion analysis is performed on the tumor boundary to determine the tumor development trend characteristics. Based on the tumor development trend characteristics, the tumor development trend is simulated to determine the tumor development risk information; Analyze the tumor development risk information to determine and output dynamic follow-up recommendations.

[0006] This solution utilizes a probabilistic boundary analysis mechanism based on multi-temporal tumor image datasets to probabilistically represent tumor boundaries within patients, constructing spatial probabilistic boundary representation information. This avoids misjudgments of tumor boundaries caused by the binary decision-making tendency of image recognition algorithms that force closed contours. Furthermore, it performs multi-dimensional feature fusion analysis on boundary features under different dimensions within the spatial probabilistic boundary representation information, obtaining tumor development trend characteristics that reflect changes in tumor status. This transforms boundary uncertainty information into quantifiable development trend indicators. Based on these characteristics, it simulates tumor development trends to clarify the risk characteristics that the current tumor development trend poses to patient disease progression, generating tumor development risk information and providing corresponding dynamic follow-up recommendations to relevant personnel. This effectively eliminates boundary illusions generated by existing technologies in analyzing tumor boundaries and improves the accuracy of tumor development trend assessment.

[0007] Optionally, acquiring the multi-temporal tumor image dataset includes: Temporal registration and noise suppression are performed on the original tumor image set to integrate and determine the multi-temporal tumor image dataset. During the time-series registration process, a non-rigid transformation algorithm is used to perform time-relative alignment processing on the original tumor images that correspond at different times within the original tumor image set. During the noise suppression process, an adaptive filtering algorithm based on the local texture features of the original tumor image is used to filter noise from the original tumor image.

[0008] This approach utilizes a non-rigid transformation algorithm and an adaptive filtering algorithm to perform temporal alignment processing and noise suppression processing on the original tumor image dataset, based on the original tumor image dataset. This results in a corresponding multi-temporal tumor image dataset, reducing the negative impact of system noise and motion artifacts on tumor status analysis and improving the accuracy of subsequent tumor status analysis results.

[0009] Optionally, the analysis of the multi-temporal tumor image dataset to determine spatial probabilistic boundary representation information includes: According to the probabilistic boundary detection strategy, probabilistic boundary detection processing is performed on each phase tumor image in the multi-temporal tumor image dataset to generate an initial boundary probability map. Based on the preset high probability interval, the initial boundary probability map is analyzed, and several high probability boundary segmentation information corresponding to each phase of tumor image is extracted. Based on the timestamp information corresponding to tumor images of different time phases, several high-probability boundary segment information corresponding to tumor images of different time phases are dynamically superimposed to determine the high-probability boundary segment outline. Based on the multi-temporal tumor image dataset and the high-probability boundary segmentation contour, the direction of the high-probability tumor boundary is analyzed, and the boundary invasion probability mapped by the corresponding discontinuous contour boundary under the high-probability boundary segmentation contour is determined. Based on the initial boundary probability map and the boundary infiltration and invasion probability, the spatial probabilistic boundary representation information is constructed.

[0010] This scheme employs a probabilistic boundary detection strategy to perform probabilistic boundary detection on tumor impacts, resulting in an initial boundary probability map. By analyzing the high-probability boundary segment information in the initial boundary probability map, high-probability boundary segment contours are extracted. Based on these contours, boundary invasion probability analysis of discontinuous contour boundaries is performed. This constructs spatial probabilistic boundary representation information that includes the initial boundary probability map and boundary invasion probability. This spatial probabilistic boundary representation information can comprehensively reflect the boundary state of tumor tissue from both explicit and implicit perspectives, improving the accuracy of subsequent tumor development status assessment.

[0011] Optionally, the probabilistic boundary detection strategy includes: By using a parallel multi-branch convolutional architecture model, the local texture gradient, regional contrast enhancement features, and contextual semantic features of each image in the multi-temporal tumor image dataset under each model branch are extracted simultaneously, and the feature response map corresponding to each image under each model branch is integrated to determine the feature response map. By using a bidirectional gating mechanism, the feature response map of each image is analyzed to suppress isolated high-response noise regions caused by imaging artifacts and tissue adhesion in each image, thereby enhancing the cross-scale feature consistency and spatial continuity of the true tumor boundary region in each image under each model branch. Based on the cross-scale feature consistency and the spatial continuity, the feature response maps corresponding to different images under different model branches are subjected to pixel-level adaptive weighted fusion processing to generate the initial boundary probability map containing several pixel boundary segment probability values. The pixel boundary segmentation probability value is used to characterize the confidence that the corresponding pixel belongs to the true boundary, rather than the geometric connectivity of the forced closed contour.

[0012] This approach utilizes a parallel multi-branch convolutional architecture model to extract complementary features and combines a bidirectional gating mechanism to purify the feature response map. This effectively identifies and strongly suppresses isolated high-response noise regions caused by imaging artifacts and tissue adhesions, significantly reducing the generation of false positive boundary points, improving the robustness of tumor boundary detection, and thus ensuring the accuracy of tumor boundary detection results.

[0013] Optionally, the step of analyzing the direction of the high-probability tumor boundary based on the multi-temporal tumor image dataset and the high-probability boundary segmentation contour, and determining the boundary invasion probability mapped by the corresponding discontinuous contour boundary under the high-probability boundary segmentation contour, includes: Based on the high-probability boundary segmentation contour, a spatiotemporal topology graph with the high-probability boundary segmentation contour as boundary nodes is constructed. Based on the spatial displacement vector and morphological variability between the boundary nodes of adjacent time phases in the spatiotemporal topology diagram, the expansion direction of the current high-probability tumor boundary in three-dimensional space is analyzed, and the direction of the high-probability tumor boundary and the discontinuous contour boundary corresponding to each segment of the high-probability boundary are determined. Based on the multi-temporal tumor image dataset, the texture heterogeneity features of the discontinuous contour boundaries in the corresponding images are extracted according to the discontinuous contour boundaries. By fusing the texture heterogeneity features through a multilayer perceptual network, the probability of boundary blurring caused by the invasive growth of tumor cells in the image region corresponding to the discontinuous contour boundary is determined, and the boundary blurring probability is used as the boundary invasive probability. The boundary invasion probability directly reflects the hidden manifestation of tumor microscopic invasion behavior in macroscopic images.

[0014] This approach constructs a spatiotemporal topology map to reflect the dynamic evolution of tumor boundaries. Combined with local influence texture analysis, it captures hidden clues (hidden manifestations) reflecting microscopic invasive behavior from macroscopic images, outputting quantitative boundary infiltration and invasion probabilities. This avoids the lag caused by analyzing only the morphological changes of tumor boundaries that are visible to the naked eye, and enables the subsequent analysis results on tumor development status to provide more effective guidance for the generation of follow-up recommendations.

[0015] Optionally, the step of performing multi-dimensional feature fusion analysis on the tumor boundary based on the spatial probabilistic boundary representation information to determine the tumor development trend characteristics includes: Based on the boundary segment probability value and the boundary infiltration and invasion probability in the spatial probabilistic boundary representation information, a three-dimensional spatiotemporal feature matrix is ​​constructed. The three-dimensional spatiotemporal feature matrix includes the spatial coordinates of boundary points, local probability confidence, and adjacent temporal phase displacement vectors; Based on a dual-channel feature extraction strategy, the three-dimensional spatiotemporal feature matrix is ​​analyzed to determine boundary morphology feature information and invasion tendency feature information. Based on the boundary morphology features and the invasion tendency features, the tumor development trend features are constructed.

[0016] This scheme constructs a three-dimensional spatiotemporal feature matrix that comprehensively reflects the characteristics of tumor boundaries based on spatial probabilistic boundary representation information. On this basis, combined with a dual-channel feature extraction strategy, the morphology and dynamics of the tumor boundary are decoupled and analyzed separately to obtain boundary morphological feature information and invasion tendency feature information, thereby constructing tumor development trend feature information. This achieves morphological-dynamic decoupling analysis of the tumor boundary, avoids the tumor morphological details being submerged by motion noise, and improves the accuracy of tumor boundary analysis.

[0017] Optionally, the dual-channel feature extraction strategy includes a morphological feature analysis strategy and an invasion tendency analysis strategy; The morphological feature analysis strategy includes: By using a multi-scale morphological topology analysis algorithm, the spatial coordinates of all the boundary points in the three-dimensional spatiotemporal feature matrix and the local probability confidence are analyzed to extract several probabilistic abrupt boundary points. The probability mutation boundary point is the boundary point where the boundary segment probability value is less than that of the adjacent boundary point and the difference is greater than the preset boundary probability difference / the boundary point where the boundary infiltration and invasion probability is greater than the preset invasion probability. Each of the probabilistic mutation boundary points is taken as a potential micro-infiltrative foci, and the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltrative foci is extracted by a topology-aware convolutional neural network. The boundary morphological feature information is constructed based on the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltration lesion. The invasion tendency analysis strategy includes: A temporal graph neural network is constructed using the spatial coordinates of the boundary points as nodes and the adjacent temporal displacement vectors as edge weights. By analyzing the temporal graph neural network using a graph propagation algorithm, the motion consistency between adjacent nodes in the temporal graph neural network is aggregated to determine the aggregated motion information of the nodes; The node aggregation motion information is analyzed to determine the boundary expansion direction corresponding to different nodes, and the invasion tendency feature information is constructed based on the boundary expansion direction corresponding to different nodes.

[0018] This approach utilizes a dual-channel feature extraction strategy comprised of morphological feature analysis and invasion tendency analysis to improve the detection rate of microinvasive lesions while reducing the misjudgment rate of invasion direction. It achieves joint scientific quantitative analysis of the "morphological changes and dynamic trends" of tumor boundaries, breaking through the limitations of traditional single-dimensional analysis methods and providing a core feature engine for accurate tumor evolution prediction.

