Liver cancer recurrence early screening system based on imagingomics

Abstract

The application discloses a liver cancer recurrence early screening system based on imageomics, relates to the technical field of medical image processing, and comprises the following modules: a data acquisition module, which acquires multi-modal medical data to generate a multi-phase image library; a liver segment grid module, which extracts microenvironment features according to the Couinaud eight segment method to generate a feature matrix; a recurrence type prior module, which outputs the probabilities of intrahepatic metastasis and new canceration types through double-flow fusion; a longitudinal dynamics module, which constructs an improved Gompertz model with Logistic chaos modulation to obtain a feature doubling time and a deviation sequence; a niche determination module, which fuses three-way data to generate a risk heat map; and a screening scheduling module, which directs review and dynamically adjusts the follow-up cycle according to the recurrence type. The application realizes liver segment-level dynamic perception and accurate follow-up resource allocation for liver cancer recurrence, and significantly improves the early detection rate and re-treatment conversion rate.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an early screening system for liver cancer recurrence based on radiomics. Background Technology

[0002] With advancements in liver surgery techniques and perioperative management, radical resection of liver cancer has become the preferred treatment for early-stage hepatocellular carcinoma. However, the recurrence rate within two years post-surgery remains as high as 40% to 60%, making recurrence monitoring a crucial factor in determining patient prognosis. Currently, the mainstream clinical follow-up model typically includes a fixed imaging follow-up plan, namely, enhanced CT or MRI scans at 1, 3, and 6 months post-surgery, and every 3 to 6 months thereafter, supplemented by tumor marker tests such as alpha-fetoprotein (AFP) for auxiliary screening, thus achieving, to a certain extent, regular monitoring of postoperative recurrence. However, existing follow-up techniques for liver cancer still have many shortcomings in practical applications. First, at the feature extraction level, current models only rely on visually visible nodular lesions on imaging to determine recurrence, lacking quantitative extraction of liver segment-level microenvironment features such as fat deposition, iron deposition, portal vein blood supply, tumor margin distance, and blood supply distance under the Couinaud eight-segment division, and also failing to characterize the longitudinal evolution trend of each liver segment feature during follow-up phases. Second, at the recurrence type discrimination level, current models do not distinguish between the differences in dissemination pathways between intrahepatic metastatic and neoplastic recurrences. Heterogeneous information such as primary tumor imaging features, microvascular invasion status, and residual liver regeneration rate lacks a unified fusion discrimination channel, making it difficult to correlate the prior recurrence type with the selection of the follow-up scope. Third, at the follow-up intensity setting level, current follow-up generally uses fixed intervals unrelated to the individual recurrence evolution rate, and there is a lack of quantitative linkage between the follow-up interval and the tumor doubling rate. Furthermore, the existing model lacks a closed-loop mechanism to write back the results of this round of review and drive the next round of screening, making it difficult for the data obtained from the review to be continuously used in subsequent judgments.

[0003] Therefore, how to provide an early screening system for liver cancer recurrence based on radiomics is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an early screening system for liver cancer recurrence based on radiomics. This invention fully integrates key steps such as multi-temporal image library construction, generation of liver segment-level microenvironment feature matrices, prior discrimination of recurrence types through dual-stream inversion coupling, longitudinal dynamic extrapolation using Gompertz chaotic modulation, and a closed-loop system of niche risk scoring and targeted screening. It constructs a full-cycle recurrence monitoring process with temporal quantitative perception of recurrence risk, differential discrimination of recurrence types, dynamic measurement of feature doubling time, visual representation of liver segment-level risk heatmaps, and adaptive adjustment of follow-up cycles. This enables accurate perception of abnormal microenvironment evolution, early differentiation of recurrence types, and targeted allocation of follow-up resources in the context of post-operative follow-up for liver cancer. This invention possesses advantages such as deep expression of liver segment microenvironment, high accuracy in recurrence type discrimination, strong prospective dynamic extrapolation, and flexible follow-up scheduling mechanism. It can significantly improve the early recurrence detection rate, the advance detection time of recurrence, and the conversion rate to radical re-treatment, thereby effectively solving problems in existing technologies such as missing microenvironment features, indiscriminate recurrence type differentiation, one-size-fits-all follow-up intensity, and blindly expanding the scope of follow-up examinations.

