A method and system for dynamically evaluating the risk of recurrence of glioma after surgery

Abstract

The application discloses a glioma postoperative recurrence risk dynamic evaluation method and system, which performs multi-modal time sequence fusion on multi-period image follow-up data and clinical detection data, identifies and distinguishes real recurrence and treatment response through image dynamic change extraction and multi-modal contradictory response, establishes a recurrence judgment rule and generates a false remission identifier; according to the recurrence judgment rule, perfusion diffusion combined analysis is performed, the recurrence characteristic index is established by capturing the continuous leading period of the peritumoral edema area in the functional area direction on the expansion rate of the active tumor area, and the infiltration advancing direction and the functional area space proximity degree are quantified; combined with the functional compensation degradation characteristics, the risk key node is jointly calibrated, the risk grade matching is completed after triggering the deterioration warning and the interval warning, and the dynamic evaluation result of the glioma postoperative recurrence risk is output, and the multi-period follow-up data is improved in the image and clinical double-dimensional recurrence risk comprehensive judgment ability.

Description

Technical Field

[0001] This invention relates to the field of medical image information processing technology, and in particular to a method and system for dynamic assessment of the risk of recurrence after glioma surgery. Background Technology

[0002] Postoperative glioma patients require long-term follow-up management. During follow-up, imaging examinations and functional status assessments are conducted simultaneously, and these two types of data naturally differ in terms of time span and collection frequency. Current follow-up assessment systems mainly rely on the independent interpretation of single imaging results and clinical scores, lacking a processing mechanism to systematically correlate dynamic changes in multi-cycle imaging with the temporal evolution of functional status. This makes it difficult to effectively distinguish between early recurrence signals and imaging changes caused by treatment response in the early stages of follow-up.

[0003] False remission is a significant confounding factor affecting the accuracy of recurrence assessment during postoperative follow-up of gliomas. Its core manifestation is the decoupling between decreased imaging enhancement and a simultaneous decline in functional status. Current methods lack a continuous tracking mechanism based on multi-phase follow-up data to identify this decoupling. Furthermore, the joint assessment of the threat posed to functional areas by the direction of the invasion front, the dynamic exhaustion process of functional compensation capacity, and spatial risk signals on imaging lacks a systematic quantitative processing method within the existing assessment framework, thus hindering further improvement in the dynamic early warning capability for postoperative recurrence risk. Summary of the Invention

[0004] This invention discloses a method and system for dynamic assessment of postoperative recurrence risk in glioma. The system aims to integrate multimodal time-series data from multi-cycle imaging follow-up and clinical testing, addressing quality differences across devices and time cycles. It establishes recurrence determination rules by combining dynamic changes in imaging with multimodal contradictory responses, distinguishing between true recurrence and imaging changes caused by treatment response. Furthermore, it establishes recurrence characteristic indices through combined perfusion-diffusion analysis and quantification of invasion direction risk, integrating functional compensation and degeneration characteristics for joint calibration. Ultimately, it triggers deterioration warnings and interval warnings, completes risk level matching, and outputs a dynamic recurrence risk assessment result that considers both imaging spatial risk and clinical functional degeneration.

[0005] The first aspect of this invention proposes a method for dynamic assessment of the risk of recurrence after glioma surgery, comprising the following steps: Collect image follow-up data and clinical test data, and perform multimodal temporal fusion on the image follow-up data and the clinical test data to generate a follow-up baseline set; The dynamic changes in images of the follow-up baseline set are extracted to determine the tumor progression area. Feature decomposition is performed in the tumor progression area to establish recurrence determination rules. Based on the contradictory response between the tumor progression area and the clinical test data, a false remission identifier is generated. According to the recurrence determination rules, perfusion-diffusion combined analysis is performed on the pseudo-remission markers to form active tumor regions and peritumoral edema regions. Based on the time period when the expansion rate of the peritumoral edema region leads that of the active tumor region, the risk of invasion is quantified to establish a recurrence characteristic index. Based on the recurrence characteristic index, key risk nodes are located within the active tumor area. Functional compensation risk markers are generated by extracting the rapid recovery segment after a sharp drop in KPS from the clinical test data. The key risk nodes and the functional compensation risk markers are jointly calibrated to generate risk stratification features and degradation feature indicators. Based on the risk stratification features and the relapse determination rules, a deterioration warning is triggered to generate a warning feature quantity. Based on the degradation feature indicators, a period of sudden drop in KPS during the follow-up interval is identified to generate an interval warning feature quantity. Based on the warning feature quantity and the interval warning feature quantity, risk level matching is performed to output a risk assessment result.

[0006] A second aspect of this invention provides a dynamic assessment system for the risk of recurrence after glioma surgery, comprising: The data integration module is used to collect image follow-up data and clinical test data, and to perform multimodal temporal fusion of the image follow-up data and the clinical test data to generate a follow-up baseline set; The feature extraction module is used to extract the dynamic changes in images of the follow-up baseline set to determine the tumor progression area, perform feature decomposition in the tumor progression area to establish recurrence determination rules, and generate false remission markers based on the contradictory response between the tumor progression area and the clinical test data. The infiltration analysis module is used to perform perfusion-diffusion joint analysis on the pseudo-remission marker according to the recurrence determination rules to form the active tumor area and the peritumoral edema area. Based on the time period when the expansion rate of the peritumoral edema area leads the active tumor area, the infiltration risk is quantified to establish the recurrence characteristic index. The risk labeling module is used to locate key risk nodes in the active area of ​​the tumor based on the recurrence characteristic index, extract the rapid recovery segment after the sharp drop in KPS from the clinical test data to generate a functional compensation risk label, and perform joint labeling on the key risk nodes and the functional compensation risk label to generate risk stratification features and degradation feature indicators. The assessment output module is used to trigger a deterioration warning based on the risk stratification features and the relapse determination rules, generate a warning feature quantity, identify the period of sudden drop in KPS during the follow-up interval based on the degradation feature indicators, generate an interval warning feature quantity, and perform risk level matching based on the warning feature quantity and the interval warning feature quantity to output a risk assessment result.

[0007] The beneficial effects of this invention are reflected in the following points: First, the multimodal temporal fusion of multi-cycle image follow-up data and clinical test data establishes a unified follow-up baseline across devices and cycles, enabling follow-up data from different sources and of different qualities to be systematically processed under the framework of fusion quality labeling; on this basis, through the extraction of dynamic changes in images and the identification mechanism of contradictory responses, the decoupling phenomenon of strengthening signal trends and temporal changes in functional states is quantified, forming false remission indicators and relapse judgment rules, realizing the effective distinction between real relapse and image changes caused by treatment response, and improving the reliability of follow-up data in the early stage of relapse judgment. Secondly, under the constraints of recurrence determination rules, a combined perfusion-diffusion analysis was performed on the time interval of pseudo-remission markers. This phasedly separated the spatial boundaries between the active tumor area and the peritumoral edema area. By capturing the period when the rate of expansion of the active tumor area by the peritumoral edema area along the functional zone's direction, the spatial relationship between the infiltration direction and the functional zone was incorporated into the risk quantification system. The established recurrence characteristic index comprehensively reflects the combined contribution of the three dimensions of infiltration range, spatial proximity, and duration to the threat of functional impairment, making the infiltration risk assessment both directionally sensitive and temporally cumulative. Finally, the joint calibration of key risk nodes and functional compensatory degradation characteristics established a systematic correspondence between imaging spatial risk and clinical functional degradation signals on the time axis. The generated risk stratification features and degradation characteristic indicators respectively carry risk evolution information in the spatial and functional dimensions. The deterioration early warning mechanism identifies high-risk mismatches between risk stratification features and recurrence determination rules, while the interval early warning mechanism continuously tracks periods of sudden drops in KPS during follow-up imaging. After risk level matching, the two types of early warning signals output dynamic risk assessment results, achieving collaborative judgment of imaging risk and clinical functional degradation signals at the early warning level. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a method for dynamically assessing the risk of recurrence after glioma surgery according to the present invention.

[0009] Figure 2 This is a structural block diagram of a dynamic assessment system for postoperative recurrence risk of glioma according to the present invention. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0012] The technical solutions of the embodiments of this application will be described below.

[0013] like Figure 1 As shown, this embodiment of the invention provides a method for dynamic assessment of the risk of recurrence after glioma surgery, including the following steps S11-S15: Step S11: Collect image follow-up data and clinical test data, and perform multimodal temporal fusion of image follow-up data and clinical test data to generate a follow-up baseline set.

[0014] Specifically, image follow-up data and clinical testing data are collected. Image follow-up data consists of MRI scans from regular follow-up examinations after glioma surgery. Due to differences in institutional conditions, the acquisition equipment model, scanning parameters, and slice thickness settings vary across different follow-up periods. During the acquisition phase, the scanning protocol version for each period needs to be recorded simultaneously to ensure the temporal comparability of image follow-up data across different equipment and periods. In some follow-up periods, incomplete sequence acquisition is common due to patient tolerance limitations or temporary referrals. It is not uncommon for only one of the enhanced T1-weighted imaging (T1-weighted imaging) or DWI sequences to be completed within the same acquisition period. In the image follow-up data, a sequence missing marker is used to replace the corresponding sequence in the results, avoiding the silent neglect of missing information during multimodal temporal fusion and preventing misjudgments of completeness. After standardized preprocessing, image follow-up data extracts image feature vectors for each period. These feature vectors cover three main dimensions: enhancement signal intensity, diffusion limitation, and perfusion peak. Nodes with significant equipment differences are simultaneously annotated with equipment difference labels, recording the equipment model and protocol differences. Clinical test data are assessed and entered by clinicians at each follow-up point. The KPS score is affected by the differences in individual physician experience, and there are subjective fluctuations in the scoring range of different physicians for the same functional status. The KPS records in the early functional recovery stage after surgery are more likely to have inconsistent scores between adjacent nodes due to the high assessment frequency. The data entry method for clinical test data is divided into two categories: structured form entry and paper transcription, depending on the level of informatization of the institution. Low confidence labels are added to items with subjective fluctuations in KPS assessment or incomplete fields in paper transcription. The low confidence label reflects the reliability of the functional status record of the item.