[0019] Optionally, the step of simulating the tumor development trend based on the tumor development trend characteristics information to determine tumor development risk information includes: Based on the boundary morphology features and invasion tendency features in the tumor development trend characteristics information, the mutation probability of tumor boundary morphology and the continuous change of invasion tendency are analyzed, and a probability-driven simulation engine for simulating the potential evolution of tumors in three-dimensional spatiotemporal space is constructed. The probability-driven simulation engine simulates the development of tumors within a preset observation period, and determines the direction of high-risk boundary expansion, the extent of microinvasive lesion spread, and the volume acceleration trend. The tumor development risk information is constructed based on the direction of expansion of the high-risk boundary, the extent of spread of the microinvasive lesions, and the volume acceleration trend.

[0020] This approach integrates the probability of boundary morphological mutations with the continuity of invasion to accurately identify high-risk regions with a combination of "high texture disorder + high directional consistency," thereby reducing the rate of missed diagnoses. Based on the analysis of the high-risk boundary expansion direction, the spread range of microinvasive lesions, and the volume acceleration trend, tumor development risk information is constructed to comprehensively and accurately reflect the tumor development trend.

[0021] Optionally, the analysis of the tumor development risk information to determine and output dynamic follow-up recommendations includes: Based on the direction of expansion of the high-risk boundary and the extent of spread of the microinvasive lesions, the subsequent high-risk intervention area of ​​the tumor is determined; Based on the high-risk areas of the tumor, several target organs for subsequent screening were identified. Based on the preset observation time period, the tumor spread rate is determined according to the volume acceleration trend; The subsequent follow-up time points will be determined based on the tumor spread rate. Based on the subsequent follow-up time points and several subsequent target screening organs, the dynamic follow-up recommendation information is constructed and output.

[0022] This solution utilizes a spatial locking mechanism for high-risk interference regions and an intelligent mapping mechanism for target screening organs to improve the detection rate of high-risk organs. Through an adaptive adjustment mechanism of follow-up time points based on tumor spread speed, it reduces ineffective re-examinations for patients with slow-progression tumors while providing timely emergency responses for patients with rapid-progression tumors. Based on subsequent follow-up time points and several subsequent target screening organs, dynamic follow-up recommendations are constructed and output using data visualization technology. This transforms abstract simulation data into actionable dynamic follow-up recommendations, improving their effectiveness.

[0023] Secondly, this application provides a tumor development analysis system based on image recognition, the system comprising: The boundary analysis module is used to acquire multi-temporal tumor image datasets, analyze the multi-temporal tumor image datasets, and determine spatial probabilistic boundary representation information. The feature analysis module is used to perform multi-dimensional feature fusion analysis on the tumor boundary based on the spatial probabilistic boundary representation information to determine the tumor development trend feature information; The development simulation module is used to simulate the tumor development trend based on the tumor development trend characteristics information and determine the tumor development risk information; The suggested output module is used to analyze the tumor development risk information and determine and output dynamic follow-up recommendations. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram illustrating an application scenario provided in one embodiment of this application; Figure 2 A flowchart illustrating a tumor development analysis method based on image recognition, provided as an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a tumor development analysis system based on image recognition, provided as an embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0027] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0028] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0029] Existing image recognition-based tumor development analysis techniques are prone to creating illusions about tumor boundaries due to the inherent tendency of image recognition technology to force closed contours during the analysis of tumor development trends. This can lead to misjudgments of tumor development trends and result in low reference value of the analysis results.

[0030] Based on this, this application provides a tumor development analysis method and system based on image recognition. Using a multi-temporal tumor image dataset, a probabilistic boundary analysis mechanism is employed to probabilistically represent the tumor boundary within the patient's body, constructing spatial probabilistic boundary representation information. This avoids the misjudgment of tumor boundaries caused by the binary decision-making tendency of image recognition algorithms that force closed contours. Furthermore, multi-dimensional feature fusion analysis is performed on the boundary features under different dimensions in the spatial probabilistic boundary representation information to obtain tumor development trend feature information reflecting changes in tumor status. This transforms boundary uncertainty information into quantifiable development trend indicators. Based on the tumor development trend feature information, the tumor development trend is simulated to clarify the risk characteristics brought by the current tumor development trend to the patient's disease development, generating tumor development risk information. Corresponding dynamic follow-up recommendations are provided to the relevant examination personnel, effectively eliminating the boundary illusion generated by existing technologies in analyzing tumor boundaries and improving the accuracy of tumor development trend assessment.

[0031] Figure 1 This is a schematic diagram illustrating an application scenario provided by this application. During tumor examination, the method provided in this application effectively eliminates the boundary illusion generated when analyzing tumor boundaries using existing techniques, thereby improving the accuracy of assessing tumor development.

[0032] Specifically, the method of this application is applied to any server that communicates with the hospital data management system. The server obtains a multi-temporal tumor image dataset provided by the hospital data management system. Based on the multi-temporal tumor image dataset, a probabilistic boundary analysis mechanism is used to probabilistically represent the tumor boundary within the patient's body, constructing spatial probabilistic boundary representation information. On this basis, multi-dimensional feature fusion analysis is performed on the boundary features under different dimensions in the spatial probabilistic boundary representation information to obtain tumor development trend feature information reflecting changes in tumor status. Based on the tumor development trend feature information, the tumor development trend is simulated to generate tumor development risk information, and corresponding dynamic follow-up recommendations are provided to the relevant examination personnel. This effectively eliminates the boundary illusion generated when analyzing tumor boundaries using existing technologies and improves the accuracy of tumor development trend assessment.

[0033] For specific implementation details, please refer to the following examples.

[0034] Figure 2This is a flowchart illustrating a tumor development analysis method based on image recognition, provided as an embodiment of this application. The method of this embodiment can be applied to the server in the above scenario. For example... Figure 2 As shown, the method includes: S201. Obtain multi-temporal tumor image dataset, analyze the multi-temporal tumor image dataset, and determine the spatial probabilistic boundary representation information.

[0035] A multi-temporal tumor image dataset can refer to an image dataset obtained by aligning and denoising a collection of multi-angle tumor images of the same patient collected at multiple time points through a hospital data management system.

[0036] Spatial probabilistic boundary representation information can be obtained by quantifying the fuzziness and uncertainty of tumor boundaries through a probabilistic model, with a probability value (0~1) representing the confidence information that each pixel belongs to the true boundary.

[0037] Specifically, existing image recognition technologies (such as deep learning-based semantic segmentation or traditional edge detection algorithms) have an inherent tendency to "force closed contours" when applied to tumor development assessment. This tendency stems from the algorithm's preference for continuous, closed boundaries (e.g., treating the tumor as a complete circle or smooth polygon). However, actual tumors in medical images (such as CT and MRI) often exhibit irregular, blurred, or partially missing boundaries, leading to "boundary illusion"—that is, the algorithm incorrectly generates non-existent closed contours (which can be considered a state of overfitting). This affects the accuracy of judging tumor size, shape, and growth trend. For example, in blurred boundary regions, the tendency to force closed contours may lead to the identification of closed boundaries, easily overlooking the invasive growth of the tumor (tumor cells gradually infiltrating normal tissue), resulting in misjudgments of tumor development. By analyzing multi-temporal tumor image datasets, a boundary probabilistic analysis mechanism is used to handle the uncertainty of tumor boundaries, avoiding rigid assumptions about tumor boundaries. This generates non-binary boundary probabilities, constructing spatial probabilistic boundary representation information, and avoiding misjudgments of tumor boundaries caused by the binary decision-making resulting from the algorithm's tendency to force closed contours.

[0038] S202. Based on the spatial probabilistic boundary representation information, perform multi-dimensional feature fusion analysis on the tumor boundary to determine the tumor development trend characteristics.

[0039] The tumor boundary can be the boundary between tumor tissue and adjacent non-tumor tissue.

[0040] Multidimensional feature fusion analysis can be a process of fusing and analyzing tumor boundary features from multiple dimensions (such as morphology and biology).

[0041] Information on the characteristics of tumor development can be used to characterize the overall development trend of a tumor (such as the direction and speed of expansion).

[0042] Specifically, the assessment of tumor development cannot be based solely on the relatively static feature of tumor boundary morphology. Tumor development requires a comprehensive consideration of morphological abnormalities (static features) and spatiotemporal consistency (dynamic features). For example, probabilistic mutations at boundary points accompanied by a high probability of invasion may indicate micrometastases; while the convergence of displacement vector directions can predict the macroscopic expansion direction. This step transforms the uncertainty information of the boundary into quantifiable development indicators through the fusion analysis of the dynamic and static features of the tumor boundary, providing core scientific input data for subsequent simulations of tumor development.

[0043] S203. Based on the characteristics of tumor development, simulate the tumor development trend to determine the risk information of tumor development.

[0044] Tumor development risk information can be a set of indicators that quantify the risk characteristics (high-risk expansion direction, extent of microinvasive lesion spread, and volume acceleration) caused by the tumor development trend.

[0045] Specifically, tumor development is highly uncertain (e.g., patients with similar tumor pathological features at the time may have significantly different subsequent progression). By using biodynamic simulation technology, the tumor development trend can be simulated based on the tumor development trend characteristics, and the development trend characteristics can be transformed into quantifiable risk characteristics. This clarifies the risk characteristics that the current tumor development trend brings to the patient's disease development, and provides a scientific basis for subsequent analysis and corresponding scientific follow-up recommendations.

[0046] S204. Analyze tumor development risk information, determine and output dynamic follow-up recommendations.

[0047] Dynamic follow-up recommendations can be individualized risk-based follow-up plans that include follow-up time and target organs for screening.

[0048] Specifically, after analyzing the tumor development risk information, in order to further improve the reference significance of the tumor development risk information for subsequent disease management procedures, based on further analysis of the tumor development risk information, the key time points and organ regions that need to be focused on during subsequent follow-up are deduced, dynamic follow-up recommendation information is constructed, and the corresponding dynamic follow-up recommendation information is provided to the relevant medical staff to provide a scientific reference plan for subsequent follow-up, avoid over-examination of low-risk patients, and prevent missed detection of high-risk patients.