[0005] The radiomics-based early screening system for liver cancer recurrence according to an embodiment of the present invention includes the following modules: The data acquisition module is used to collect multimodal medical data from patients and standardize the processing to generate a multi-temporal image library; The liver segment grid module is used to divide the liver according to the Couinaud eight-segment method, extract fat deposition, tumor margin distance, blood supply distance and regeneration slope features segment by segment, generate liver segment-level microenvironment feature matrix, and save the outline of each liver segment to obtain liver segment spatial mask map. The recurrence type prior module is used to extract the primary tumor image feature list, microvascular invasion status and residual liver regeneration rate sequence and input them into the dual-stream attention fusion network to obtain the recurrence type prior vector, and output the probability of intrahepatic metastasis and the probability of new cancer. The dual-stream attention fusion network includes a dual-stream feature extraction module and an inverse coupling cross-modal bridge module. The inverse coupling cross-modal bridge module embeds a cross-modal resonance fusion mechanism. The longitudinal dynamics module is used to extract feature values ​​segment by segment of the liver and input them into the improved Gompertz model to obtain feature doubling time and expected feature values. The improved Gompertz model includes a state initialization module, a dynamic evolution module, an inverse coupling prior injection module, and an extrapolation deviation calculation module. The inverse coupling prior injection module embeds a chaotic modulation prior mechanism and generates a sequence of measured deviations based on the measured images. The niche determination module is used to fuse the liver segment-level microenvironment feature matrix, recurrence type prior vector and expected feature value with the measured deviation sequence, calculate the recurrence probability score, map it to the liver segment spatial mask map to generate a niche risk heat map, and determine the warning level label for each liver segment. The screening scheduling module is used to generate re-examination instructions using risk heatmaps when the warning level label is high-risk. It selects the corresponding liver segment for re-examination based on the probability of intrahepatic metastasis and the probability of new cancerous transformation, adjusts the follow-up cycle based on the feature doubling time, generates a dynamic follow-up plan, and drives the next round of screening.

[0006] Optionally, the operation of the data acquisition module includes: Preoperative enhanced CT / MRI and multimodal medical data were collected from the patient, including postoperative follow-up images at various time points and clinical laboratory data. Using the patient’s preoperative enhanced CT / MRI as the registration benchmark, the images of each time phase are sequentially mapped to the same anatomical coordinate system through rigid registration and non-rigid registration to eliminate the positional offset between the images of each time phase. After registration, the gray values ​​of each temporal image are mapped to a standard distribution with a mean of 0 and a standard deviation of 1 by applying the Z-Score normalization formula to each layer, thereby unifying the gray dimension of each temporal image. The standardized images of each time phase are sorted and stored according to the patient number and the acquisition time, and together with the patient's preoperative enhanced CT / MRI, a multi-time phase image library is generated.

[0007] Optionally, the operation of the liver segment grid module includes: Read the liver images of the patient’s preoperative enhanced CT / MRI from the multi-temporal image library, and divide the liver into eight liver segments according to the Couinaud eight-segment method, with the hepatic vein and portal vein branches as the boundary markers; Eight liver segments were traversed. The signal intensity ratio of fat within each segment was measured as fat deposition, the magnetic susceptibility-sensitive signal intensity within each segment was measured as iron deposition, the proportion of portal vein branch blood flow within each segment to the total blood flow of the liver was measured as portal vein blood supply, the shortest distance between the edge of the liver segment and the edge of the primary tumor was measured as the tumor margin distance, and the shortest path length between the center of the liver segment and the main blood supply vessels within the liver was measured as the blood supply distance. Linear fitting was performed on the characteristic changes of each liver segment at different time points before and after surgery, and the fitting slope was taken as the regeneration slope characteristic. The fat deposition, iron deposition, portal vein blood supply, tumor margin distance, blood supply distance and regeneration slope characteristics of each liver segment are arranged into rows according to the liver segment number. The six characteristics are spliced ​​into a feature row in a fixed order. The feature rows of all liver segments are stacked vertically to generate a liver segment-level microenvironment feature matrix. The outlines of the eight liver segments are drawn layer by layer as binary masks, and the mask values ​​are labeled according to the liver segment number to generate a spatial mask map of the liver segments.

[0008] Optionally, the operation of the recurrence type prior module includes: Images of the liver segment containing the primary tumor were extracted from a multi-temporal image database. The maximum diameter, marginal morphology, and arterial phase enhancement of the primary tumor were measured and summarized into a list of imaging features of the primary tumor. The microvascular invasion status was read from the pathology report recorded in the clinical test data. The volume change rate between adjacent time phases was calculated from the residual liver volume measurement values ​​at each time phase and arranged by time phase to obtain the residual liver regeneration rate sequence. Input the list of primary tumor imaging features into the image stream of the dual-stream feature extraction module, and input the microvascular invasion status and residual liver regeneration rate sequence into the clinical stream of the dual-stream feature extraction module. After extracting features from the two data streams, output the dual-stream features. The dual-flow features are input into the inverse coupling cross-modal bridge module. The cross-modal resonance fusion mechanism multiplies the image flow features and clinical flow features dimension by dimension and injects them into the contralateral flow with inverse weights. After iterative coupling, the fused feature vector is output. The fused feature vector is mapped to a three-dimensional vector through a fully connected layer. After being normalized by Softmax, it is used as the probability of intrahepatic metastasis and the probability of new cancerous transformation. The third dimension is output as the prior vector of recurrence type.