[0015] Multimodal temporal fusion of image follow-up data and clinical test data is performed to generate a follow-up baseline set. The image feature vectors of the image follow-up data and the KPS values ​​of the clinical test data are paired and fused according to the principle of temporal nearest neighbor. The normalized value of the pairing time interval, Q_align = 1 - |Δt| / Δt_max, is used as the temporal alignment quality component of that node, where Δt is the time interval between image acquisition and KPS assessment, and Δt_max is the set maximum acceptable interval. A higher Q_align indicates greater temporal synchronization between the two, and the degree of synchronization directly affects the fusion reliability of that node in the follow-up baseline set. The follow-up baseline set uses this quality component as the basis for adjusting the fusion weights of each node. Nodes in the image follow-up data with sequence missing markers or device difference markers undergo weight reduction on the corresponding image feature dimensions. The reduction magnitude is positively correlated with the type of missing sequence and the degree of device difference. The weight reduction is greatest for the enhancement feature dimension of nodes with missing T1 sequences, and secondary reduction for the diffusion feature dimension of nodes with missing DWI sequences. In clinical testing data, the KPS fusion weights of nodes corresponding to low-confidence labeled entries are synchronously reduced. Low-confidence KPS values ​​are used in fusion instead of point estimates, with the interval width reflecting the uncertainty of the node's functional status record. KPS fluctuations of clinical event-related labeled nodes are corrected for during fusion. In patients with postoperative pulmonary infection, KPS drops sharply during the infection period and then recovers rapidly. Without event correction, this fluctuation produces a waveform in the fusion result highly similar to the KPS drop-compensation stabilization pattern caused by tumor infiltration and compression of functional areas, leading to directional errors in KPS trend analysis. Each node in the follow-up baseline set carries an image feature vector, a fused KPS value, a temporal alignment quality component, and a comprehensive fusion quality label. The comprehensive fusion quality label summarizes the quality status of both the image follow-up data and the clinical testing data. Nodes with high quality on both sides are written with high-confidence labels, while nodes with missing or low-confidence conditions on either side are written with composite quality labels.

[0016] Step S12: Extract the dynamic changes of images from the follow-up baseline set to determine the tumor progression area, perform feature decomposition in the tumor progression area to establish recurrence determination rules, and generate false remission markers based on the contradictory response between the tumor progression area and clinical test data.

[0017] Specifically, the dynamic changes in images from the follow-up baseline set are extracted to determine the tumor progression area. The dynamic changes in images are extracted using the difference in image feature vectors between adjacent nodes in the follow-up baseline set. The difference calculation covers three main feature dimensions: enhancement signal intensity, diffusion restriction degree, and perfusion peak. The direction and magnitude of change in these three dimensions jointly determine the image state offset of the current node relative to the previous period. Nodes with low fusion quality in the follow-up baseline set need to be corrected for the corresponding feature dimensions in conjunction with quality annotations during difference calculation. If a systematic increase in signal intensity caused by equipment upgrades is directly included in the difference statistics without correction, it will produce a false offset in the dynamic changes of images that is highly similar to the actual increase in enhancement. In this case, the tumor progression area may be misjudged and expanded. During the postoperative concurrent chemoradiotherapy phase, a common scenario is that the enhancement signal intensity shows a brief increase in the third to fifth follow-up period and then falls back spontaneously, while the diffusion restriction degree does not change significantly during the same period. The offset of the three dimensions combined in this period is naturally compressed due to the inconsistent direction of change in each dimension. Solely relying on the difference in enhancement dimension to determine the tumor progression area will lead to overjudgment in this period. The dynamic changes in images at each follow-up period node form a change amplitude sequence in three feature dimensions. Time nodes in the sequence whose change amplitude exceeds a set threshold are marked as dynamic abnormal nodes. The voxel range corresponding to the dynamic abnormal node in the spatial dimension is the candidate range of the tumor progression area. The candidate range needs to be verified by the spatial consistency of adjacent nodes in the follow-up baseline. Isolated candidate ranges that appear in a single period but are not reflected in the previous and later periods are marked with low stability. The tumor progression area is generated by merging the candidate ranges that have passed the consistency verification. The dynamic change amplitude and dimensional source of each voxel position are recorded synchronously.

[0018] In some embodiments, the step of performing feature decomposition in the tumor progression region to establish recurrence determination rules includes: performing multi-scale gradient decomposition on the tumor progression region to obtain an inter-layer differential response map; extracting the maximum offset decomposition layer from the inter-layer differential response map layer by layer to determine the structural remodeling candidate layer; extracting remodeling vectors consistent with the functional region direction based on the structural remodeling candidate layer to generate a directional remodeling distribution map; and establishing recurrence determination rules by calibrating high remodeling direction intervals based on the directional remodeling distribution map.

[0019] Multi-scale gradient decomposition was performed on the tumor progression region to obtain inter-slice differential response maps. The image intensity distribution within the tumor progression region was decomposed into two types of gradient components: coarse-scale and fine-scale. The coarse-scale gradient captured large-scale enhancement boundary drift across slices, while the fine-scale gradient reflected signal perturbations of local microstructures within the slice. The superposition pattern of these two types of gradient components in the tumor progression region can distinguish between two progression patterns: homogeneous edema expansion and invasive remodeling of the borders. Gradient decomposition was performed independently along the axial, coronal, and sagittal planes. The decomposition results in different directions spatially reflect the anisotropic expansion characteristics of the tumor progression region, and gradient changes along the functional area's direction are preferentially preserved in the inter-slice differential response maps. In the coronal plane, the tumor progression region adjacent to the corpus callosum often exhibits focal abrupt increases in fine-scale gradients. The corresponding coarse-scale gradient amplitude in the axial plane is relatively stable. This asymmetry in the decomposition results suggests a directional drift of the enhancement boundary along the white matter fiber bundle direction. Such anisotropic patterns form high-amplitude clusters in local areas of the inter-layer differential response map, making them key targets for maximum offset decomposition layer extraction. Voxels with amplitudes below the noise threshold in the inter-layer differential response map are not included in the candidate selection during the maximum offset decomposition layer extraction stage. The pseudo-increase in gradient at the edge of the tumor progression region due to partial volume effects is compressed using a slice thickness-related correction coefficient during inter-layer differential response map generation, ensuring that the amplitude distribution of the inter-layer differential response map accurately reflects the structural dynamics within the tumor.

[0020] The maximum offset decomposition layer is extracted layer by layer from the inter-layer differential response map to determine the candidate layers for structural remodeling. Based on the deviation of the amplitude peak of each layer from the adjacent layer in the inter-layer differential response map, when the amplitude peak shows a cross-layer surge and the surge occurs synchronously in at least two of the three decomposition directions, the corresponding layer is identified as the maximum offset decomposition layer and included in the candidate layers for structural remodeling. The requirement of synchronous occurrence in directions excludes false surges caused by noise from unidirectional acquisition. In the area adjacent to the surgical cavity of intracranial tumors, surgical scar tissue and recurrent tumors are difficult to distinguish on imaging. The inter-layer differential response map of this area often shows high amplitude aggregation in the early postoperative period. The amplitude contribution of voxels at the surgical cavity boundary is attenuated by anatomical prior distance during the layer-by-layer extraction process. The voxel closer to the surgical cavity has a larger attenuation coefficient to reduce the influence of surgical residual cavity signal interference on the determination of candidate layers for structural remodeling. In the inter-layer difference response diagram, nodes with systematic amplitude increases due to equipment upgrades or changes in layer thickness need to be corrected for the corresponding layer amplitude during maximum offset extraction by combining the equipment difference annotations of the node in the follow-up baseline set. Uncorrected equipment change nodes may generate false candidate segments spanning multiple layers in the structural remodeling candidate layer. The structural remodeling candidate layer adds suspected equipment factor annotations to candidate segments from such sources. In the directional remodeling vector extraction stage, a stricter directional consistency verification threshold is applied to candidate segments with such annotations to reduce the interference of equipment factor confusion on the establishment of recurrence determination rules.

[0021] An orientation remodeling distribution map is generated by extracting remodeling vectors consistent with functional area orientation from the structural remodeling candidate layer. Remodeling vectors are extracted from the gradient directions of each voxel position within the structural remodeling candidate layer. Voxels in the structural remodeling candidate layer with an angle θ lower than the consistency threshold θ_thresh are identified as functional area orientation-consistent voxels, and their gradient direction vectors are included as remodeling vectors in the corresponding positions of the orientation remodeling distribution map. The orientation consistency score is defined as S_dir = cos(θ) × A_grad, where θ is the angle between the voxel's gradient direction and the functional area's direction of travel, and A_grad is the normalized gradient magnitude value at the voxel's position. A higher S_dir indicates a more significant projection intensity of the voxel's remodeling vector onto the functional area direction. In tumor progression regions adjacent to the motor cortex, when the remodeling vectors show a higher concentration of S_dir perpendicular to the cortical direction, the orientation remodeling distribution map forms a high-density clustering band pointing towards the motor functional area in this region. The spatial width of the clustering band and the mean S_dir together reflect the activity level of the tumor progression region extending towards this functional area. The reference baseline for functional area orientation is derived from pre-stored annotations of white matter fiber bundle courses in preoperative images. When postoperative functional area deviation occurs due to intraoperative cavity remodeling and edema compression, the reference baseline is corrected using estimated functional area positions from the most recent follow-up baseline. An uncertain position interval is added to the correction value to reflect the estimation error of functional area course changes. After orientation screening, voxel regions with high S_dir and spatial clustering in the orientation remodeling distribution map reflect the most active local locations extending towards functional areas within the tumor progression region. The distribution density of such regions constitutes the core input for high remodeling orientation interval labeling.

[0022] Based on the directional remodeling distribution map, high remodeling direction intervals were identified to establish recurrence determination rules. Continuous voxel regions with an S_dir mean exceeding 1.5 times the global mean were selected from the directional remodeling distribution map as candidate regions for high remodeling direction intervals. If the consistency score of this continuous region in the functional area direction remained consistently high, it was identified as a high remodeling direction interval and included in the spatial triggering conditions of the recurrence determination rules. The amplitude threshold was set at 1.5 times the global mean rather than a more stringent multiple, mainly considering that the directional remodeling amplitude in the early stages of tumor progression had not yet reached a significant peak; an overly strict threshold would lead to overall missed detection of high remodeling direction intervals in early cases. Tumor progression in the frontal lobe white matter region often appears as a continuous high-value band of S_dir along the pyramidal tract direction in the directional remodeling distribution map. This band-like distribution triggers the priority determination of the functional area direction when high remodeling direction intervals are identified. As the interval boundary extends towards the pyramidal tract, the spatial coverage of the recurrence determination rules expands accordingly. The extension segment projected from the distal pyramidal tract to the motor cortex is simultaneously included in the triggering conditions, allowing the trend of tumor expansion along the distal fiber tract to be pre-identified within the rule coverage area. The recurrence determination rule is triggered by three parameters: spatial location of the high remodeling direction interval, amplitude threshold, and directional consistency threshold. The trigger intensity I_rule = w1 × R_spatial + w2 × R_amplitude + w3 × R_direction, where R_spatial is the spatial location matching degree, R_amplitude is the degree of amplitude threshold satisfaction, and R_direction is the degree of directional consistency threshold satisfaction. The sum of the three weights w1, w2, and w3 is 1. When any parameter is lower than the corresponding threshold, the corresponding R value is set to zero. I_rule decreases with the number of missing parameters. When all three parameters are satisfied, the recurrence determination rule outputs the highest trigger intensity. Under the highest trigger intensity state, the rule has the most significant effect on regulating the voxel assignment threshold of subsequent perfusion-diffusion joint analysis.

[0023] In some embodiments, the step of generating a false remission identifier based on the contradictory response identification between the tumor progression region and the clinical test data includes: performing statistical analysis on enhancement signal changes in the tumor progression region to determine the enhancement trend distribution; performing a KPS synchronicity test based on the enhancement trend distribution and the clinical test data to obtain a functional response checklist; screening contradictory pairs from the functional response checklist that show weakened enhancement signals but a synchronous decrease in KPS to identify a set of contradictory responses; and generating a false remission identifier based on the duration of the contradictory response candidate set.