[0049] This solution utilizes a probabilistic boundary analysis mechanism based on multi-temporal tumor image datasets to probabilistically represent tumor boundaries within patients, constructing spatial probabilistic boundary representation information. This avoids misjudgments of tumor boundaries caused by the binary decision-making tendency of image recognition algorithms that force closed contours. Furthermore, it performs multi-dimensional feature fusion analysis on boundary features under different dimensions within the spatial probabilistic boundary representation information, obtaining tumor development trend characteristics that reflect changes in tumor status. This transforms boundary uncertainty information into quantifiable development trend indicators. Based on these characteristics, it simulates tumor development trends to clarify the risk characteristics that the current tumor development trend poses to patient disease progression, generating tumor development risk information and providing corresponding dynamic follow-up recommendations to relevant personnel. This effectively eliminates boundary illusions generated by existing technologies in analyzing tumor boundaries and improves the accuracy of tumor development trend assessment.

[0050] In some embodiments, temporal registration and noise suppression processing are performed on the original tumor image set to integrate and determine a multi-temporal tumor image dataset; during the temporal registration process, a non-rigid transformation algorithm is used to perform temporal alignment processing on the original tumor images corresponding to different times within the original tumor image set; during the noise suppression process, an adaptive filtering algorithm based on the local texture features of the original tumor images is used to filter noise from the original tumor images.

[0051] The original tumor image set can be a collection of tumor image data corresponding to the current patient, provided by the hospital's data management system.

[0052] Temporal registration is the process of aligning tumor images acquired at different time points in space to eliminate image position deviations caused by factors such as changes in patient position and organ movement.

[0053] Noise suppression processing is a technique that reduces random noise (such as device noise and quantum noise) in images to improve the signal-to-noise ratio of the target area.

[0054] Non-rigid transformation algorithms can be image alignment algorithms that allow local deformation. By establishing an elastic deformation field to adapt to the physiological deformation of organs (such as breathing and peristalsis), computer vision algorithms such as the Demons algorithm and B-spline transformation algorithm can be used as non-rigid transformation algorithms in this scheme.

[0055] Local texture features can be the spatial distribution patterns of pixel gray values ​​within a local area of ​​an image (such as gradients and co-occurrence matrices), reflecting the heterogeneity of the tissue structure.

[0056] Adaptive filtering algorithms can be noise reduction methods that dynamically adjust filtering parameters based on local texture. The Wiener filtering algorithm can be used as the adaptive filtering algorithm in this scheme.

[0057] Specifically, because the patient's position can vary by centimeters in each scan, the same tumor may be spuriously displaced in the temporal images. Furthermore, due to the influence of system noise and motion artifacts in medical imaging equipment, high-response artifacts are easily formed in the tissue boundary area, which can be confused with the real invasive lesions (especially in low-contrast images). If the original tumor images that have not been aligned and denoised are used directly for tumor status analysis, it will lead to an amplified misjudgment rate of tumor status. Therefore, accurate judgment of the tumor development trend requires precise alignment and denoising of cross-temporal images.

[0058] A non-rigid transformation algorithm is used to extract feature points (such as SIFT keypoints) between the floating image and the reference image, establish an elastic deformation model, and minimize the feature point distance by optimizing the energy function (such as gradient descent) to generate a deformation field. The floating image is then mapped pixel by pixel to the coordinate system of the reference image, achieving time-relative alignment of the original tumor images corresponding to different times within the original tumor image set. The non-rigid transformation algorithm is used here mainly to adapt to the dynamic local deformation of tumor tissue (such as the deformation rate of liver tumors reaching 15% due to respiration), and to reduce the error in judging tumor tissue displacement caused by the rigid transformation algorithm.

[0059] The specific steps for feature point extraction include: constructing a Gaussian pyramid scale space and detecting local extrema as candidate feature points; removing low-contrast points using the Hessian matrix and retaining feature points with high stability; calculating the principal direction of the feature points (based on the peak value of the gradient histogram) and generating a 128-dimensional SIFT descriptor.

[0060] The specific steps for constructing the deformation field include: using a B-spline non-rigid transformation model to divide the image into a uniform control grid (grid spacing set to 15mm); optimizing the energy function using normalized mutual information (NMI) as a similarity measure, with the optimization objective being to minimize the feature point distance between the reference image and the floating image; iterative optimization: using the L-BFGS algorithm (finite memory quasi-Newton method) to adjust the control point displacement, with an initial step size of 0.1 and a maximum of 100 iterations.

[0061] The specific steps of the influence mapping include: performing bicubic interpolation on the floating image based on the optimized deformation field, and mapping it pixel by pixel to the reference image coordinate system.

[0062] Furthermore, the image data after relative alignment is processed into blocks (e.g., 16×16 pixel windows). Texture indicators (contrast and entropy values ​​in the gray-level co-occurrence matrix) that affect each data block are extracted. For highly homogeneous and low-heterogeneous regions (low contrast / low entropy), strong filtering parameters (large convolution kernels) are used, while for low-homogeneous and highly heterogeneous regions (high contrast / high entropy), weak filtering parameters (small convolution kernels) are used. Based on the corresponding filtering parameters, an adaptive filtering algorithm, such as the Wiener filtering algorithm, is adopted to adjust the filtering intensity according to the local variance, and the corresponding denoised image is output, thereby achieving noise filtering of the original tumor image.

[0063] The contrast ratio in the gray-level co-occurrence matrix reflects the intensity of local gray-level changes, and its calculation formula is as follows: ; The entropy value in the gray-level co-occurrence matrix is ​​used to measure texture randomness, and the calculation formula is as follows: ; in Represents grayscale combination The probability of its occurrence, It represents the grayscale difference between adjacent pixels. The greater the difference, the higher the value (highlighting abrupt changes such as boundaries and noise).

[0064] Low heterogeneous regions (contrast < 30 and entropy < 2.0): Wiener filtering with large convolution kernel (9×9) is used to enhance the smoothing effect; High heterogeneous regions (contrast ≥ 30 or entropy ≥ 2.0): Wiener filtering with small convolution kernels (3×3) is used to preserve edge details; The filtering strength is adaptively adjusted based on the local variance: the filtering is enhanced when the variance is greater than 1000, and weakened when the variance is less than 500.

[0065] This approach utilizes a non-rigid transformation algorithm and an adaptive filtering algorithm to perform temporal alignment processing and noise suppression processing on the original tumor image dataset, based on the original tumor image dataset. This results in a corresponding multi-temporal tumor image dataset, reducing the negative impact of system noise and motion artifacts on tumor status analysis and improving the accuracy of subsequent tumor status analysis results.

[0066] In some embodiments, according to a probabilistic boundary detection strategy, probabilistic boundary detection processing is performed on each phase of the tumor image in the multi-temporal tumor image dataset to generate an initial boundary probability map; based on a preset high-probability interval, the initial boundary probability map is analyzed to extract several high-probability boundary segment information corresponding to each phase of the tumor image; based on the timestamp information corresponding to different phases of the tumor image, several high-probability boundary segment information corresponding to different phases of the tumor image are dynamically superimposed to determine the high-probability boundary segment contours; based on the multi-temporal tumor image dataset and the high-probability boundary segment contours, the direction of the high-probability tumor boundary is analyzed to determine the boundary invasion probability mapped by the non-continuous contour boundary corresponding to the high-probability boundary segment contours; based on the initial boundary probability map and the boundary invasion probability, spatial probabilistic boundary representation information is constructed.

[0067] Probabilistic boundary detection processing can be a process that characterizes tumor boundaries in the form of probability values.

[0068] The initial boundary probability map can be a two-dimensional matrix used to characterize the probability values ​​of the tumor boundary, where each probability value represents the confidence level of the probability that the corresponding position belongs to the true boundary of the tumor.

[0069] The preset high probability interval can be a preset range of probability values ​​used to measure whether the current boundary can be regarded as the actual boundary of the tumor.

[0070] High-probability boundary segmentation information can be local boundary segments extracted from the initial boundary probability map, and the probability values ​​corresponding to the boundaries are all within a preset high-probability interval.

[0071] The timestamp information can be the creation timestamp data corresponding to the tumor image.

[0072] Dynamic overlay can be a process of weighted fusion of high-probability boundaries from different time phases in three-dimensional space according to timestamp order.

[0073] High-probability boundary segmentation contours can be a three-dimensional (spatial + temporal) probabilistic representation of tumor boundaries formed by dynamic superposition, reflecting the evolution path of the boundaries over time.

[0074] The trajectory of a tumor's high-probability boundary can indicate the trajectory of its subsequent high-probability expansion.

[0075] Discontinuous contour boundaries can be segments in the boundary contour where the probability value drops sharply or there is a geometric break (such as burr-like or lobed edges), pointing to potential microinvasive foci.

[0076] The probability of boundary invasion can be used to quantify the likelihood (0~1) of tumor cell micro-infiltration in discontinuous boundary regions, and is directly related to pathological invasion behavior.

[0077] Specifically, tumor boundaries exhibit invasive growth at the microscopic level (i.e., tumor cells infiltrate normal tissue). Due to image resolution limitations, the actual infiltrated area appears as fragmented boundary segments in the image, and the infiltrated area is accompanied by changes in tissue density, manifesting as local texture disorder (such as spiculation in mammograms). Therefore, in the process of probabilistic analysis of tumor boundaries, the corresponding probabilistic boundary detection strategy is invoked to analyze multi-temporal tumor image datasets, outputting an initial boundary probability map (matrix size is the same as the image, and pixel values ​​are probability values) for each tumor image. For each initial boundary probability map, a preset high-probability interval (e.g., 0.7~0.99) is applied for threshold segmentation to obtain a binary mask. Continuous pixel regions in the mask are extracted through connected component analysis, and each connected component is recorded as a high-probability boundary segment. Based on the timestamp information, the high-probability boundary segment information of tumor images at different temporal phases is sorted by time, and the high-probability segments corresponding to different temporal phases are then sorted. Boundary segmentation information is mapped to the same spatial coordinate system. The pixel probability values ​​of overlapping positions in different high-probability boundary segments are accumulated to generate a comprehensive probability distribution map. The high-probability boundary segment contours are formed by contour fitting. Furthermore, discontinuous regions (such as fracture endpoints or isolated fragments) in the high-probability boundary segment contours are identified. With the discontinuous regions as the center, local texture heterogeneity features (contrast and entropy values ​​of the gray-level co-occurrence matrix) are extracted from the original image. The texture features are input into a pre-trained multilayer perceptron and output the boundary infiltration probability value of the region. The initial boundary probability map and the corresponding boundary infiltration probability are integrated to obtain spatial probabilistic boundary representation information.