[0009] Optionally, the operation of the longitudinal dynamics module includes: Images of each liver segment at each follow-up time point were read from the multi-temporal image library. The values ​​of fat deposition, tumor margin distance, blood supply distance and regeneration slope were extracted for each liver segment and each time phase. The images were arranged in chronological order of follow-up time to obtain the sequence of omics characteristic values ​​for each liver segment. The characteristic values ​​of the earliest follow-up time point of each liver segment were used as the initial characteristic values. The upper limit parameter and the growth rate parameter were set, and the Gompertz growth equation with time as the independent variable and characteristic values ​​as the dependent variable was constructed in the dynamic evolution module. The probability of intrahepatic metastasis and the probability of neonatal carcinogenesis in the prior vector of recurrence type are input into the Logistic chaotic mapping. The growth rate parameter is modulated phase by phase using the generated chaotic sequence. The modulated parameter is then injected into the Gompertz growth equation to obtain the modulated evolution equation. The modulated evolution equation is solved point by point along the follow-up time axis. The solution of the equation is taken as the expected characteristic value at each follow-up time point. The time required for the expected characteristic value to increase to twice the initial characteristic value is taken as the characteristic doubling time. The difference between the expected feature value and the measured feature value extracted from the measured image at each follow-up time point is calculated, and the values ​​are arranged in chronological order of follow-up time to generate a sequence of expected and measured deviations.

[0010] Optionally, the working process of the niche determination module includes: Read the feature rows corresponding to each liver segment in the liver segment-level microenvironment feature matrix, read the recurrence type prior vector and the expected feature value and expected and measured deviation sequence of each liver segment, and concatenate the three data of the same liver segment to obtain the fusion input vector of each liver segment. The fused input vectors of each liver segment are input into a fully connected layer. After weighted summation and nonlinear activation, a single numerical value is output. The Sigmoid function is used to map the numerical value to the interval between 0 and 1 to obtain the recurrence probability score of each liver segment. The recurrence probability score of each liver segment is compared with the preset high-risk threshold and intermediate-risk threshold. The liver segment with a recurrence probability score not lower than the high-risk threshold is labeled as high-risk, the liver segment with a recurrence probability score not lower than the intermediate-risk threshold but lower than the high-risk threshold is labeled as intermediate-risk, and the liver segment with a recurrence probability score lower than the intermediate-risk threshold is labeled as low-risk. The recurrence probability scores of each liver segment are filled into the corresponding mask area in the liver segment spatial mask map according to the liver segment number. The mask area is then colored according to the recurrence probability scores to generate a liver segment-level recurrence niche risk heatmap.

[0011] Optionally, the working process of the screening and scheduling module includes: The liver segment level labels output by the niche determination module are read in round by round. When there is a liver segment with a warning level label of high risk, the targeted re-examination process is triggered. Read the liver segment-level recurrence niche risk heatmap, select the liver segment with the highest recurrence probability score, and encapsulate the liver segment number, re-examination time phase, and imaging modality into a targeted imaging re-examination instruction; Compare the numerical values ​​of the probability of intrahepatic metastasis and the probability of neo-cancer transformation in the prior vector of recurrence type. When the probability of intrahepatic metastasis is dominant, prioritize the liver segment adjacent to the high-risk liver segment for re-examination. When the probability of neo-cancer transformation is dominant, prioritize the liver segment with the same blood supply as the high-risk liver segment for re-examination. Read the characteristic doubling time of high-risk liver segments, adjust the follow-up period to no more than one-third of the characteristic doubling time, the shorter the characteristic doubling time, the shorter the follow-up period, and generate a dynamic follow-up plan table that records the re-examination time points and re-examination contents of each liver segment; The dynamic follow-up plan is written back to the multi-temporal image database, and the next round of screening is triggered according to the most recent re-examination time point in the plan.

[0012] Optionally, the operation of the state initialization module and the extrapolation deviation calculation module in the longitudinal dynamics module includes: The state initialization module reads the omics feature value of the earliest follow-up time point of each liver segment as the initial feature value, reads the maximum value of the omics feature value of all follow-up time points of each liver segment, multiplies the maximum value by a preset multiple as the growth limit parameter, reads the omics feature values ​​of the first two follow-up time points, calculates the natural logarithm of the ratio of the two as the growth rate parameter, and writes the initial feature value, growth limit parameter and growth rate parameter into the Gompertz growth equation to complete the state initialization. The extrapolation deviation calculation module calculates the difference between the expected feature value and the measured feature value extracted from the measured image at each follow-up time point. The difference is divided by the measured feature value at the corresponding follow-up time point to obtain the relative deviation. The relative deviations at each follow-up time point are arranged in chronological order of follow-up time to generate a sequence of expected and measured deviations.