[0024] Statistical analysis of enhancement signal changes in the tumor progression region was performed to determine the enhancement trend distribution. The enhancement T1 signal intensity at each voxel location within the tumor progression region was statistically analyzed along the time axis of the follow-up baseline set. The signal intensity change amplitude was normalized between adjacent nodes to form a signal change rate sequence. The mean trajectory of the signal change rate sequence across all voxels constitutes the overall enhancement trend of the tumor progression region. The enhancement trend distribution is not a single mean curve, but a statistical description encompassing both the mean trajectory and fluctuation ranges. The fluctuation range reflects the spatial heterogeneity of signal changes at each voxel location within the tumor progression region; periods with continuously increasing fluctuation range widths indicate that the spatial distribution of enhancement changes within the tumor progression region is becoming increasingly uneven. In the weeks following the end of temozolomide maintenance therapy, some patients exhibited a biphasic waveform in the voxel signal change rate sequence around the central necrotic lesion in the tumor progression area, initially decreasing and then increasing. Conversely, the signal change rate sequence around the necrotic lesion showed a continuous decline during the same period. This separation of the two sets of voxel signal change directions led to a sharp expansion of the fluctuation range during this time. The enhancement trend distribution added an increased spatial heterogeneity marker to this biphasic waveform period. During the KPS synchronicity test, this marker could be used to identify a hidden pattern where some areas within the enhancement trend distribution showed substantial progression while the overall mean still showed a decline. Nodes with composite low-quality markers in the follow-up baseline set had their corresponding voxel change amplitudes substituted with corrected estimates when calculating the signal change rate. Directly including uncorrected nodes in the statistics would cause inaccurate signal peaks in the enhancement trend distribution for the corresponding time periods. The quality source of each node in the enhancement trend distribution was marked for each time period. Based on this, the KPS synchronicity test adaptively adjusted the trend reference weights for low-quality time periods.

[0025] A functional response comparison table was obtained by performing a KPS synchronicity test based on the enhancement trend distribution and clinical test data. The goal of the KPS synchronicity test is to quantify the temporal consistency between the signal change direction of the enhancement trend distribution and the KPS change direction in the clinical test data. When the two directions are consistent, it reflects that the changes in imaging and functional status are in a synchronous response state. When the directions are contradictory, it indicates that there is a decoupling phenomenon between the two. Decoupling phenomena tend to occur in the typical course of pseudo-remission within several weeks after the end of the treatment response period. The synchronicity test is conducted by comparing the directional signs of the mean signal change rate of the enhancement trend distribution and the KPS change at the corresponding node in the clinical test data within the same time window. When the mean signal change rate decreases while the KPS change is negative at the same time, it is recorded as a directional deviation event. The distribution density and duration of directional deviation events at all time nodes constitute the core content of the functional response comparison table. When periods marked with increased spatial heterogeneity in the enhanced trend distribution were included in the KPS synchronicity test, the synchronicity test distinguished between two levels: the overall mean direction and the direction of local high-signal areas, comparing them with the changes in KPS. Periods where the overall mean was in the same direction as KPS but the local high-signal areas were in the opposite direction to KPS were marked with a local decoupling label in the functional response control table. The local decoupling label suggests that the false remission may only occur in a local area of ​​tumor progression rather than the entire tumor. Each node in the functional response control table recorded the mean rate of change of signal, the change in KPS, and the direction sign, comparing the results with the paired confidence of the corresponding node. Directional deviation events of low-confidence nodes were marked as low-confidence deviations in the functional response control table. The weight of low-confidence deviations in the contradiction pairing screening stage was lower than that of normal-confidence nodes. The overall proportion of low-confidence deviations in the functional response control table reflects the data quality of the current follow-up data at the contradiction response identification level.

[0026] For example, the step of screening contradictory pairings to identify a candidate set of contradictory responses from the functional response comparison table where the reinforcement signal weakens but the KPS decreases concurrently includes: performing statistical analysis on the directional differences between reinforcement changes and KPS changes in the functional response comparison table to obtain a directional divergence distribution; identifying continuous reverse periods of reinforcement-KPS from the directional divergence distribution to generate a set of divergence duration intervals; sorting the set of divergence duration intervals by confidence to form a divergence location set; and filtering based on the duration threshold of the divergence location set to generate a candidate set of contradictory responses.

[0027] The directional divergence distribution was obtained by statistically analyzing the directional differences between enhancement changes and KPS changes in the functional response control table. The mean sign of the signal change rate and the sign of the KPS change were compared node by node in the functional response control table. When the signs were the same, it was recorded as a directional state; when the signs were opposite, it was recorded as a directional divergence state. The divergence magnitude was represented by the normalized product of the decrease in signal change rate and the absolute decrease in KPS. Nodes with larger divergence magnitudes were represented with higher weighting density in the directional divergence distribution, allowing priority to be given to periods where both signal and functional divergence were significant during the divergence persistence interval identification phase. During the transition period from the end of radiotherapy to the start of maintenance chemotherapy, the image enhancement signal often showed a phased decrease due to blood-brain barrier repair, while the patient's KPS continued to decline due to cumulative radiotherapy fatigue. directional divergence nodes during this period appeared densely and continuously in the functional response control table, forming high-density clusters in the directional divergence distribution. The peak position of the density waveform in these clusters directly indicated the concentrated period of contradictory responses on the time axis, and the waveform morphology of the directional divergence distribution in these clusters directly determined the boundary location of the divergence persistence interval set. In the functional response comparison table, the weighted density of low-confidence divergence nodes is reduced according to confidence level, and the divergence amplitude of clinical event-related labeled nodes is compressed proportionally based on the duration of the event. These two types of reduction processing make the shape of the directional divergence distribution more reflective of the true distribution law of tumor-derived contradictory responses. After reduction, the weighted density value of each node is separately labeled in the directional divergence distribution. The boundary delineation strategy for high-quality clustering areas and low-quality clustering areas is based on this during the divergence duration interval set generation stage.

[0028] This method identifies reinforced KPS reverse continuous periods from directional divergence distributions, generating a set of divergence persistence intervals. Reverse continuous periods are identified using the weighted density waveform of the directional divergence distribution. Periods in the directional divergence distribution that continuously exceed a set density threshold are marked as reverse continuous periods. Within a period, at least two adjacent nodes must satisfy directional divergence and both weighted densities must exceed the threshold. A single isolated high-density node does not constitute a continuous period; it is recorded in an appendix as a single-point divergence but does not participate in the main path's duration selection. The interval boundaries of continuous periods are determined by the threshold crossing points of the directional divergence distribution density waveform. Nodes whose density crosses upwards from below the threshold are the interval's starting boundary, and nodes whose density falls back below the threshold are the interval's ending boundary. The proportion of low-confidence divergence and event-associated nodes within the interval is simultaneously recorded in the divergence persistence interval set. This proportion data directly participates in the penalty term calculation during the confidence ranking stage. Within weeks of temozolomide dose adjustment, the enhancement signal caused by transient changes in blood-brain barrier permeability is a non-tumor-derived decrease. If the patient's KPS decreases concurrently due to myelosuppression side effects, the resulting inverse continuous time intervals are clinically considered a treatment response rather than a pseudo-remission. The deviation from the continuous interval set adds adjacent treatment event labels to the intervals occurring within a set period after the treatment adjustment record. During the confidence ranking stage, the interval ranking score with this label is reduced by a fixed percentage to avoid contradictory pairings within the treatment adjustment window being processed with high priority and mixed into the generation path of pseudo-remission labels.

[0029] The set of persistent divergence intervals is ranked by confidence to form a divergence location set. A comprehensive score is calculated for the data quality and signal characteristics of each interval within the persistent divergence interval set to form a confidence ranking. The ranking score is synthesized from three indicators: the weighted density mean of nodes within the interval, the average paired confidence of nodes covered by the interval, and the proportion of low-confidence divergence nodes. Intervals with higher weighted density mean and higher average paired confidence scores are ranked higher. Intervals with a high proportion of low-confidence divergence nodes are penalized, with the penalty magnitude positively correlated with the proportion. Intervals adjacent to treatment events are further reduced by a fixed percentage after the three indicators are synthesized. Intervals that still exceed a set quantile threshold after reduction are retained in the divergence location set for duration screening. Intervals that are below the threshold after reduction are moved to an appendix for archiving and are not included in the main path for generating the candidate set of contradictory responses. For a patient's postoperative follow-up period, the intervals of divergence persistence, due to the overall low average paired confidence of the covered nodes and a high proportion of low-confidence divergences, resulted in ranking scores falling below the threshold under the penalty effect. These intervals were archived as low-scoring intervals and can be included as extended candidate sources for evaluation when the overall quality of the follow-up baseline data is low. The divergence location set is arranged in descending order of ranking score. A highly concentrated score distribution in the high-score segment indicates a relatively uniform quality of the current contradictory response signal, while a dispersed distribution indicates significant differences in data quality across intervals. During the duration screening phase, caution should be exercised regarding the screening conclusions for low-scoring intervals.

[0030] A candidate set of contradictory responses is generated based on a duration threshold screening method using the deviation location set. The duration threshold is set according to the typical time-series characteristics of pseudo-remission. In clinical observation, pseudo-remission usually lasts for several weeks to several months. Too short a duration is more likely to reflect normal imaging fluctuations or individual differences in physician assessments rather than the true course of pseudo-remission. If the threshold is too low, the candidate set of contradictory responses will include a large number of noisy intervals, reducing the final credibility of the pseudo-remission label. The duration of each interval in the deviation location set is calculated based on the interval's start and end boundary and the distance between follow-up time nodes. Intervals with a distance that reaches or exceeds the duration threshold are directly added to the candidate set of contradictory responses. Intervals with a distance below the threshold but a ranking score in the top quantile of the deviation location set and a significantly higher internal weighted density mean are added to the candidate set of contradictory responses with low duration and high density labeling instead of being directly discarded. Low duration and high density labeling indicates that the contradictory response in this interval is atypical in time but has reference value in intensity. During the pseudo-remission label generation stage, such entries are converted to equivalent duration using a density compensation mechanism before performing rank mapping. Each entry in the candidate set of conflict responses carries a sorting score and duration screening pass type from the deviation location set. The type distinguishes between two situations: qualified writing and low-duration high-density writing. The two types of entries have different duration labeling weights in the pseudo-relief mark generation stage, ensuring that the information granularity of the candidate set of conflict responses can support the pseudo-relief mark to differentiate the degree of relief duration.