[0078] This scheme employs a probabilistic boundary detection strategy to perform probabilistic boundary detection on tumor impacts, resulting in an initial boundary probability map. By analyzing the high-probability boundary segment information in the initial boundary probability map, high-probability boundary segment contours are extracted. Based on these contours, boundary invasion probability analysis of discontinuous contour boundaries is performed. This constructs spatial probabilistic boundary representation information that includes the initial boundary probability map and boundary invasion probability. This spatial probabilistic boundary representation information can comprehensively reflect the boundary state of tumor tissue from both explicit and implicit perspectives, improving the accuracy of subsequent tumor development status assessment.

[0079] In some embodiments, a parallel multi-branch convolutional architecture model is used to simultaneously extract local texture gradients, regional contrast enhancement features, and contextual semantic features of each image in a multi-temporal tumor image dataset under each model branch, and integrate them to determine the feature response map corresponding to each image under each model branch. Through a bidirectional gating mechanism, each feature response map is analyzed to suppress isolated high-response noise regions in each image caused by imaging artifacts and tissue adhesion, thereby enhancing the cross-scale feature consistency and spatial continuity of the true tumor boundary region in each image under each model branch. Based on the cross-scale feature consistency and spatial continuity, the feature response maps corresponding to different images under different model branches are subjected to pixel-level adaptive weighted fusion processing to generate an initial boundary probability map containing several pixel boundary segment probability values. The pixel boundary segment probability values ​​are used to characterize the confidence that the corresponding pixel belongs to the true boundary, rather than the geometric connectivity of the forced closed contour.

[0080] Parallel multi-branch convolutional architecture models can be multi-path deep learning models that improve feature extraction efficiency through parallel computation, with each branch independently extracting image features (local / regional / global) at different scales.

[0081] Local texture gradients can be the intensity and direction features of pixel grayscale changes within a local area of ​​an image, used to identify minute boundaries (such as the edges of tumor spicules).

[0082] Regional contrast enhancement features can be features that quantify the gray-level differences between adjacent regions of an image, thereby strengthening the boundary between tumors and normal tissues.

[0083] Contextual semantic features can represent the semantic association of pixels in the global image (such as the spatial relationship between the tumor region and surrounding organs).

[0084] A feature response map can refer to the output (figure) after processing by a convolutional neural network (network), which is usually a two-dimensional or three-dimensional matrix related to the spatial size of the input image.

[0085] A bidirectional gating mechanism can be a dynamic feature filtering mechanism that evaluates the reliability of feature responses through forward / backward paths, suppresses noise, and enhances effective features.

[0086] Imaging artifacts can be the shadows or effects caused by minute displacements of organs and tissues in tumor images.

[0087] Tissue adhesions can be seen as adhesions between different tissues in tumor imaging.

[0088] Isolated high-response noise regions can be areas in the tumor that have high response values ​​in the feature response map but are spatially isolated, lack continuity with the surrounding areas, and are false signal regions caused by non-tumor factors (such as imaging artifacts and blurred boundaries caused by tissue adhesion). Identifying and suppressing these regions is the key to improving the accuracy of boundary detection.

[0089] Cross-scale feature consistency can be the logical correlation in the spatial distribution of features at the same location extracted from different branches (such as the coordinated response of microinvasive foci in local and global features).

[0090] Spatial continuity can be the topological connectivity property of the real tumor boundary in geometric space (such as the chain-like adjacency relationship of boundary points).

[0091] Pixel-level adaptive weighted fusion processing can dynamically assign weights to the outputs of different branches based on feature confidence, thereby achieving pixel-level feature fusion.

[0092] Specifically, multi-temporal tumor images are input into a parallel multi-branch convolutional architecture model. This model contains multiple independent processing paths (model branches). Each branch contains a series of convolutional layers and non-linear activation functions. The first branch focuses on calculating and enhancing the local texture gradient of the image, capturing subtle intensity change patterns within the pixel neighborhood. The second branch focuses on calculating regional contrast enhancement features, strengthening the tissue boundary between the tumor and the background by comparing the average characteristics (such as grayscale and texture statistics) of adjacent larger regions. The third branch is dedicated to extracting contextual semantic features, using convolutional layers or attention modules with large receptive fields to understand the semantic role of the current pixel in the entire tumor region and even the surrounding anatomical structures. Each branch processes the input image independently and finally outputs a feature response map that reflects its specifically extracted features.

[0093] The first branch uses a 3-layer convolutional network (3×3 kernels, stride 1) to extract the gray-level gradient in the neighborhood of a pixel and is enhanced by the Sobel operator. The second branch uses dilated convolution (dilation rate=2) to expand the receptive field and calculates the average gray-level difference between adjacent 10×10 regions; The third branch uses an attention mechanism to weight global features, focusing on strengthening the region where the tumor is associated with surrounding organs; Each branch outputs a feature map of the same size as the input image, with pixel values ​​representing the feature response intensity at that location.

[0094] The feature response map generated for each branch is refined using a bidirectional gating mechanism. This mechanism consists of two parallel gating units: Spatial Gating Unit: This unit analyzes the response relationship between each location in the feature response map and its neighboring locations (such as top, bottom, left, and right). It calculates a "spatial continuity score" to evaluate whether the current point's response pattern is consistent and smooth with its neighborhood response. For isolated high-response noise regions (i.e., points with high response values ​​but low or inconsistent response values ​​around them, often caused by imaging artifacts or blurring due to tissue adhesion), this unit will significantly reduce their response values ​​(calculate the 8-neighborhood response variance pixel by pixel in the feature response map; if the variance > threshold (empirical value 0.25), it is determined to be an isolated noise point, and its response value is reduced to 0).

[0095] Scale-gated unit: This unit analyzes the response values ​​of the same spatial location on the feature response maps of different branches (or different levels / scales within the same branch). It calculates a "cross-scale consistency score" to evaluate whether the response of the point is consistent across multiple abstract levels or feature dimensions. For points that have a strong response at one scale but a weak response or conflicting patterns at other scales, this unit will also reduce their response value (calculate the standard deviation of the response values ​​of the same location in the three branches; if the standard deviation is >0.3, it is judged as a cross-scale inconsistency point, and the response value is reduced proportionally).

[0096] Meanwhile, for regions that are spatially continuous and respond consistently across scales (more likely to be true tumor boundaries), the bidirectional gating mechanism identifies and enhances the response signals of these regions, making them more significant. After bidirectional gating, the noise in the response maps of each branch is effectively suppressed, and the true boundary features are enhanced.

[0097] This approach utilizes a parallel multi-branch convolutional architecture model to extract complementary features and combines a bidirectional gating mechanism to purify the feature response map. This effectively identifies and strongly suppresses isolated high-response noise regions caused by imaging artifacts and tissue adhesions, significantly reducing the generation of false positive boundary points, improving the robustness of tumor boundary detection, and thus ensuring the accuracy of tumor boundary detection results.

[0098] In some embodiments, a spatiotemporal topology map is constructed based on high-probability boundary segment contours, with high-probability boundary segment contours as boundary nodes. Based on the spatial displacement vectors and morphological variability between adjacent temporal boundary nodes in the spatiotemporal topology map, the expansion directionality of the current high-probability tumor boundary in three-dimensional space is analyzed, and the direction of the high-probability tumor boundary and discontinuous contour boundaries corresponding to each high-probability boundary segment contour are determined. Based on a multi-temporal tumor image dataset, texture heterogeneity features of discontinuous contour boundaries in the corresponding images are extracted. By fusing texture heterogeneity features through a multilayer perceptron, the boundary blurring probability caused by invasive growth of tumor cells within the image region corresponding to the discontinuous contour boundary is determined, and the boundary blurring probability is used as the boundary invasion probability. The boundary invasion probability directly reflects the hidden manifestation of tumor microscopic invasion behavior in macroscopic images.

[0099] Spatiotemporal topology graphs can be graph structures used to characterize the dynamic evolution of tumor boundary points in time and space dimensions.

[0100] Boundary nodes can be graph nodes in a spatiotemporal topology graph that represent a high-probability boundary point in a certain temporal image. Each node contains spatial coordinates and its corresponding temporal information, which are directly derived from points in the high-probability boundary segment contour.

[0101] A spatial displacement vector can refer to a directed line segment (vector) in a spatiotemporal topology diagram that connects two boundary nodes of adjacent time phases (such as time phase t and time phase t+1). The magnitude and direction of this vector quantitatively describe the spatial movement (displacement) of the corresponding boundary point within the time interval, and its calculation is derived from the difference in coordinates of adjacent time phase nodes.

[0102] Morphological variability can be a numerical index that quantifies the degree of difference in shape (such as curvature, concavity and convexity) between local boundary segments represented by two boundary nodes connected by spatial displacement vectors in adjacent temporal phases. It is derived from the calculation of the geometric properties of the local contours around the nodes (such as the rate of change of curvature).

[0103] The direction of expansion can be the direction of the tumor's spread from its boundaries to other tissues.

[0104] The high-probability boundary orientation of a tumor can refer to the dominant expansion direction of the tumor boundary in three-dimensional space, determined by statistical analysis based on the set of spatial displacement vectors of boundary nodes in a spatiotemporal topology graph. This direction originates from the aggregation analysis of the set of spatial displacement vectors.