[0031] False remission indicators are generated based on the duration calibration of the candidate set of contradictory responses. Duration calibration is performed on the duration values ​​of each item in the candidate set of contradictory responses. The calibration results are then weighted by quality, taking into account the ranking score of the corresponding item. Two items with the same duration value will have different calibration intensities due to different ranking scores; the item with the higher ranking score has a higher calibration intensity, and its generated false remission indicator level is correspondingly improved. For items in the candidate set of contradictory responses that are labeled with low duration and high density, a density compensation mechanism is used during duration calibration. The internal weighted density mean is converted to an equivalent duration using a set conversion factor, and then weighted and synthesized with the actual duration. The converted comprehensive duration participates in the level mapping. In some patients at the end of the treatment response period, due to the sparse distribution of follow-up nodes, the duration of the contradictory response interval is insufficient, but the deviation amplitude remains high. The density compensation mechanism ensures that the false remission signal of such items is not completely suppressed due to the duration gap. The duration threshold for pseudo-progression window entries is adjusted upwards during duration calibration. The adjusted duration requirement is higher than that for non-pseudo-progression window entries. This ensures that contradictory pairings during peak pseudo-progression periods require a longer duration to trigger the same pseudo-remission marker as non-pseudo-progression periods, thus preventing short-term imaging contradictions during radiotherapy reaction periods from being misjudged as pseudo-remissions. Pseudo-remission markers are output along with their corresponding time intervals. The levels are divided into low-suspection, moderate-suspection, and high-suspection categories. The comprehensive score range for duration calibration for each category is determined based on the distribution characteristics of pseudo-remission cases in historical follow-up data. The time interval for pseudo-remission markers is aligned with the follow-up baseline set nodes.

[0032] Step S13: Perform perfusion-diffusion combined analysis on pseudoremission markers according to recurrence determination rules to form active tumor areas and peritumoral edema areas. Quantify the risk of infiltration and establish a recurrence characteristic index based on the time period when the expansion rate of peritumoral edema areas precedes that of active tumor areas.

[0033] Specifically, perfusion-diffusion combined analysis was performed on pseudo-remission markers according to recurrence criteria to form active tumor regions and peritumoral edema regions. Within the time interval corresponding to the pseudo-remission marker, perfusion and diffusion parameters were extracted periodically from the follow-up baseline set. Perfusion parameters were primarily based on relative cerebral blood volume, while diffusion parameters were primarily based on the apparent diffusion coefficient. The combined distribution of these two types of parameters at the same voxel location determined the voxel's regional classification. The trigger strength of the recurrence criteria modulates the voxel classification threshold. Spatial locations with high trigger strength, where the recurrence criteria already provide a clear spatial orientation, allow single-parameter dominance for classification to avoid over-removal of truly active voxels. Locations with low trigger strength employ stricter two-parameter consistency requirements to prevent mis-inclusion. This differentiated strategy is particularly important within the low-probability range of pseudo-remission markers, where the voxel characteristics of true recurrence and treatment response are highly overlapping, making single-parameter dominance prone to cross-regional classification errors. The spatial trigger conditions of the recurrence criteria simultaneously constrain the candidate voxel range during classification; voxels outside the trigger condition coverage area do not participate in the classification of active tumor regions and peritumoral edema regions. Voxels with relatively high cerebral blood volume and low apparent diffusion coefficient were classified as tumor active regions, while voxels with moderate cerebral blood volume but slightly limited apparent diffusion coefficient were classified as peritumoral edema regions. The combined two-parameter distribution of voxels surrounding the central necrosis zone often lies at the classification boundary. These voxels are subject to additional boundary fuzzy labeling. During the quantification of infiltration risk, the contribution of the expansion rate of voxels with fuzzy boundaries is handled using interval estimation instead of point estimation. The spatial boundaries between tumor active regions and peritumoral edema regions are generated independently at each follow-up period, and the direction and magnitude of boundary drift are recorded in the spatial boundary sequence for each follow-up period.

[0034] In some embodiments, the step of quantifying the risk of invasion and establishing a recurrence characteristic index based on the time period when the expansion rate of the peritumoral edema region precedes that of the active tumor region includes: calculating the expansion rate of the peritumoral edema region and the active tumor region respectively to obtain a dual-region rate sequence; determining the directional advance time period by performing a functional area directional expansion ratio test on the dual-region rate sequence; generating an invasion grade distribution by classifying the invasion intensity based on the directional advance time period and the peritumoral edema region; and establishing the recurrence characteristic index by performing risk threshold mapping based on the invasion grade distribution.

[0035] The expansion rate was calculated separately for the peritumoral edema area and the active tumor area to obtain a dual-region rate sequence. The expansion rate was calculated based on the periodic displacement of the spatial boundaries of the two regions at each follow-up period. The displacement was defined as the normalized value of the newly added coverage area of ​​the current boundary voxel set relative to the previous boundary voxel set. The newly added coverage area was calculated in three-dimensional space along the axial, coronal, and sagittal directions and then the weighted average was taken. The expansion rate v_exp=Σ_k cos(φ_k)×ΔA_k, where k∈{ax,cor,sag} corresponds to the axial, coronal, and sagittal directions, respectively, ΔA_k is the normalized value of the newly added coverage area in the k-th direction, φ_k is the angle between the k-th direction and the functional area's direction of travel, and cos(φ_k) is used as the directional weight to maximize the weight of the expansion displacement in the functional area's direction of travel. In the early stages of glioma recurrence, peritumoral edema often exhibits a precessive pattern where its expansion rate is consistently higher than that of the active tumor area. The physiological mechanism lies in the fact that when tumor cells infiltrate along white matter fiber bundles, they preferentially manifest as edema rather than enhancement on imaging. The enhanced area only gradually catches up with the edema boundary several weeks after infiltration. The capture of this temporal difference by dual-region rate sequences is the core basis for identifying directional lead periods. In a patient during the eighth to twelfth postoperative follow-up period, the expansion rate component of the peritumoral edema area along the pyramidal tract was consistently higher than the component of the active tumor area in the same direction, while the overall expansion rate of the active tumor area remained relatively stable during the same period. The dual-region rate sequence exhibited a typical asymmetric pattern of edema preceding enhancement during this period. This pattern can directly trigger the marking of continuous lead states during the directional lead period identification stage. In the dual-region rate sequence, the rate values ​​corresponding to the low-quality labeled nodes in the follow-up baseline set are written as estimated intervals instead of point estimates. The width of the estimated interval reflects the uncertainty of the expansion rate at that stage. During the directional lead period identification stage, uncertainty labels are added to the lead judgment conclusions involving nodes of the estimated interval.

[0036] The functional zone directional expansion ratio test was used to determine the directional advance period in the dual-region rate sequence. For each phase of the dual-region rate sequence, the expansion rate component of the peritumoral edema area in the functional zone direction was compared with the expansion rate component of the active tumor area in the same direction. In the dual-region rate sequence, a phase with a positive difference exceeding a set advance threshold was recorded as an advance phase; a negative difference or not exceeding the threshold was recorded as a non-advanced phase. Phase sequences with consecutive advance phases constituted the candidate range for directional advance periods. Isolated single-phase advance phases did not constitute directional advance periods and were recorded as notes but not included in the invasion intensity grading path. The advance threshold setting needed to balance sensitivity and specificity. A threshold that was too low would misjudge normal edema fluctuations as invasion advance, while a threshold that was too high would miss early low-amplitude invasion progression signals. The distribution of directional expansion differences in confirmed relapse cases in historical follow-up data was the main basis for determining the advance threshold. In cases adjacent to the motor functional area, the rate component of peritumoral edema along the pyramidal tract often precedes the increase in enhancement in the active tumor area several weeks prior. The directional lead phase exhibits a clear, continuous lead sequence in the dual-region rate sequences of these cases. The lead amplitude and the duration of the lead phase jointly determine the input intensity during the invasiveness grading stage. The output of the directional lead phase also carries the lead difference sequence for each phase. The amplitude change trend of the difference sequence distinguishes between two invasive progression patterns: one where the lead amplitude increases over time, and the other where it plateaus.

[0037] Infiltration intensity grading was generated based on the directional advance time period and the peritumoral edema area to create an infiltration grading distribution. Infiltration intensity grading was performed voxel-by-voxel on the peritumoral edema area during each follow-up period covered by the directional advance time period. Grading was based on a normalized intensity score derived from the product of the cumulative expansion displacement of each voxel location within the directional advance time period and the proportion of directional expansion towards the functional area. Voxel locations with high intensity scores reflected that the infiltration at that location was continuously advancing towards the functional area with a significant cumulative magnitude. Infiltration intensity levels were divided into three categories: mild, moderate, and severe. Mild corresponds to an intensity score below the global mean, moderate corresponds to an intensity score above the global mean but below the set upper limit, and severe corresponds to an intensity score above the set upper limit. These three levels are presented in the infiltration grading distribution as voxel-level grade labels. The spatial aggregation degree and distribution direction of voxels at each level together constitute the morphological characteristics of the infiltration grading distribution. When the cumulative expansion displacement of peritumoral edema voxels adjacent to areas with dense white matter fiber bundles is the same, they receive higher intensity scores due to the higher proportion of expansion in the functional area direction. This weighting mechanism makes the infiltration grade distribution more sensitive to the infiltration progression in the functional area direction. In cases where the magnitude of directional advancement increases during the directional advancement period, the spatial range of severe voxels in the infiltration grade distribution continues to expand with the progress of follow-up periods. The higher the consistency between the expansion direction of the severe voxel aggregation area and the functional area direction, the higher the corresponding recurrence characteristic index output value. The overall proportion of severe voxels and the spatial morphology of the aggregation area in the infiltration grade distribution are output in the form of voxel level labels.

[0038] A recurrence characteristic index was established by performing risk threshold mapping based on the infiltration grade distribution. Risk threshold mapping transforms the spatial distribution characteristics of voxels at each grade in the infiltration grade distribution into a quantifiable and comparable recurrence risk index. The mapping process uses three parameters as inputs: the proportion of severely infiltrated voxels in the total number of voxels in the peritumoral edema area (R_severe), the normalized value of the nearest distance between the severely infiltrated voxel aggregation area and the functional zone boundary (D_inv), and the normalized value of the duration of the directional lead time (T_lead). Here, D_inv = d_min_ref / d_min, where d_min is the measured nearest distance, d_min_ref is the set reference distance benchmark, and the normalized D_inv is the reference distance. _inv is a dimensionless value. The recurrence characteristic index I_rec is a weighted composite of three parameters: I_rec = α × R_severe + β × D_inv + γ × T_lead, where α, β, and γ are weighting coefficients, and their sum is 1. The β coefficient is usually higher than α and γ because the spatial proximity of the severely infiltrated voxel to the functional area has the most direct indicative significance for clinical risk. A high R_severe but a low D_inv usually corresponds to an early stage where the infiltration range is wide but has not yet approached the functional area boundary. The I_rec value is suppressed by the β component and does not trigger the highest warning level. The recurrence characteristic index is normalized to the range of 0 to 1. The higher the value, the higher the overall severity of the infiltration risk in the current follow-up data. In the infiltration gradation distribution, voxels with ambiguous boundary labels are weighted according to their assignment probability when calculating R_severe and included in the proportion of severe voxels. Voxel groups with assignment probabilities near the threshold around the central necrotic lesion contribute between 0 and 1 to R_severe after weighting, and do not cause discontinuous jumps in I_rec due to minor deviations in boundary determination. The recurrence characteristic index carries the original values ​​of three input parameters. During the degeneration characteristic index extraction stage, this can be used to decompose the contribution sources of I_rec, distinguishing the differentiated treatment needs in early warning strategies between high-risk areas dominated by D_inv and those dominated by T_lead.

[0039] Step S14: Based on the recurrence characteristic index, locate key risk nodes in the active tumor area, extract the rapid stabilization segment after the sharp drop in KPS from the clinical test data to generate functional compensation risk markers, and perform joint calibration on the key risk nodes and functional compensation risk markers to generate risk stratification features and degradation feature indicators.

[0040] Specifically, risk-critical nodes are located within the active tumor region based on the recurrence characteristic index. The spatial proximity component of the recurrence characteristic index corresponds to a region of heavily infiltrating voxels, which is mapped at the voxel level as a candidate range. Each voxel position within the candidate range inherits the recurrence characteristic index value of its respective region. This value, along with the density of heavily infiltrating voxels along the functional zone's direction, jointly determines whether a voxel is ultimately identified as a risk-critical node. This combined determination avoids misclassifying inactive enhancing voxels as high-risk nodes based solely on spatial proximity. Voxel positions with inherited recurrence characteristic index values ​​exceeding the global mean and clustering densities exceeding a set density threshold are marked as risk-critical nodes. The spatial distribution of risk-critical nodes within the active tumor region reflects the current concentration trend of recurrence risk. When nodes are highly clustered near the functional zone boundary, it indicates that the tumor invasion front has entered a critical range for functional protection. The localization process also needs to exclude areas with artificially high recurrence characteristic indices caused by the boundary effect of surgical cavity remnants within the active tumor region. The relative cerebral blood volume of voxels adjacent to the surgical cavity is systematically higher due to postoperative blood-brain barrier disruption. An exclusion buffer is set at the surgical cavity boundary during the localization phase; voxels within this buffer do not participate in the final confirmation of critical risk nodes to avoid spatial contamination of the node localization results by surgical cavity remnant signals. The I_rec value, spatial coordinates, and cluster identifier of each critical risk node are jointly written into the output. The cluster identifier is used to group spatially adjacent critical risk nodes into the same calibration unit. The merged calibration unit is represented by its centroid coordinates and the highest I_rec value within the unit.

[0041] In some embodiments, the step of extracting a functional compensation risk identifier from the rapid stabilization segment after a sharp drop in KPS from the clinical test data includes: performing KPS time-series change statistics on the clinical test data to determine the KPS fluctuation distribution; performing a recovery rate test after a sharp drop based on the KPS fluctuation distribution to obtain a sharp drop-stabilization pair set; screening segments where the steady-state KPS baseline continues to decline after stabilization from the sharp drop-stabilization pair set to identify a baseline decline candidate set; and generating a functional compensation risk identifier based on the cumulative baseline drop magnitude of the baseline decline candidate set.

[0042] KPS time-series variation statistics were performed on clinical test data to determine the KPS fluctuation distribution. After fusing the clinical test data at each node of the follow-up baseline set, the KPS values ​​were differentially analyzed period by period to form a KPS change sequence between adjacent nodes. The amplitude and direction distributions of these change sequences together constitute the basic form of the KPS fluctuation distribution. The amplitude distribution reflects the severity of functional state fluctuations, while the direction distribution reflects the overall evolution trend of functional state. The comprehensive fusion quality labeling of each node of the clinical test data corrects the reliability of the corresponding changes in the KPS fluctuation distribution. KPS changes corresponding to nodes with low-quality labels are further labeled with low reliability. Changes with low reliability labels do not directly trigger a sudden drop judgment during the recovery rate test phase after a sudden drop, to avoid false fluctuations caused by data quality issues being misidentified as genuine functional drops. When the cumulative mean of negative changes in the KPS fluctuation distribution shows a continuous downward trend, the corresponding drop threshold is appropriately tightened during the testing phase to improve the sensitivity of drop identification in the context of overall functional decline. This dynamic threshold adjustment mechanism ensures that the drop judgment criteria maintain an adaptive correspondence with the patient's current functional baseline level. For patients with a postoperative follow-up period of more than one year and a continuous and slow functional decline, the static threshold may systematically miss true drop events due to the overall downward shift in the baseline level. The dynamic tightening strategy has the most significant improvement effect in the identification of such patients. The changes corresponding to the clinical event-related annotation nodes in the clinical test data are marked as event interference changes in the KPS fluctuation distribution. During the recovery rate testing phase, event-related annotations are added to the drop pair triggered by event interference changes. Based on this, the rank mapping phase adopts stricter duration requirements for such pairings to distinguish between non-tumor events and true functional compensation.

[0043] The sudden drop-stabilization pairing set is obtained by testing the recovery rate after a sudden drop based on the KPS fluctuation distribution. Nodes in the KPS fluctuation distribution with negative changes exceeding the sudden drop threshold trigger backward tracking. The time period where continuous positive changes first appear after a sudden drop and the cumulative recovery reaches a set stabilization ratio is identified. Time periods meeting this condition, along with their corresponding sudden drop nodes, constitute a sudden drop-stabilization pairing and are written into the sudden drop-stabilization pairing set. The recovery rate V_rec = ΔK / Δt_rec, where ΔK is the cumulative recovery from the sudden drop node to the stabilization target node, and Δt_rec is the corresponding time interval. Pairings with high V_rec indicate rapid activation of the compensatory mechanism, while pairings with low V_rec indicate delayed initiation of compensation. These two types of pairings correspond to different baseline stability testing strategies during the baseline decline candidate set identification stage. Continuous pairing sequences with persistently low V_rec are clinically common in late-stage patients whose functional compensation reserves are nearing depletion, and their stabilization speed gradually increases with each sudden drop. If a second sudden drop occurs during the follow-up process after a sudden drop but before stabilization is achieved, the follow-up is terminated and the current pairing is marked as interrupted stabilization. Interrupted stabilization pairings are written with a special annotation into the sudden drop-stabilization pairing set. These pairings reflect situations where the compensatory mechanism is suppressed due to accelerated tumor progression after the initiation of compensation. The dense occurrence of interrupted stabilization pairings is an important basis for improving the functional compensation risk level for the corresponding period. The amplitude values ​​of each sudden drop node in the KPS fluctuation distribution are written into the sudden drop-stabilization pairing set along with the pairing. The magnitude of the sudden drop directly participates in risk quantification in the functional compensation risk identification level mapping. The absolute value of the KPS of the first stable node after stabilization provides a reference anchor for steady-state baseline comparison during the baseline decline candidate set identification stage.

[0044] For example, the step of screening segments of sustained downward shift of steady-state KPS baseline after stabilization from the sudden drop-stabilization pairing set to identify a candidate set of baseline decay includes: performing a baseline stability test on the sudden drop-stabilization pairing set to generate a baseline stability sequence; extracting historical first non-stabilization events after sustained stabilization from the baseline stability sequence to generate a subset of compensatory failure events; performing compensatory failure time node fitting calculation on the subset of compensatory failure events to generate a set of compensatory failure nodes; and identifying segments of sustained downward shift of baseline based on the set of compensatory failure nodes to generate a candidate set of baseline decay.

[0045] A baseline stability test is performed on the plummeting-stabilizing pair set to generate a baseline stability sequence. The standard deviation of the KPS values ​​of consecutive nodes after stabilization in each pair within the plummeting-stabilizing pair set is inversely normalized and used as the baseline stability value for that pair. A smaller standard deviation indicates a more stable steady state after stabilization, and the corresponding stability value in the baseline stability sequence is higher. The window length for the stability test extends forward by a set number of periods from the stabilizing node. Pairs with insufficient nodes within the window in the plummeting-stabilizing pair set are marked with the number of available nodes and calculated based on the actual number. Incomplete window markings are added to the stability value. A conservative threshold is used for stability judgment of incomplete window pairs during the subset extraction stage of compensatory failure events to reduce the risk of misjudgment. In baseline stability sequences, the stability values ​​of each pair, together with their corresponding V_rec values, reflect the activation efficiency and homeostasis maintenance capacity of the compensatory mechanism. Pairs with high V_rec but low stability indicate rapid compensatory activation but insufficient homeostasis maintenance. Pair sequences with both V_rec and stability continuously declining indicate that the compensatory mechanism is declining simultaneously in both activation and maintenance dimensions. This double-decline pattern is a strong predictive signal that the compensatory reserve is nearing depletion. When the stability values ​​of several consecutive sharp drops followed by stabilization show a monotonically decreasing trend, it reflects that the homeostasis maintenance capacity after each compensation is lower than the previous one. This decreasing trend is an important precursor signal before the first failure to stabilize occurs. In patients with long follow-up of more than two years post-surgery, the appearance of this decreasing trend often precedes the clear recurrence signal on imaging by several weeks.

[0046] A subset of compensatory failure events is generated by extracting the first non-stabilization event after a period of sustained stabilization from the baseline stability sequence. Sustained stabilization refers to paired sequences in the plunge-stabilization pair set that continuously meet the stabilization criteria and whose stability values ​​are consistently above the test threshold in the baseline stability sequence. The first non-stabilization event refers to the event where the first plunge after the termination of this continuous sequence fails to trigger stabilization. The transition boundary between these two events is the basis for locating the timing of compensatory failure initiation. This transition boundary corresponds to the node position on the baseline stability sequence where the stability value crosses the threshold. Accurate identification of this position directly determines the starting quality of the fitting of the compensatory failure node set. If the cumulative increase in KPS within the tracking window after a plunge does not reach the stabilization ratio, it is considered non-stabilized. If stabilization still hasn't occurred by the end of the tracking window, the corresponding plunge event is recorded as a non-stabilized event and added to the compensatory failure event subset. The time interval between non-stabilization and the historical sustained stabilization sequence reflects the transition speed from maintenance to failure of compensatory capacity. Patients with a rapid transition speed indicate that neural plasticity reserves are quickly depleted, and the clinical warning time window is relatively limited. The decreasing trend of the baseline stability sequence often shows a stability value approaching the test threshold at the end of the historical sustained stabilization period. This pattern suggests that the compensation capacity is in a critical state. For the subset of compensation failure events, pairings with preceding decline labels are added for events whose stability values ​​have been continuously below the threshold before the occurrence of the stabilization event. During the fitting stage of the compensation failure time node, events with this label are given higher fitting weights. The higher the proportion of preceding decline labeled events, the higher the confidence of the fitting results.

[0047] A subset of compensatory failure events is fitted with time nodes to generate a set of compensatory failure nodes. The occurrence time series of each unrecoverable event in the subset are then subjected to piecewise linear regression to identify the slope inflection point where unrecoverable events transition from sporadic to concentrated occurrences. The physical meaning of this slope inflection point is the time boundary of accelerated compensatory failure; functional degradation after this boundary enters an accelerated phase. Events in the subset with precursory decline labels are doubled in weight during fitting, resulting in higher temporal location reliability than isolated unrecoverable events without precursor signals. Isolated unrecoverable events may originate from sporadic non-tumor health events; reducing their weight can decrease their interference with the fitting inflection point location. For cases with irregular follow-up node intervals, time axis normalization is performed during the fitting process. When follow-up intervals are significantly prolonged due to hospitalization or referral, the density of unrecoverable events within the corresponding period is comparable to the density of events in normal follow-up intervals after normalization, avoiding systematic bias in the fitting of compensatory failure nodes due to differences in follow-up intervals. The time stamp and fitting confidence value of each accelerated compensation failure node are written into the compensation failure node set. The confidence of each node in the compensation failure node set is determined by the proportion of precursor decay labeled events and the fitting residual. The node confidence C_node = p_pre × (1 - ε_fit), where p_pre is the proportion of precursor decay labeled events in the fitting segment corresponding to the node, and ε_fit is the normalized residual of piecewise linear regression. When p_pre is high and ε_fit is low, C_node is close to the upper limit. The tracking segment corresponding to the low confidence node adopts a more conservative cumulative drop threshold to prevent low-quality fitting results from causing false baseline decay candidate segments to enter the functional compensation risk label level mapping path.