[0105] Discontinuous contour boundaries refer to regions within high-probability boundary segmentation where geometric connections are broken due to missing boundary points, a sudden drop in probability values, or morphological mutations. These regions are often potential locations for microinvasive growth of tumor cells. This concept originates from continuity analysis and mutation point detection of high-probability boundary segmentation contours.

[0106] Texture heterogeneity features can refer to a set of quantitative features extracted from medical images that reflect the non-uniformity of pixel grayscale, gradient, and pattern distribution in local image regions where non-continuous contour boundaries are located. These features originate from medical image texture analysis techniques (such as Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP)).

[0107] Invasive growth of tumor cells can be an aggressive expansion mode in which tumor cells break through the basement membrane or the boundary of the primary lesion, invade and destroy the surrounding normal tissues.

[0108] Boundary blur probability can be a numerical value (between 0 and 1) output by a multilayer perceptron, which quantifies the possibility that the interface between normal tissue and tumor tissue will be blurred due to the infiltrative growth of tumor cells in the image region corresponding to the discontinuous contour boundary. It comes from the calculation results of texture heterogeneity features by the multilayer perceptron.

[0109] Specifically, invasive behavior often occurs at the microscopic level of the tumor boundary. Traditional imaging methods mainly rely on doctors' visual observation of macroscopic changes in boundary morphology (such as spiculation and lobulation) to indirectly infer the risk of invasion. This method has significant limitations: it is highly subjective and lagging. Macroscopic morphological changes are usually the result of the accumulation of microscopic invasion. At this time, the tumor may have already entered the progression stage and missed the best intervention period. It cannot quantify occult invasion. Tumor cells often grow infiltratively along the interstitial space in the form of single cells or small clusters. This microscopic behavior is only manifested in macroscopic images as slight blurring, discontinuity or subtle changes in texture (heterogeneity) in local boundary areas, which are difficult to detect with the naked eye and difficult to quantify. Based on the high-probability boundary segment contours corresponding to tumor images at different time phases obtained from the analysis in the aforementioned embodiments, each high-probability boundary point on each time phase contour is defined as a boundary node. The node attributes include its three-dimensional spatial coordinates and the timestamp of the corresponding time phase. Based on the principle of spatial proximity (matching boundary points of adjacent time phases according to Euclidean distance, with distances <5mm considered as the same node), boundary nodes under adjacent time phases are connected to form connecting edges (edge ​​weight = displacement vector magnitude × direction consistency (cosine similarity > 0.8 is retained)). Based on all boundary nodes and corresponding connecting edges (edge ​​weights are marked on the connecting edges), a spatiotemporal topology graph is constructed. On this basis, the spatial displacement vector of each connecting edge is calculated (vector = target node coordinates - ...). The source node coordinates are used to quantify the morphological variability of each connecting edge (characterized by local curvature changes; a 5×5 pixel region centered on the node is taken, and the rate of curvature change is calculated: a curvature difference > 0.05 radians / mm is marked as a morphological mutation). Using clustering algorithms, such as principal component analysis, the spatial displacement vector set of all connecting edges in the spatiotemporal topology graph is analyzed (partial or overall analysis is possible). This determines the high-probability tumor boundary orientation in three-dimensional space. Based on this, the morphological variability corresponding to the high-probability tumor boundary orientation is further analyzed (high grayscale uniformity (grayscale standard deviation > 4)). Boundary regions with 0) or gradient direction disorder (direction variance > 60°) are marked as discontinuous contour boundaries. Based on multi-temporal tumor image datasets, image data corresponding to the regions where discontinuous contour boundaries are located are extracted. The gray-level distribution data within the corresponding image data is used as texture heterogeneity features. On this basis, the gray-level standard deviation algorithm is used to analyze the gray-level distribution non-uniformity and quantify the boundary blur probability as the boundary invasion probability (gray-level distribution non-uniformity is positively correlated with the boundary blur probability; the higher the gray-level distribution non-uniformity, the more disordered the corresponding boundary is, and the greater the invasion probability).

[0110] Calculation process of boundary infiltration attack probability: Texture heterogeneity feature extraction: Extract 20×20 pixel ROI regions around discontinuous boundary points, calculate the contrast, correlation, and entropy of the gray-level co-occurrence matrix (GLCM), and extract the kurtosis coefficients of the local binary mode (LBP) histogram.

[0111] Multilayer perceptron structure: Input layer: 8-dimensional texture feature vector; Hidden layers: 3 fully connected layers (16→8→4 neurons), ReLU activated; Output layer: The Sigmoid function outputs probability values; Training data: Using imaging blocks of infiltrative lesions confirmed by existing pathological data; Positive samples were marked as areas with indistinct borders and evidence of pathological infiltration.

[0112] This approach constructs a spatiotemporal topology map to reflect the dynamic evolution of tumor boundaries. Combined with local influence texture analysis, it captures hidden clues (hidden manifestations) reflecting microscopic invasive behavior from macroscopic images, outputting quantitative boundary infiltration and invasion probabilities. This avoids the lag caused by analyzing only the morphological changes of tumor boundaries that are visible to the naked eye, and enables the subsequent analysis results on tumor development status to provide more effective guidance for the generation of follow-up recommendations.

[0113] In some embodiments, a three-dimensional spatiotemporal feature matrix is ​​constructed based on the boundary segment probability value and boundary invasion probability in the spatial probabilistic boundary representation information. The three-dimensional spatiotemporal feature matrix includes the spatial coordinates of boundary points, local probability confidence, and adjacent temporal displacement vectors. Based on a dual-channel feature extraction strategy, the three-dimensional spatiotemporal feature matrix is ​​analyzed to determine the boundary morphology feature information and invasion tendency feature information. Based on the boundary morphology feature information and invasion tendency feature information, tumor development trend feature information is constructed.

[0114] A three-dimensional spatiotemporal feature matrix can be a three-dimensional data structure that integrates spatial location, time dimension and probabilistic attributes, and is used to uniformly characterize the core feature parameters of tumor boundaries in spatiotemporal evolution.

[0115] The spatial coordinates of the boundary point can be the position information (x, y, z) of the tumor boundary point in the three-dimensional image coordinate system.

[0116] The local probability confidence level can be the probability value (0~1) that a single boundary point belongs to the true tumor boundary, derived from the initial boundary probability map of the aforementioned embodiment.

[0117] The displacement vectors of adjacent time phases can be the spatial movement vectors (Δx, Δy, Δz) of the same boundary point in continuous time phase images.

[0118] The dual-channel feature extraction strategy can serve as a framework for performing morphological and dynamic analysis in parallel.

[0119] Boundary morphology features can be a set of parameters describing the geometric heterogeneity of tumor boundaries and reflecting the intensity of local invasion.

[0120] Invasive tendency characteristics can be a set of parameters that quantify the expansion trend of boundary points and predict the tumor evolution path.

[0121] Specifically, existing tumor boundary analysis techniques suffer from a lack of decoupling between morphological and dynamic features: tumor boundaries contain both geometric morphological information (such as contour breaks and concavity / convexity) and implicit temporal motion information (such as directional expansion). If these are directly mixed for analysis (such as processing with a single convolutional network), morphological details will be obscured by motion noise. For example, the local texture disorder (morphological features) of microinvasive lesions and the overall boundary displacement (invasiveness features) have different physical meanings, and mixed processing will reduce the interpretability of both. First, the three-dimensional spatial coordinates of all boundary points are extracted, and a local probability confidence score (simultaneously introducing the boundary segment probability value and the corresponding boundary invasion probability) and adjacent temporal displacement vectors (the coordinate difference of the same boundary point in adjacent temporal images) are bound to each boundary point to construct a three-dimensional spatiotemporal feature matrix. Based on this, a dual-channel feature extraction strategy is used to perform morphological and dynamic decoupling analysis on the tumor boundary according to the three-dimensional spatiotemporal feature matrix, respectively, to obtain boundary morphological feature information and invasiveness tendency feature information, in order to construct tumor development trend feature information.

[0122] Implementation details of morphological feature analysis strategy: Probabilistic mutation point detection: Slide a 5mm window along the boundary line, and mark the probability mutation point when the average probability within the window decreases by more than 0.25 or the infiltration probability increases by more than 0.65.

[0123] Topology-aware convolutional networks: Input: A 31×31 pixel region centered on the mutation point; Use circular convolution kernels to detect radial texture patterns; Output: Variance values ​​of the texture gradient direction histogram (0-180°).

[0124] Details of the implementation of the invasion tendency analysis strategy: Graph Neural Network Construction: Node features: [coordinates, probability value, time phase]; Edge weight: cosine similarity of displacement vectors.

[0125] Graph propagation algorithm: Initially: Each node stores its own displacement vector; Iteration 1: Collect 1-hop neighborhood vectors and calculate the weighted average; Iteration 2: Filter outlier vectors that deviate from the mean by more than 25°; Output: Principal direction vector + aggregate confidence (0-1).

[0126] This scheme constructs a three-dimensional spatiotemporal feature matrix that comprehensively reflects the characteristics of tumor boundaries based on spatial probabilistic boundary representation information. On this basis, combined with a dual-channel feature extraction strategy, the morphology and dynamics of the tumor boundary are decoupled and analyzed separately to obtain boundary morphological feature information and invasion tendency feature information, thereby constructing tumor development trend feature information. This achieves morphological-dynamic decoupling analysis of the tumor boundary, avoids the tumor morphological details being submerged by motion noise, and improves the accuracy of tumor boundary analysis.