[0048] Based on the set of compensatory failure nodes, a baseline decline candidate set is generated by identifying segments with continuously accumulating baseline shifts. Each accelerated compensatory failure node in the set serves as a starting point for tracking. The KPS steady-state baseline sequence of the corresponding node in the follow-up baseline set is examined periodically. When the steady-state baseline sequence is consistently lower than the historical baseline mean before the compensatory failure node within the tracking window, and the magnitude of the drop monotonically increases with node progression, the corresponding tracking segment is marked as a segment with continuously accumulating baseline shifts and added to the baseline decline candidate set. Tracking segments corresponding to low-confidence starting nodes in the set of compensatory failure nodes are marked with low-confidence starting points in the baseline decline candidate set. During the drop magnitude calibration stage, conservative estimation is used for such segments to control the excessive influence of low-quality fitting starting points on the final functional compensation risk identification level. If clinical event-related markers appear during steady-state baseline sequence tracking, the KPS values ​​during the event are corrected for the event before being used in the amplitude calculation. When the steady-state baseline is suppressed by non-tumor factors after acute events such as infection or stress, the uncorrected steady-state baseline values ​​will overestimate the true amplitude of tumor-related functional decline. Correction processing makes the amplitude of the baseline decline candidate set more accurately reflect the degree of independent consumption of functional reserves by tumor progression. Segments where the amplitude of the amplitude is still monotonically increasing at the end of the tracking window are marked as open decline segments. The extrapolated estimation results of the amplitude of the amplitude at the end of the open decline segment are supplemented with a prediction confidence interval. The cumulative amplitude value and time range of each item are directly used in the calibration intensity calculation. The low confidence starting point marker triggers the substitution processing of the lower bound of the estimation interval, and the open decline marker triggers the substitution processing of the lower bound of the extrapolation confidence interval.

[0049] Functional compensation risk markers are generated based on the cumulative baseline drop amplitude of the baseline decline candidate set. The cumulative drop amplitude values ​​of each item in the baseline decline candidate set are time-weighted by the number of covered pairs to form a comprehensive calibration strength. Items with a large number of covered pairs reflect a continuous accumulation of baseline decline over time, and the time-weighted comprehensive calibration strength is higher than that of items with the same drop amplitude but fewer pairs. The comprehensive calibration strength S_deg = F_drop × ln(1 + N_pair), where F_drop is the normalized cumulative drop amplitude, and N_pair is the number of covered pairs. The logarithmic term compresses the nonlinear gain when the number of pairs is extremely large, avoiding uncontrolled overestimation of the calibration strength due to the accumulation of pairs in patients with excessively long follow-up periods. Logarithmic compression ensures that S_deg can provide reasonable intensity estimates for both short-term high-amplitude decline and long-term low-amplitude decline. In the baseline decline candidate set, the cumulative difference of the estimated segments was used in the S_deg calculation during calibration, with the lower bound of the estimated interval. For open decline segments, the lower bound of the extrapolated estimated confidence interval was substituted. These two conservative approaches jointly maintained the overall robustness of the functional compensation risk markers under data uncertainty. The high-compensation marker corresponds to segments where S_deg exceeds the upper threshold, indicating that the patient's functional reserve has been continuously depleted to a low level. The marker results carry the corresponding time interval. In the joint calibration stage, higher joint calibration weights were assigned to the risk key nodes within the time period corresponding to the high-compensation marker. The overall distribution of functional compensation risk marker levels reflects the cumulative degree and duration of functional decline signals in the current follow-up data.

[0050] In some embodiments, the step of jointly calibrating the risk-critical nodes and the functional compensation risk identifier to generate risk stratification features and degradation feature indicators includes: calculating the proximity distance of each node based on the risk-critical nodes; jointly mapping the distance values ​​of each node with the functional compensation risk identifier to obtain a distance association table; selecting pairs of nodes that are close to the functional area and have high compensation intensity from the distance association table to identify effective calibration nodes; and extracting and generating risk stratification features and degradation feature indicators based on the distance attributes and functional attributes of the effective calibration nodes.

[0051] The distance values ​​of each node are determined by calculating the proximity distance to the functional area based on the risk-critical nodes. The shortest Euclidean distance from each risk-critical node to the functional area boundary marker is normalized to form the distance value of each node. The normalization is defined as D_v=d(v,F) / d_ref, where d(v,F) is the shortest Euclidean distance from the risk-critical node v to the functional area boundary marker, and d_ref is the maximum distance within the candidate range of the current follow-up period as the normalization benchmark. The lower the D_v, the closer the risk-critical node is to the functional area boundary in space, and the higher the direct threat of tumor invasion risk to functional impairment at the corresponding location. The functional area boundary markers are derived from the pre-existing markers of the white matter fiber bundle course before surgery and the estimated functional area location values ​​of the most recent follow-up nodes. When there is a positional deviation between the two types of boundary markers, the estimated value is used as the main factor and the pre-existing markers are used as the auxiliary factor to perform a weighted average. The weighted comprehensive boundary marker is supplemented with an uncertain position interval. When calculating the distance value of each node, the outer boundary of the uncertain interval is used as a conservative distance benchmark to ensure that the distance value of each node is not systematically underestimated due to the deviation of the functional area location estimation. When calculating the distance value of adjacent nodes clustered in the same calibration unit in the risk-critical nodes, the centroid coordinates of the cluster area are used instead of the independent coordinates of each node. This avoids the concentrated and distorted distribution of distance values ​​caused by multiple independent calculations for the boundary segments of the same functional area by dense nodes in the cluster area. The centroid distance value is used as the representative value of all nodes in the unit during the joint mapping stage to participate in the pairing determination. When the range of I_rec values ​​of each node in the unit exceeds the set range, an internal heterogeneity label is added. During the joint calibration stage, the pairing weight of such units is appropriately reduced.

[0052] A distance correlation table is obtained by jointly mapping the distance values ​​of each node with the functional compensation risk indicator. The distance values ​​of each node are paired with the corresponding functional compensation risk indicator level for each review period on a timeline. Each record simultaneously carries the distance value, compensation level, review period time indicator, and I_rec value. The I_rec value serves as an auxiliary weight in determining the boundary of effective calibrated nodes. When multiple key risk nodes exist within the same review period, the distance correlation table establishes independent records for each node. The distribution of distance values ​​for each node in multiple records reflects the spatial distribution of tumor infiltration around the functional area in that period. When the distance values ​​of each node are concentrated within the nearest neighbor threshold, it indicates that multiple risk nodes are simultaneously approaching the functional area boundary. This concentrated distribution pattern triggers a high spatial concentration of calibration weights during the effective calibrated node screening stage. When there is a cross-node correspondence between the time interval of the functional compensation risk marker and the review period of the risk critical node, the distance association table adds a cross-period alignment label to the pair. The weight of the record with the cross-period alignment label is reduced according to the alignment time interval during the effective labeling node screening stage. Patients with irregular follow-up intervals have a significantly higher frequency of cross-period alignment labels than patients with regular follow-up intervals. For such patients, the distance association table needs to be evaluated at the overall level to determine whether the cross-period alignment ratio affects the joint mapping quality. Risk critical nodes outside the time intersection range are recorded as records without corresponding labels in the distance association table and are retained separately as isolated spatial risk nodes during the degradation feature index extraction stage.

[0053] The distance association table is used to select pairs of nodes that are close to the functional area and have high compensation strength to identify valid labeled nodes. Records in the distance association table that meet the following criteria and whose distance value is below the set nearest neighbor threshold and whose corresponding functional compensation risk level reaches moderate compensation or above are identified as valid labeled nodes. Records in the distance association table that meet only one criterion are marked as weakly valid and written into an appendix table, retained as supplementary input for degradation feature indicators. Weakly valid labeled records can be used as extended candidates to participate in the auxiliary calculation of risk stratification features in batches where the overall quality of the follow-up baseline data is low. The nearest neighbor threshold is dynamically adjusted based on the width of the uncertain interval of the functional area boundary. During the review period when the uncertain interval is wide, the nearest neighbor threshold is appropriately relaxed to avoid the exclusion of truly high-risk nodes due to slight deviations in the functional area location estimation. The dynamic adjustment range is positively correlated with the width of the uncertain interval. For records in the distance association table whose distance values ​​are near the nearest neighbor threshold boundary but whose I_rec values ​​are significantly high, I_rec auxiliary weights are introduced to perform boundary determination during the identification of effective calibration nodes. Nodes whose comprehensive scores exceed the set boundary determination threshold are included in the effective calibration nodes with effective boundary labeling. The effective boundary labeling performs appropriate reduction on the stratification contribution of the corresponding node during the risk stratification feature extraction stage. The number of effective calibration nodes and the spatial distribution density together reflect the degree of joint consistency between image spatial risk and clinical functional degradation signals. When nodes are highly concentrated in a certain functional area boundary segment, that area is marked as a high-priority spatial risk concentration location.

[0054] Risk stratification features and degradation feature indices are extracted and generated based on the distance and functional attributes of effective calibrated nodes. Effective calibrated nodes with lower distance values ​​and higher grades are assigned a higher contribution weight in the risk stratification features. When nodes cluster towards the low-distance-high-grade quadrant, the corresponding stratification features can more accurately hit the high-risk range of the recurrence determination rules. The risk stratification features include three components: distance-weighted mean, grade-weighted mean, and joint distribution concentration index. The distance-weighted mean reflects the overall spatial proximity of the imaging risk node to the functional area, the grade-weighted mean reflects the overall degree of functional compensation and degradation, and the concentration index reflects the degree of node clustering in the distance-grade two-dimensional space. In high-risk recurrence patients near the motor functional area, the three components typically exhibit the following characteristics: the distance-weighted mean is near the nearest neighbor threshold, the grade-weighted mean falls into the high-compensation range, and the concentration index shows clear focus within the distance-grade quadrant. The degradation characteristic index is extracted based on the cumulative gap magnitude sequence of functional compensation risk markers corresponding to effectively calibrated nodes. Nodes with a continuously accelerating rate of change in the gap magnitude are classified as accelerated degradation components, while nodes with a relatively stable rate of change are classified as plateau degradation components. The proportions of these two types of components together constitute the trend type labeling of the degradation characteristic index. In the late follow-up of a patient with advanced disease, the proportion of the accelerated degradation component was significantly increased, and the concentration index was also at a high level. This pattern corresponds to the synchronous evolution stage of imaging spatial risk and clinical functional degradation, which is a typical combination for priority notification of deterioration warning. If the concentration index has reached a high level but the proportion of the accelerated degradation component is still low, it suggests an evolution pattern in which imaging risk precedes and functional compensation still has a short-term buffer.

[0055] Step S15: Based on the risk stratification characteristics and recurrence judgment rules, trigger deterioration warning and generate warning feature quantity; identify the period of sudden drop in KPS during the follow-up interval based on the degradation characteristic indicators and generate inter-interval warning feature quantity; and match the warning feature quantity with the inter-interval warning feature quantity to output the risk assessment result.

[0056] In some embodiments, the step of generating warning feature quantities based on the risk stratification features and the recurrence determination rule to trigger deterioration warning includes: performing hierarchical matching on the risk stratification features and the recurrence determination rule to obtain hierarchical feature matching regions; filtering pairs with low hierarchical levels but high-risk feature components through the hierarchical feature matching regions to determine mismatch warning regions; sorting the warning levels based on the mismatch warning regions to form a warning location set; and extracting warning feature quantities by classifying the warning intensity of the warning location set.

[0057] A hierarchical feature matching region is obtained by performing stepwise matching on risk stratification features and recurrence determination rules. The process traverses the hierarchical structure of the recurrence determination rules from the highest trigger strength level downwards, checking whether each component of the risk stratification feature meets the trigger conditions of the corresponding level. Levels that meet the conditions are recorded as matches, and the matched levels and corresponding positions constitute the basic members of the hierarchical feature matching region. The matching process also incorporates the adjustment of the matching window width by the risk stratification feature concentration index. A high concentration index narrows the matching window to improve matching accuracy, while a low concentration index moderately widens the matching window to improve coverage. This dynamic window mechanism enables the hierarchical feature matching region to have adaptive matching capabilities among patients with different risk signal distribution patterns. In the recurrence determination rule, the layers corresponding to high remodeling direction intervals are preferentially compared with the spatial proximity distribution of risk stratification features during matching. High-remodeling layers within the protection range of the motion functional area, where the weighted average distance value is low, form a high-intensity match. These matches are distinguished in the hierarchical feature matching area by functional area hazard proximity markers. During the mismatch warning area screening stage, matches marked with functional area hazard proximity are given the highest mismatch warning priority. The hit density and distribution layers in the hierarchical feature matching area together reflect the overall fit between risk stratification features and the recurrence determination rule. When hits are concentrated in high-trigger-intensity layers, it indicates that the recurrence risk features in the current follow-up data meet the strict triggering conditions of the recurrence determination rule. When hits are dispersed across multiple low- and medium-level layers, it indicates that the recurrence risk features are within the rule coverage boundary.

[0058] The mismatch warning zone is determined by screening pairs with low-level but high-risk feature components within the hierarchical feature matching zone. For pairs within the hierarchical feature matching zone that hit a low to medium level but whose corresponding risk stratification feature component values ​​exceed the upper limit of the historical case distribution for the same level, these pairs do not formally meet the high-level triggering conditions of the recurrence determination rules, but the absolute values ​​of their feature components are already within the historical high-risk distribution range. The historical distribution upper limit of the feature components is determined statistically based on the feature values ​​of the corresponding hit level in similar examination data. The statistical window typically covers historical cases with the same pathological type, surgical procedure, and similar follow-up period to ensure that the upper limit estimation is comparable to the current patient's case background. Pairs outside the distribution upper limit within the hierarchical feature matching zone are marked as feature out-of-bounds pairs and included in the mismatch warning zone. The degree of feature out-of-bounds is quantified by the multiple of the standard deviation of the component value exceeding the upper limit. The higher the multiple, the higher the warning level of the pair in the mismatch warning zone. Pairs with a multiple remaining at a high level for multiple consecutive follow-up periods indicate a persistent misalignment between the patient's risk characteristics and the rule hierarchical structure; the cumulative nature of this misalignment is itself an important warning signal. The weighted mean of a patient's risk stratification features reached the 97th percentile of its historical distribution within the mid-level hit, indicating that the patient's functional deterioration has significantly exceeded the typical range for mid-level relapse risk cases. Even if the spatial proximity component has not yet triggered a high-level match, the exceptional performance of the functional dimension is sufficient to include this pair in the mismatch warning zone and assign it a high warning level. Each pair in the mismatch warning zone carries the hit level, the degree of feature component boundary crossing, and the functional area danger proximity labeling status. These three attributes jointly determine the base value of the pair's ranking score in the subsequent warning level ranking stage.

[0059] A warning location set is formed by ranking warning levels based on mismatch warning zones. The ranking score is synthesized by combining the hit level, feature component boundary crossing degree, and functional area hazard proximity labeling status of each pair within the mismatch warning zone. Pairs with higher hit levels have higher base ranking scores. Feature component boundary crossing degree positively weights the ranking score. Pairs with functional area hazard proximity labels receive an additional score boost after the three indicators are combined, ensuring that mismatch warnings in high-risk proximity locations within functional areas are prioritized over similar-level pairs far from functional areas in the warning location set. For pairs with similar ranking scores in the mismatch warning zone, a concentration index of risk stratification features is used as a secondary ranking criterion. Pairs with higher concentration indices indicate a clearer focus on risk signals, and the warning location set prioritizes these pairs over pairs with similar concentration indices but dispersed component distributions. The introduction of the concentration index prioritizes clinical attention towards pairs with clear spatial patterns of risk signals, preventing dispersed pairs from competing for push resources when their primary ranking scores are the same. The size of the early warning location set is constrained by the overall quality of the current follow-up data. For batches with a high proportion of low-quality labeled nodes in the follow-up baseline, the threshold for screening mismatch warning areas is adjusted upwards. After adjustment, the number of pairs entering the early warning location set is relatively reduced, avoiding the misjudgment of inflated feature components caused by data quality issues as genuine mismatch warnings. The early warning location set is arranged in descending order of score. A concentration of high-score segments indicates that the quality of mismatch warning signals in the current follow-up data is relatively uniform. A tailing of the score distribution towards the middle and lower segments indicates that some warning signals originate from fuzzy matching of hierarchical structure boundaries. The priority of handling items in the tailing segment is correspondingly lowered during the risk level matching stage.

[0060] Early warning intensity grading was performed on the early warning location set to extract early warning feature quantities. The ranking score of each item in the early warning location set was discretized into three intensity levels: low warning, medium warning, and high warning, using continuous input. The level boundaries were determined based on the ranking score distribution of confirmed deterioration cases in historical follow-up data. Items near the level boundaries were assigned to the corresponding level after being marked with an additional boundary fuzzy label. The boundary fuzzy label triggered a conservative treatment strategy during the risk level matching stage to avoid risk level jumps caused by slight deviations in level switching for items near the boundaries. There is a difference in the intensity grading weight between spatially driven early warning and function-driven early warning. Items whose triggering component in the early warning location set is dominated by distance attributes have a higher grading weight in the high warning level than items dominated by level attributes. This is because the time window for risk signals where spatial proximity has reached the high-risk range is usually more urgent for functional impairment than the situation where functional deterioration precedes it. High warnings driven by spatial proximity often reflect that the tumor invasion front has completed a critical advancement towards the functional area within the current follow-up period. The early warning feature quantity is composed of the distribution of the number of items at each intensity level in the early warning location set, the corresponding time node, and the source type of the triggering component. When the proportion of high-alert items in the top percentile of the early warning location set is high, the early warning feature quantity shows a high-level clustering pattern in the intensity dimension, indicating that the recurrence risk at the current follow-up node has entered the range requiring active clinical intervention and assessment. The overall shape of the early warning feature quantity reflects the concentration trend of deterioration warning signals in the intensity dimension in the current follow-up data. When the concentration trend continues to cluster in the high segment over multiple follow-up periods, it indicates that the overall risk evolution of the patient is in a rapid escalation stage. When the concentration trend fluctuates with the period, it indicates that the stability of the risk signal needs to be observed cumulatively at more follow-up nodes.

[0061] Based on degradation characteristic indicators, a sudden drop in KPS during the follow-up interval is identified, generating interval warning features. The follow-up interval refers to the time period between two adjacent image re-examinations. During this period, only the dynamic updates of clinical test data are performed, without new image feature input. The trend type label carried by the degradation characteristic indicators becomes the main basis for identifying the risk of a sudden drop in KPS at this stage. The identification process prioritizes scanning follow-up intervals where the proportion of accelerated degradation components in the degradation characteristic indicators exceeds a set ratio. Within the priority scanning interval, a sudden drop period detection is performed on the KPS time-series changes in clinical test data. The criteria for determining the sudden drop period are consistent with the sudden drop-stabilization pairing set generation stage. However, in the follow-up interval identification, the cumulative drop rate of the degradation characteristic indicators is additionally introduced as a priori weight for the probability of sudden drop. The KPS sudden drop judgment threshold in intervals with high cumulative drop rates is appropriately lowered to improve the sensitivity of the interval warning in capturing sudden drop events in the rapid degradation stage. In the degradation characteristic indicators, the follow-up intervals dominated by the platform degradation component correspond to a stage where KPS changes tend to stabilize. For sudden drops within these intervals, a standard threshold is used for detection. Sudden drops in KPS occurring within the platform degradation interval are marked as unexpected drops in the interval warning characteristic quantity, indicating that the drop occurred against a backdrop of relatively stable functional degradation. These sudden drops need to be differentiated from sudden drops in accelerated degradation intervals during the risk level matching stage. The interval warning characteristic quantity covers the scan results of all follow-up intervals. Intervals without sudden drops are recorded as low-risk. Intervals with identified sudden drops carry the location of the sudden drop, the corresponding interval identifier, and the magnitude of the drop. The magnitude of the drop is directly obtained by taking the difference in KPS before and after the drop point. Intervals with a higher proportion of accelerated degradation components in the degradation characteristic indicators enjoy higher priority in the subsequent risk level matching stage.

[0062] Risk assessment results are output based on risk level matching using early warning feature quantities and inter-period early warning feature quantities. Risk level matching incorporates the intensity level distribution of early warning feature quantities and the sudden drop period distribution of inter-period early warning feature quantities into the matching process, examining the synchronization degree of the two types of early warning signals in the time dimension. When a high-level early warning item and a sudden drop period overlap in time, it is identified as a double early warning trigger, with the risk level exceeding the upper limit of any single input trigger. When triggering by overlap, the source type of the trigger component further determines the subdivision of the risk level, with levels dominated by distance attributes taking precedence over those dominated by level attributes. The matching weight of items with ambiguous boundaries in the early warning feature quantities is reduced. For items with unexpected sudden drops in the inter-period early warning feature quantities, the risk reversion probability of similar historical scenarios is introduced as an auxiliary judgment criterion; items with higher reversion probabilities correspond to a moderately higher risk level. The risk assessment result consists of a risk level, corresponding confidence level, and a detailed list of risk sources. The risk level is divided into three categories: low, medium, and high. The detailed list of risk sources summarizes the warning time point, corresponding interval markers, and the magnitude of the sudden drop, facilitating the location of risk signals during clinical follow-up. The confidence level is calculated using the formula C_result = (1 - r_img) × (1 - r_clin) × C_base, where r_img and r_clin represent the proportion of low-quality annotations in the warning feature quantity and interval warning feature quantity, respectively, and C_base is the base confidence score for the matching process. Taking a high-risk patient near a motor function area as an example, the warning feature quantity in the tenth follow-up period shows a high-risk item dominated by distance attributes, while the interval warning feature quantity simultaneously captures an unexpected sudden drop. After matching, a high-risk level and corresponding confidence level are output, and the detailed list of risk sources points to the location of the sudden drop in the tenth follow-up period and the adjacent interval. High-risk outputs with high confidence levels indicate that the risk of recurrence has reached a reliable warning level in both imaging and clinical dimensions. For medium-risk cases, it is recommended to shorten the follow-up interval to monitor the risk evolution trend. For low-risk cases, maintain the routine follow-up strategy. Historical assessment results are continuously accumulated with each batch of examinations.