[0127] In some embodiments, the dual-channel feature extraction strategy includes a morphological feature analysis strategy and an invasion tendency analysis strategy. The morphological feature analysis strategy includes: analyzing the spatial coordinates and local probability confidence of all boundary points in the three-dimensional spatiotemporal feature matrix using a multi-scale morphological topology analysis algorithm, and extracting several probability mutation boundary points; the probability mutation boundary points are boundary points whose boundary segment probability value is less than that of adjacent boundary points and the difference is greater than a preset boundary probability difference / boundary points whose boundary infiltration invasion probability is greater than a preset invasion probability; each probability mutation boundary point is used as a potential micro-infiltration foci, and each potential micro-infiltration foci are extracted using a topology-aware convolutional neural network. The disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltrating lesion is determined. Boundary morphology feature information is constructed based on the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltrating lesion. An invasion tendency analysis strategy is employed, including: constructing a temporal graph neural network using the spatial coordinates of boundary points as nodes and adjacent temporal displacement vectors as edge weights; analyzing the temporal graph neural network using a graph propagation algorithm, aggregating the motion consistency between adjacent nodes in the temporal graph neural network, and determining the node aggregation motion information; analyzing the node aggregation motion information, determining the boundary expansion direction corresponding to different nodes, and constructing invasion tendency feature information based on the boundary expansion direction corresponding to different nodes.

[0128] Morphological feature analysis strategies can include analyzing tumor infiltration characteristics from the perspective of boundary morphology.

[0129] Invasive tendency analysis strategies can be used to analyze tumor invasive characteristics from the perspective of boundary dynamics.

[0130] Multiscale morphological topology analysis algorithms can be algorithms that integrate multiscale spatial analysis and topological principles to identify local morphological mutation features in tumor boundaries caused by microscopic invasion.

[0131] A probabilistic mutation boundary point can be a boundary node in the boundary that leads to an increase in the degree of boundary disorder.

[0132] The preset boundary probability difference can be a probability difference threshold used to quantify whether the boundary probability of a boundary point has undergone a significant change.

[0133] The preset invasion probability can be the probability value used to determine whether the invasion probability of a boundary point is abnormal.

[0134] Potential microinvasive foci can be the region where probabilistic mutation boundary points are located, used to characterize the starting point where tumor cells may infiltrate surrounding tissues.

[0135] Topology-aware convolutional neural networks can be convolutional neural networks designed and trained for the topology of medical images, used to quantify the degree of tissue disorder by sensing the spatial correlation of local texture gradient directions.

[0136] The degree of disorder in the direction of the surrounding texture gradient can be used as an indicator to quantify the degree of disorder in the texture direction of the image around the microinvasive lesion. The higher the value, the more disordered the arrangement of tissue cells.

[0137] Temporal graph neural networks can be dynamic graph structures with boundary points as nodes and adjacent temporal displacement vectors as edge weights, used to model the motion trajectory of boundaries over time.

[0138] Graph propagation algorithms can be algorithms that aggregate the motion information of adjacent nodes through iterative message passing, and are used to extract the overall trend of boundary expansion.

[0139] Motion consistency can be the consistency pattern of adjacent boundary points in the direction of motion.

[0140] The node aggregation motion information can be generated by fusing the displacement vectors of adjacent nodes through a graph propagation algorithm, which represents the overall motion trend of the boundary region.

[0141] The boundary expansion direction can be the tumor invasion dominant direction vector obtained by clustering node aggregation motion information.

[0142] Specifically, the morphological feature analysis strategy focuses on calculating the disorder of local texture gradient directions, isolating displacement interference, and assessing the invasive characteristics of tumors from a morphological perspective. A multi-scale morphological topology analysis algorithm (scale division can be 1mm / 3mm / 5mm) is used to scan all boundary points. Boundary points with a boundary segment probability value less than that of adjacent boundary points and a difference greater than a preset boundary probability difference, or boundary points with a boundary invasion probability greater than a preset invasion probability, are designated as probability mutation boundary points. A 5×5×3 (width×height×temporal) image region is extracted centered on each probability mutation point. A topology-aware convolutional neural network extracts the texture gradient direction (0°~360°) of all pixels within the region, and the gradient direction variance is calculated as the texture gradient direction disorder (e.g., variance > 45° = high risk). Boundary morphological feature information is constructed based on the disorder of each texture gradient direction.

[0143] The invasion tendency analysis strategy specializes in motion consistency aggregation, avoiding morphological noise contamination of direction prediction. It maps each boundary point in the 3D spatiotemporal feature matrix to a node, with node features defined as [spatial coordinates, local probability confidence]. If two points belong to the same boundary in a continuous temporal phase, the edge weight is calculated as the displacement vector similarity (cosine value > 0.8 for edge construction). Based on a temporal graph neural network, a graph propagation algorithm is used for three iterations: First, nodes receive neighborhood displacement vectors and calculate the weighted average direction; second, abnormal nodes with direction deviations > 20° are filtered out; third, the aggregated motion information of the nodes (main direction vector + confidence) is output. K-means directional clustering (K=3, divided according to blood vessel / organ boundaries / free regions) is performed on the aggregated motion information, outputting the dominant expansion direction of each cluster. Based on the boundary expansion direction corresponding to different nodes, invasion tendency feature information is constructed.

[0144] This approach utilizes a dual-channel feature extraction strategy comprised of morphological feature analysis and invasion tendency analysis to improve the detection rate of microinvasive lesions while reducing the misjudgment rate of invasion direction. It achieves joint scientific quantitative analysis of the "morphological changes and dynamic trends" of tumor boundaries, breaking through the limitations of traditional single-dimensional analysis methods and providing a core feature engine for accurate tumor evolution prediction.

[0145] In some embodiments, based on the boundary morphology characteristics and invasive tendency characteristics in the tumor development trend information, the mutation probability of the tumor boundary morphology and the continuous changes in invasive tendency are analyzed to construct a probability-driven simulation engine for simulating the potential evolution of the tumor in three-dimensional spatiotemporal space. Through the probability-driven simulation engine, the development state of the tumor is simulated within a preset observation period to determine the direction of high-risk boundary expansion, the spread range of microinvasive lesions, and the volume acceleration trend. Based on the direction of high-risk boundary expansion, the spread range of microinvasive lesions, and the volume acceleration trend, tumor development risk information is constructed.

[0146] A probability-driven simulation engine can be a dynamic simulation engine that simulates the evolutionary path of tumors in three-dimensional spacetime by quantifying the probability of boundary morphological mutations and the continuous changes in invasive tendency.

[0147] The implementation details of the probability-driven simulation engine include: Input parameters: Boundary morphological characteristics: location of microinfiltrates and degree of texture disorder (>45° is high risk).

[0148] Invasion tendency characteristics: Clustering results of the boundary expansion direction (confidence of the main direction >80%).

[0149] Evolutionary rules: Micro-infiltrative foci generation: Within a unit time step (7 days), new foci are randomly generated according to the proportion of texture disorder (for every 10% increase in disorder, the generation probability increases by 20%).

[0150] Diffusion direction calibration: If the motion direction offset is >20°, reset the vector according to the main direction of the neighboring node.

[0151] Volume acceleration calculation: Base volume growth: linear model (historical volume × 0.05 / week).

[0152] Contribution of micro-wetting foci: Volume increase per foci = diffusion range diameter³ × 0.1.

[0153] The preset observation period can be defined as the time span that clinicians pre-set based on the individual patient's condition to simulate the future development of the tumor. It serves as the time boundary of the simulation to constrain the range of tumor evolution prediction.

[0154] The direction of high-risk boundary expansion can be the spatial direction in which the tumor is most likely to invade rapidly. It is calculated by superimposing the boundary expansion direction in the invasive tendency feature information with the mutation probability.

[0155] The diffusion range of microinvasive lesions can be the maximum spatial area in which potential microinvasive lesions corresponding to probabilistic abrupt change boundary points may spread within a preset period.

[0156] The volume acceleration trend can be the rate of acceleration of tumor volume change over time (such as uniform / exponential growth), which is determined by the synergistic effect of boundary morphological stability and invasion direction.

[0157] Specifically, after obtaining boundary morphology and invasion tendency information, a simulation time step is set according to a preset observation period (e.g., 7 days per simulation step); texture disorder is converted into the probability value of boundary mutation per unit time; the higher the disorder, the greater the mutation probability; node aggregation motion information is converted into the invasion direction maintenance probability; the higher the motion consistency, the stronger the direction continuity; based on the mutation probability of the current time step, new micro-infiltrates are randomly generated (the location is determined by the probability mutation boundary point), and the diffusion vector of the micro-infiltrates is adjusted according to the invasion direction maintenance probability (recalibration is triggered when the direction shifts). Based on this, the changes in tumor volume are derived (basal volume growth is extrapolated from historical acceleration trends, plus additional volume increments caused by the spread of microinvasive lesions). The invasion direction with the highest frequency in the simulation (frequency > 80% is considered high risk) is used as the high-risk boundary expansion direction. The largest envelope region of microinvasive lesions in all simulated paths (such as a continuous region with a diameter ≥ 5 mm) is taken as the microinvasive lesion spread range. The second derivative of the volume-time curve is fitted, and the volume acceleration trend (linearly stable or exponentially accelerating) is output. Based on the high-risk boundary expansion direction, the microinvasive lesion spread range, and the volume acceleration trend, tumor development risk information is constructed.

[0158] This approach integrates the probability of boundary morphological mutations with the continuity of invasion to accurately identify high-risk regions with a combination of "high texture disorder + high directional consistency," thereby reducing the rate of missed diagnoses. Based on the analysis of the high-risk boundary expansion direction, the spread range of microinvasive lesions, and the volume acceleration trend, tumor development risk information is constructed to comprehensively and accurately reflect the tumor development trend.

[0159] In some embodiments, the high-risk intervention area of ​​the tumor is determined based on the expansion direction of the high-risk boundary and the spread range of the microinvasive lesions; several subsequent target screening organs are determined based on the subsequent high-risk intervention area of ​​the tumor; the tumor spread rate is determined based on the volume acceleration trend according to the preset observation time period; the subsequent follow-up time nodes are determined based on the tumor spread rate; and dynamic follow-up recommendation information is constructed and output based on the subsequent follow-up time nodes and several subsequent target screening organs.

[0160] The high-risk intervention area for subsequent tumor intervention can be the anatomical area that requires priority intervention, which is the overlapping area between the organ boundary pointed to by the expansion direction of the high-risk boundary and the spread of the microinvasive lesion.

[0161] The subsequent target screening organs can be a list of organs directly associated with the high-risk interference area.

[0162] The rate of tumor spread can be the spatial spread rate of microinvasive lesions per unit time (mm / month), which is derived from the correlation between volumetric acceleration and the spread range of microinvasive lesions.

[0163] The follow-up time point can be dynamically adjusted (e.g., 3 months → shortened to 1 month), triggered by the severity of the spread.

[0164] Specifically, traditional follow-up relies solely on volume changes (such as RECIST criteria) to set fixed re-examination cycles, ignoring spatial invasive characteristics. For example, if the volume is stable but the high-risk direction points to the brainstem, a routine 6-month re-examination may be performed, potentially missing the window of neurological invasion; if the volume increases rapidly but the invasive range is limited to the fat space, excessively shortening the follow-up interval may waste medical resources. This approach achieves joint quantification of "invasion orientation and invasion range" by fusing the expansion direction of the high-risk boundary (spatial vector) with the spread range of the microinvasive lesion (spatial region) into a subsequent high-risk intervention area for the tumor. A cone-shaped risk zone is generated by extending the spread range of the microinvasive lesion along the axis of the high-risk boundary expansion direction. This risk zone is superimposed on the patient's anatomical model, and areas with an overlap greater than 30% are marked as subsequent high-risk intervention areas for the tumor (e.g., "anterior margin of liver S8 segment - portal vein branch area"). An organ relationship database is queried to extract all organs within 10mm of the high-risk intervention area; secondary organs are filtered, retaining those with a high risk of metastasis (e.g., liver > mesenteric fat); a list of subsequent target screening organs is output (e.g., "liver, head of pancreas"). A velocity calculation model is selected based on the volume acceleration trend: linear stability → diffusion velocity = maximum diameter of microinvasive lesion / preset observation time period; exponential acceleration → diffusion velocity = (microinvasive lesion volume² × acceleration coefficient) / The time period is used to output the tumor spread rate (in mm / month); based on the tumor spread rate, the severity is graded: low-speed spread (≤3 mm / month) → follow-up node = standard period (e.g., 3 months); medium-speed spread (3-6 mm / month) → follow-up node = standard period × 0.7; high-speed spread (>6 mm / month) → follow-up node = standard period × 0.4; the subsequent follow-up time nodes are quantified; by integrating the subsequent follow-up time nodes and several subsequent target screening organs, dynamic follow-up recommendation information is constructed and output through data visualization technology.

[0165] The specific details of the mapping of high-risk interference regions are as follows: Using the high-risk expansion direction as the axis, a conical risk zone with a half angle of 30° is constructed and superimposed on the patient's CT three-dimensional reconstruction model. Overlapping areas are marked, and areas with an overlap rate >30% and containing critical organs are classified as high-risk interference zones.

[0166] Data visualization techniques: High-risk organs were displayed using a three-dimensional heat map, and key follow-up time points were marked using a time axis.

[0167] This solution utilizes a spatial locking mechanism for high-risk interference regions and an intelligent mapping mechanism for target screening organs to improve the detection rate of high-risk organs. Through an adaptive adjustment mechanism of follow-up time points based on tumor spread speed, it reduces ineffective re-examinations for patients with slow-progression tumors while providing timely emergency responses for patients with rapid-progression tumors. Based on subsequent follow-up time points and several subsequent target screening organs, dynamic follow-up recommendations are constructed and output using data visualization technology. This transforms abstract simulation data into actionable dynamic follow-up recommendations, improving their effectiveness.

[0168] Figure 3 A schematic diagram of the structure of an image recognition-based tumor development analysis system provided in one embodiment of this application is shown below. Figure 3 As shown, the tumor development analysis system 300 based on image recognition in this embodiment includes: a boundary analysis module 301, a feature analysis module 302, a development simulation module 303, and a suggestion output module 304.

[0169] Boundary analysis module 301 is used to acquire a multi-temporal tumor image dataset, analyze the multi-temporal tumor image dataset, and determine spatial probabilistic boundary representation information. Feature analysis module 301 is used to perform multi-dimensional feature fusion analysis on tumor boundaries based on the spatial probabilistic boundary representation information to determine tumor development trend feature information; The development simulation module 301 is used to simulate the tumor development trend based on the tumor development trend characteristics information and determine the tumor development risk information. The suggested output module 301 is used to analyze the tumor development risk information and determine and output dynamic follow-up recommendation information.

[0170] Optionally, when acquiring a multi-temporal tumor image dataset, the boundary analysis module 301 is specifically used for: Temporal registration and noise suppression are performed on the original tumor image set to integrate and determine the multi-temporal tumor image dataset. During the time-series registration process, a non-rigid transformation algorithm is used to perform time-relative alignment processing on the original tumor images that correspond at different times within the original tumor image set. During the noise suppression process, an adaptive filtering algorithm based on the local texture features of the original tumor image is used to filter noise from the original tumor image.

[0171] Optionally, the feature analysis module 302 is specifically used for: According to the probabilistic boundary detection strategy, probabilistic boundary detection processing is performed on each phase tumor image in the multi-temporal tumor image dataset to generate an initial boundary probability map. Based on the preset high probability interval, the initial boundary probability map is analyzed, and several high probability boundary segmentation information corresponding to each phase of tumor image is extracted. Based on the timestamp information corresponding to tumor images of different time phases, several high-probability boundary segment information corresponding to tumor images of different time phases are dynamically superimposed to determine the high-probability boundary segment outline. Based on the multi-temporal tumor image dataset and the high-probability boundary segmentation contour, the direction of the high-probability tumor boundary is analyzed, and the boundary invasion probability mapped by the corresponding discontinuous contour boundary under the high-probability boundary segmentation contour is determined. Based on the initial boundary probability map and the boundary infiltration and invasion probability, the spatial probabilistic boundary representation information is constructed.

[0172] Optionally, the probabilistic boundary detection strategy in the feature analysis module 302 is specifically used for: By using a parallel multi-branch convolutional architecture model, the local texture gradient, regional contrast enhancement features, and contextual semantic features of each image in the multi-temporal tumor image dataset under each model branch are extracted simultaneously, and the feature response map corresponding to each image under each model branch is integrated to determine the feature response map. By using a bidirectional gating mechanism, the feature response map of each image is analyzed to suppress isolated high-response noise regions caused by imaging artifacts and tissue adhesion in each image, thereby enhancing the cross-scale feature consistency and spatial continuity of the true tumor boundary region in each image under each model branch. Based on the cross-scale feature consistency and the spatial continuity, the feature response maps corresponding to different images under different model branches are subjected to pixel-level adaptive weighted fusion processing to generate the initial boundary probability map containing several pixel boundary segment probability values. The pixel boundary segmentation probability value is used to characterize the confidence that the corresponding pixel belongs to the true boundary, rather than the geometric connectivity of the forced closed contour.

[0173] Optionally, when the feature analysis module 302 analyzes the orientation of the high-probability tumor boundary based on the multi-temporal tumor image dataset and the high-probability boundary segmentation contour, and determines the boundary invasion probability mapped by the corresponding discontinuous contour boundary under the high-probability boundary segmentation contour, it is specifically used for: Based on the high-probability boundary segmentation contour, a spatiotemporal topology graph with the high-probability boundary segmentation contour as boundary nodes is constructed. Based on the spatial displacement vector and morphological variability between the boundary nodes of adjacent time phases in the spatiotemporal topology diagram, the expansion direction of the current high-probability tumor boundary in three-dimensional space is analyzed, and the direction of the high-probability tumor boundary and the discontinuous contour boundary corresponding to each segment of the high-probability boundary are determined. Based on the multi-temporal tumor image dataset, the texture heterogeneity features of the discontinuous contour boundaries in the corresponding images are extracted according to the discontinuous contour boundaries. By fusing the texture heterogeneity features through a multilayer perceptual network, the probability of boundary blurring caused by the invasive growth of tumor cells in the image region corresponding to the discontinuous contour boundary is determined, and the boundary blurring probability is used as the boundary invasive probability. The boundary invasion probability directly reflects the hidden manifestation of tumor microscopic invasion behavior in macroscopic images.

[0174] Optionally, when the feature analysis module 302 performs multi-dimensional feature fusion analysis on the tumor boundary based on the spatial probabilistic boundary representation information to determine the tumor development trend feature information, it is specifically used for: Based on the boundary segment probability value and the boundary infiltration and invasion probability in the spatial probabilistic boundary representation information, a three-dimensional spatiotemporal feature matrix is ​​constructed. The three-dimensional spatiotemporal feature matrix includes the spatial coordinates of boundary points, local probability confidence, and adjacent temporal phase displacement vectors; Based on a dual-channel feature extraction strategy, the three-dimensional spatiotemporal feature matrix is ​​analyzed to determine boundary morphology feature information and invasion tendency feature information. Based on the boundary morphology features and the invasion tendency features, the tumor development trend features are constructed.

[0175] Optionally, the dual-channel feature extraction strategy in the feature analysis module 302 includes a morphological feature analysis strategy and an invasion tendency analysis strategy; The morphological feature analysis strategy is specifically used for: By using a multi-scale morphological topology analysis algorithm, the spatial coordinates of all the boundary points in the three-dimensional spatiotemporal feature matrix and the local probability confidence are analyzed to extract several probabilistic abrupt boundary points. The probability mutation boundary point is the boundary point where the boundary segment probability value is less than that of the adjacent boundary point and the difference is greater than the preset boundary probability difference / the boundary point where the boundary infiltration and invasion probability is greater than the preset invasion probability. Each of the probabilistic mutation boundary points is taken as a potential micro-infiltrative foci, and the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltrative foci is extracted by a topology-aware convolutional neural network. The boundary morphological feature information is constructed based on the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltration lesion. The invasion tendency analysis strategy is specifically used for: A temporal graph neural network is constructed using the spatial coordinates of the boundary points as nodes and the adjacent temporal displacement vectors as edge weights. By analyzing the temporal graph neural network using a graph propagation algorithm, the motion consistency between adjacent nodes in the temporal graph neural network is aggregated to determine the aggregated motion information of the nodes; The node aggregation motion information is analyzed to determine the boundary expansion direction corresponding to different nodes, and the invasion tendency feature information is constructed based on the boundary expansion direction corresponding to different nodes.