[0063] To implement the above-described method embodiments, a dynamic assessment method for postoperative recurrence risk of glioma is proposed to achieve the corresponding functional and technical effects. See also... Figure 2 , Figure 2 This application provides a structural block diagram of a dynamic assessment system for postoperative recurrence risk of glioma, comprising: Data integration module 201 is used to collect image follow-up data and clinical test data, and to perform multimodal temporal fusion of the image follow-up data and the clinical test data to generate a follow-up baseline set; Feature extraction module 202 is used to extract the dynamic changes of images in the follow-up baseline set to determine the tumor progression area, perform feature decomposition in the tumor progression area to establish recurrence determination rules, and generate false remission markers based on the contradictory response between the tumor progression area and the clinical test data. Infiltration analysis module 203 is used to perform perfusion-diffusion joint analysis on the pseudo-remission marker according to the recurrence determination rules to form a tumor active area and a peritumoral edema area, and to establish a recurrence characteristic index by quantifying the infiltration risk based on the time period when the expansion rate of the peritumoral edema area precedes that of the tumor active area. The risk labeling module 204 is used to locate key risk nodes in the active area of ​​the tumor based on the recurrence characteristic index, extract the rapid recovery segment after the sharp drop in KPS from the clinical test data to generate a functional compensation risk label, and perform joint labeling on the key risk nodes and the functional compensation risk label to generate risk stratification features and degradation feature indicators. The assessment output module 205 is used to trigger a deterioration warning based on the risk stratification features and the relapse judgment rules to generate a warning feature quantity, identify the period of sudden drop in KPS during the follow-up interval based on the degradation feature indicators to generate an interval warning feature quantity, and perform risk level matching based on the warning feature quantity and the interval warning feature quantity to output a risk assessment result.

[0064] The aforementioned dynamic assessment system for postoperative recurrence risk of glioma can implement one of the methods described in the above embodiments for dynamic assessment of postoperative recurrence risk of glioma. The options described in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining contents of this application's embodiments can be referred to the contents of the above method embodiments, and will not be repeated in this embodiment.

[0065] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for dynamic assessment of the risk of recurrence after glioma surgery, characterized in that, include: Collect image follow-up data and clinical test data, and perform multimodal temporal fusion on the image follow-up data and the clinical test data to generate a follow-up baseline set; The dynamic changes in images of the follow-up baseline set are extracted to determine the tumor progression area. Feature decomposition is performed in the tumor progression area to establish recurrence determination rules. Based on the contradictory response between the tumor progression area and the clinical test data, a false remission identifier is generated. According to the recurrence determination rules, perfusion-diffusion joint analysis is performed on the pseudo-remission markers to form active tumor regions and peritumoral edema regions. Based on the time period when the expansion rate of the peritumoral edema region precedes that of the active tumor region, an invasion risk quantification is performed to establish a recurrence characteristic index. This includes: calculating the expansion rate of the peritumoral edema region and the active tumor region separately to obtain a dual-region rate sequence; performing a functional area directional expansion ratio test on the dual-region rate sequence to determine the directional lead time period; generating an invasion grade distribution by grading the invasion intensity based on the directional lead time period and the peritumoral edema region; and establishing the recurrence characteristic index by performing risk threshold mapping based on the invasion grade distribution. Based on the recurrence characteristic index, key risk nodes are located within the active tumor region. Functional compensation risk markers are generated by extracting rapid stabilization segments after a sharp drop in KPS from the clinical test data. This includes: performing KPS time-series change statistics on the clinical test data to determine the KPS fluctuation distribution; performing a recovery rate test after a sharp drop based on the KPS fluctuation distribution to obtain a sharp drop-stabilization pair set; screening segments where the steady-state KPS baseline continuously declines after stabilization from the sharp drop-stabilization pair set to identify a baseline decline candidate set; generating a functional compensation risk marker based on the cumulative baseline drop amplitude of the baseline decline candidate set; and jointly calibrating the key risk nodes and the functional compensation risk markers to generate risk stratification features and degradation feature indicators. Based on the risk stratification features and the relapse determination rules, a deterioration warning is triggered to generate a warning feature quantity. Based on the degradation feature indicators, a period of sudden drop in KPS during the follow-up interval is identified to generate an interval warning feature quantity. Based on the warning feature quantity and the interval warning feature quantity, risk level matching is performed to output a risk assessment result.

2. The method according to claim 1, characterized in that, The step of establishing recurrence determination rules by performing feature decomposition in the tumor progression region includes: Multiscale gradient decomposition was performed on the tumor progression region to obtain interlayer differential response maps; The maximum offset decomposition layer is extracted layer by layer from the interlayer difference response map to determine the candidate layer for structural remodeling; Based on the structural remodeling candidate layer, remodeling vectors consistent with the functional area direction are extracted to generate a directional remodeling distribution map; Based on the aforementioned directional remodeling distribution map, high remodeling direction intervals are calibrated to establish recurrence determination rules.

3. The method according to claim 1, characterized in that, The generation of false remission markers based on the contradictory response identification between the tumor progression region and the clinical test data includes: Statistical analysis of enhancement signal changes was performed on the tumor progression region to determine the enhancement trend distribution; Based on the enhanced trend distribution and the clinical test data, a functional response comparison table was obtained by performing a KPS synchronicity test. The candidate set of contradictory responses is identified by screening contradictory pairs that show a weakening of the enhancement signal but a simultaneous decrease in KPS from the functional response checklist. A false mitigation identifier is generated based on the duration of the candidate set of contradictory responses.

4. The method according to claim 1, characterized in that, The method of jointly labeling the key risk nodes and the functional compensation risk identifier to generate risk stratification features and degradation feature indicators includes: Based on the calculation of the proximity distance between the functional areas of the aforementioned risk-critical nodes, the distance value of each node is determined. A distance association table is obtained by jointly mapping the distance values ​​of each node with the functional compensation risk identifier; Select effective calibration nodes by filtering pairs that are close to the functional area and have high compensation strength from the distance association table; Risk stratification features and degradation feature indicators are extracted and generated based on the distance and functional attributes of the effective calibrated nodes.

5. The method according to claim 1, characterized in that, The generation of warning feature quantities based on the risk stratification features and the recurrence determination rules to trigger deterioration warning includes: A hierarchical feature matching region is obtained by performing a step-by-step matching on the risk stratification features and the recurrence determination rules; The mismatch warning area is determined by filtering pairs with low-level but high-risk feature components through the hierarchical feature matching area. Based on the mismatched warning areas, the warning levels are sorted to form a warning location set; The warning location set is subjected to warning intensity classification and extraction to generate warning feature quantities.

6. The method according to claim 3, characterized in that, The step of screening contradictory pairings from the functional response lookup table to identify a candidate set of contradictory responses where the enhancement signal weakens but the KPS decreases concurrently includes: Perform directional difference statistics on the functional response comparison table to obtain directional divergence distribution; From the directional divergence distribution, identify the reinforcement-KPS reverse continuous time period to generate a set of divergence persistence intervals; The set of persistent deviation intervals is sorted by confidence to form a deviation location set; Based on the deviation location set, a candidate set of contradictory responses is generated by filtering based on the duration threshold.

7. The method according to claim 1, characterized in that, The process of identifying a baseline decay candidate set by filtering segments of the steady-state KPS baseline that continue to decline after stabilization from the drop-to-stabilization pairing set includes: Perform a baseline stability test on the plummeting-stabilizing paired set to generate a baseline stability sequence; Extract the first non-stabilization event after a period of sustained stabilization from the baseline stability sequence to generate a subset of compensatory failure events; The compensation failure node set is generated by fitting the compensation failure time nodes to the subset of compensation failure events. Based on the set of compensatory failure nodes, a baseline decay candidate set is generated by identifying the continuously accumulating segments of baseline decline.

8. A dynamic assessment system for the risk of recurrence after glioma surgery, characterized in that, include: The data integration module is used to collect image follow-up data and clinical test data, and to perform multimodal temporal fusion of the image follow-up data and the clinical test data to generate a follow-up baseline set; The feature extraction module is used to extract the dynamic changes in images of the follow-up baseline set to determine the tumor progression area, perform feature decomposition in the tumor progression area to establish recurrence determination rules, and generate false remission markers based on the contradictory response between the tumor progression area and the clinical test data. The infiltration analysis module is used to perform a combined perfusion-diffusion analysis on the pseudo-remission markers according to the recurrence determination rules to form active tumor regions and peritumoral edema regions. Based on the time period when the expansion rate of the peritumoral edema region precedes that of the active tumor region, it quantifies the infiltration risk and establishes a recurrence characteristic index. This includes: calculating the expansion rate of the peritumoral edema region and the active tumor region respectively to obtain a dual-region rate sequence; performing a functional area directional expansion ratio test on the dual-region rate sequence to determine the directional lead time period; generating an infiltration grade distribution by grading the infiltration intensity based on the directional lead time period and the peritumoral edema region; and establishing the recurrence characteristic index by performing risk threshold mapping based on the infiltration grade distribution. The risk labeling module is used to locate key risk nodes within the active tumor region based on the recurrence characteristic index, and to extract functional compensation risk identifiers from the rapid stabilization segment after a sharp drop in KPS based on the clinical test data. This includes: performing KPS time-series change statistics on the clinical test data to determine the KPS fluctuation distribution; performing a recovery rate test after a sharp drop based on the KPS fluctuation distribution to obtain a sharp drop-stabilization pair set; screening segments where the steady-state KPS baseline continuously declines after stabilization from the sharp drop-stabilization pair set to identify a baseline decline candidate set; labeling and generating a functional compensation risk identifier based on the cumulative baseline drop magnitude of the baseline decline candidate set; and performing joint labeling on the key risk nodes and the functional compensation risk identifier to generate risk stratification features and degradation feature indicators. The assessment output module is used to trigger a deterioration warning based on the risk stratification features and the relapse determination rules, generate a warning feature quantity, identify the period of sudden drop in KPS during the follow-up interval based on the degradation feature indicators, generate an interval warning feature quantity, and perform risk level matching based on the warning feature quantity and the interval warning feature quantity to output a risk assessment result.

Images