[0176] Optionally, the development simulation module 303 is specifically used for: Based on the boundary morphology features and invasion tendency features in the tumor development trend characteristics information, the mutation probability of tumor boundary morphology and the continuous change of invasion tendency are analyzed, and a probability-driven simulation engine for simulating the potential evolution of tumors in three-dimensional spatiotemporal space is constructed. The probability-driven simulation engine simulates the development of tumors within a preset observation period, and determines the direction of high-risk boundary expansion, the extent of microinvasive lesion spread, and the volume acceleration trend. The tumor development risk information is constructed based on the direction of expansion of the high-risk boundary, the extent of spread of the microinvasive lesions, and the volume acceleration trend.

[0177] Optionally, the suggestion output module 304 is specifically used for: Based on the direction of expansion of the high-risk boundary and the extent of spread of the microinvasive lesions, the subsequent high-risk intervention area of ​​the tumor is determined; Based on the high-risk areas of the tumor, several target organs for subsequent screening were identified. Based on the preset observation time period, the tumor spread rate is determined according to the volume acceleration trend; The subsequent follow-up time points will be determined based on the tumor spread rate. Based on the subsequent follow-up time points and several subsequent target screening organs, the dynamic follow-up recommendation information is constructed and output.

[0178] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A tumor development analysis method based on image recognition, characterized in that, include: Acquire a multi-temporal tumor image dataset, analyze the multi-temporal tumor image dataset, and determine the spatial probabilistic boundary representation information; Based on the spatial probabilistic boundary representation information, a multi-dimensional feature fusion analysis is performed on the tumor boundary to determine the tumor development trend characteristics. Based on the tumor development trend characteristics, the tumor development trend is simulated to determine the tumor development risk information; Analyze the tumor development risk information to determine and output dynamic follow-up recommendations.

2. The method according to claim 1, characterized in that, The acquisition of the multi-temporal tumor image dataset includes: Temporal registration and noise suppression are performed on the original tumor image set to integrate and determine the multi-temporal tumor image dataset. During the time-series registration process, a non-rigid transformation algorithm is used to perform time-relative alignment processing on the original tumor images that correspond at different times within the original tumor image set. During the noise suppression process, an adaptive filtering algorithm based on the local texture features of the original tumor image is used to filter noise from the original tumor image.

3. The method according to claim 2, characterized in that, The analysis of the multi-temporal tumor image dataset to determine spatial probabilistic boundary representation information includes: According to the probabilistic boundary detection strategy, probabilistic boundary detection processing is performed on each phase tumor image in the multi-temporal tumor image dataset to generate an initial boundary probability map. Based on the preset high probability interval, the initial boundary probability map is analyzed, and several high probability boundary segmentation information corresponding to each phase of tumor image is extracted. Based on the timestamp information corresponding to tumor images of different time phases, several high-probability boundary segment information corresponding to tumor images of different time phases are dynamically superimposed to determine the high-probability boundary segment outline. Based on the multi-temporal tumor image dataset and the high-probability boundary segmentation contour, the direction of the high-probability tumor boundary is analyzed, and the boundary invasion probability mapped by the corresponding discontinuous contour boundary under the high-probability boundary segmentation contour is determined. Based on the initial boundary probability map and the boundary infiltration and invasion probability, the spatial probabilistic boundary representation information is constructed.

4. The method according to claim 3, characterized in that, The probabilistic boundary detection strategy includes: By using a parallel multi-branch convolutional architecture model, the local texture gradient, regional contrast enhancement features, and contextual semantic features of each image in the multi-temporal tumor image dataset under each model branch are extracted simultaneously, and the feature response map corresponding to each image under each model branch is integrated to determine the feature response map. By using a bidirectional gating mechanism, the feature response map of each image is analyzed to suppress isolated high-response noise regions caused by imaging artifacts and tissue adhesion in each image, thereby enhancing the cross-scale feature consistency and spatial continuity of the true tumor boundary region in each image under each model branch. Based on the cross-scale feature consistency and the spatial continuity, the feature response maps corresponding to different images under different model branches are subjected to pixel-level adaptive weighted fusion processing to generate the initial boundary probability map containing several pixel boundary segment probability values. The pixel boundary segmentation probability value is used to characterize the confidence that the corresponding pixel belongs to the true boundary, rather than the geometric connectivity of the forced closed contour.

5. The method according to claim 4, characterized in that, The step of analyzing the direction of high-probability tumor boundaries based on the multi-temporal tumor image dataset and the high-probability boundary segmentation contours, and determining the boundary invasion probability mapped by the corresponding discontinuous contour boundaries under the high-probability boundary segmentation contours, includes: Based on the high-probability boundary segmentation contour, a spatiotemporal topology graph with the high-probability boundary segmentation contour as boundary nodes is constructed. Based on the spatial displacement vector and morphological variability between the boundary nodes of adjacent time phases in the spatiotemporal topology diagram, the expansion direction of the current high-probability tumor boundary in three-dimensional space is analyzed, and the direction of the high-probability tumor boundary and the discontinuous contour boundary corresponding to each segment of the high-probability boundary are determined. Based on the multi-temporal tumor image dataset, the texture heterogeneity features of the discontinuous contour boundaries in the corresponding images are extracted according to the discontinuous contour boundaries. By fusing the texture heterogeneity features through a multilayer perceptual network, the probability of boundary blurring caused by the invasive growth of tumor cells in the image region corresponding to the discontinuous contour boundary is determined, and the boundary blurring probability is used as the boundary invasive probability. The boundary invasion probability directly reflects the hidden manifestation of tumor microscopic invasion behavior in macroscopic images.

6. The method according to claim 5, characterized in that, The step of performing multi-dimensional feature fusion analysis on the tumor boundary based on the spatial probabilistic boundary representation information to determine the tumor development trend characteristics includes: Based on the boundary segment probability value and the boundary infiltration and invasion probability in the spatial probabilistic boundary representation information, a three-dimensional spatiotemporal feature matrix is ​​constructed. The three-dimensional spatiotemporal feature matrix includes the spatial coordinates of boundary points, local probability confidence, and adjacent temporal phase displacement vectors; Based on a dual-channel feature extraction strategy, the three-dimensional spatiotemporal feature matrix is ​​analyzed to determine boundary morphology feature information and invasion tendency feature information. Based on the boundary morphology features and the invasion tendency features, the tumor development trend features are constructed.

7. The method according to claim 6, characterized in that, The dual-channel feature extraction strategy includes a morphological feature analysis strategy and an invasion tendency analysis strategy; The morphological feature analysis strategy includes: By using a multi-scale morphological topology analysis algorithm, the spatial coordinates of all the boundary points in the three-dimensional spatiotemporal feature matrix and the local probability confidence are analyzed to extract several probabilistic abrupt boundary points. The probability mutation boundary point is the boundary point where the boundary segment probability value is less than that of the adjacent boundary point and the difference is greater than the preset boundary probability difference / the boundary point where the boundary infiltration and invasion probability is greater than the preset invasion probability. Each of the probabilistic mutation boundary points is taken as a potential micro-infiltrative foci, and the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltrative foci is extracted by a topology-aware convolutional neural network. The boundary morphological feature information is constructed based on the disorder of the surrounding texture gradient direction corresponding to each potential micro-infiltration lesion. The invasion tendency analysis strategy includes: A temporal graph neural network is constructed using the spatial coordinates of the boundary points as nodes and the adjacent temporal displacement vectors as edge weights. By analyzing the temporal graph neural network using a graph propagation algorithm, the motion consistency between adjacent nodes in the temporal graph neural network is aggregated to determine the aggregated motion information of the nodes; The node aggregation motion information is analyzed to determine the boundary expansion direction corresponding to different nodes, and the invasion tendency feature information is constructed based on the boundary expansion direction corresponding to different nodes.

8. The method according to claim 7, characterized in that, The step of simulating the tumor development trend based on the tumor development trend characteristics information to determine the tumor development risk information includes: Based on the boundary morphology features and invasion tendency features in the tumor development trend characteristics information, the mutation probability of tumor boundary morphology and the continuous change of invasion tendency are analyzed, and a probability-driven simulation engine for simulating the potential evolution of tumors in three-dimensional spatiotemporal space is constructed. The probability-driven simulation engine simulates the development of tumors within a preset observation period, and determines the direction of high-risk boundary expansion, the extent of microinvasive lesion spread, and the volume acceleration trend. The tumor development risk information is constructed based on the direction of expansion of the high-risk boundary, the extent of spread of the microinvasive lesions, and the volume acceleration trend.

9. The method according to claim 8, characterized in that, The analysis of the tumor development risk information, and the determination and output of dynamic follow-up recommendations, include: Based on the direction of expansion of the high-risk boundary and the extent of spread of the microinvasive lesions, the subsequent high-risk intervention area of ​​the tumor is determined; Based on the high-risk areas of the tumor, several target organs for subsequent screening were identified. Based on the preset observation time period, the tumor spread rate is determined according to the volume acceleration trend; The subsequent follow-up time points will be determined based on the tumor spread rate. Based on the subsequent follow-up time points and several subsequent target screening organs, the dynamic follow-up recommendation information is constructed and output.

10. A tumor development analysis system based on image recognition, characterized in that, include: The boundary analysis module is used to acquire multi-temporal tumor image datasets, analyze the multi-temporal tumor image datasets, and determine spatial probabilistic boundary representation information. The feature analysis module is used to perform multi-dimensional feature fusion analysis on the tumor boundary based on the spatial probabilistic boundary representation information to determine the tumor development trend feature information; The development simulation module is used to simulate the tumor development trend based on the tumor development trend characteristics information and determine the tumor development risk information; The suggested output module is used to analyze the tumor development risk information and determine and output dynamic follow-up recommendations.

Images