[0013] The beneficial effects of this invention are: This invention uses a data acquisition module with preoperative images as the registration benchmark. Through rigid registration, non-rigid registration, and layer-by-layer Z-Score standardization, it unifies the anatomical coordinates and gray dimensions of images at different time phases, generating a complete multi-temporal image library. A liver segment raster module extracts features of fat deposition, iron deposition, portal vein blood supply, tumor margin distance, blood supply distance, and regeneration slope segment by segment using the Couinaud eight-segment method, generating a liver segment-level microenvironment feature matrix and a liver segment spatial mask map carrying liver segment numbers. This expands the criteria for recurrence determination from visible nodules to the quantitative expression of microenvironment indicators. A recurrence type prior module deeply couples image flow and clinical flow features through dual-flow feature extraction and inverted coupling cross-modal bridge cross-modal resonance fusion, outputting the probabilities of intrahepatic metastatic and neo-cancer types, providing prior guidance for differentiated segment expansion. The longitudinal dynamics module constructs the Gompertz growth equation and modulates the growth rate parameter with a Logistic chaotic sequence to extrapolate the expected feature values ​​of each liver segment, quantifies the feature doubling time, and generates a sequence of expected and measured deviations, enabling a prospective assessment of the recurrence evolution rate and early capture of the "pre-image state." The niche determination module integrates the three data streams into a liver segment score and classifies them into high-risk, intermediate-risk, and low-risk levels, mapping them to a liver segment spatial mask to generate a recurrence niche risk heatmap, achieving a visual closed loop from numerical determination to anatomical localization of recurrence risk. The screening scheduling module selects adjacent liver segments or segments with homologous blood supply based on the direction of recurrence type dominance for targeted follow-up, adjusting the follow-up cycle to no more than one-third of the feature doubling time, and writing the dynamic follow-up plan back to the multi-temporal image library to drive the next round of screening. Ultimately, this achieves accurate perception of postoperative recurrence risk in liver cancer, differentiated discrimination of recurrence types, and adaptive closed-loop scheduling of follow-up resources, effectively improving the early recurrence detection rate, interpretation efficiency, and conversion rate to radical re-treatment. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the radiomics-based early screening system for liver cancer recurrence proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figure 1 The radiomics-based early screening system for liver cancer recurrence includes the following modules: The data acquisition module is used to collect multimodal medical data from patients and standardize the processing to generate a multi-temporal image library; The liver segment grid module is used to divide the liver according to the Couinaud eight-segment method, extract fat deposition, tumor margin distance, blood supply distance and regeneration slope features segment by segment, generate liver segment-level microenvironment feature matrix, and save the outline of each liver segment to obtain liver segment spatial mask map. The recurrence type prior module is used to extract the primary tumor image feature list, microvascular invasion status and residual liver regeneration rate sequence and input them into the dual-stream attention fusion network to obtain the recurrence type prior vector, and output the probability of intrahepatic metastasis and the probability of new cancer. The dual-stream attention fusion network includes a dual-stream feature extraction module and an inverse coupling cross-modal bridge module. The inverse coupling cross-modal bridge module embeds a cross-modal resonance fusion mechanism. The longitudinal dynamics module is used to extract feature values ​​segment by segment of the liver and input them into the improved Gompertz model to obtain feature doubling time and expected feature values. The improved Gompertz model includes a state initialization module, a dynamic evolution module, an inverse coupling prior injection module, and an extrapolation deviation calculation module. The inverse coupling prior injection module embeds a chaotic modulation prior mechanism and generates a sequence of measured deviations based on the measured images. The niche determination module is used to fuse the liver segment-level microenvironment feature matrix, recurrence type prior vector and expected feature value with the measured deviation sequence, calculate the recurrence probability score, map it to the liver segment spatial mask map to generate a niche risk heat map, and determine the warning level label for each liver segment. The screening scheduling module is used to generate re-examination instructions using risk heatmaps when the warning level label is high-risk. It selects the corresponding liver segment for re-examination based on the probability of intrahepatic metastasis and the probability of new cancerous transformation, adjusts the follow-up cycle based on the feature doubling time, generates a dynamic follow-up plan, and drives the next round of screening.

[0017] In this embodiment, the operation of the data acquisition module includes: Preoperative contrast-enhanced CT / MRI and multimodal medical data were collected from patients. The multimodal medical data included images at various time points during postoperative follow-up and clinical laboratory data. The images at each time point were collected according to the postoperative follow-up plan, covering follow-up points at 1, 3, and 6 months postoperatively, and every 3 to 6 months thereafter, to cover the peak window for liver cancer recurrence. The clinical laboratory data included pathology reports and tumor marker results such as alpha-fetoprotein, providing supplementary evidence beyond imaging for determining the type of recurrence.

[0018] Using the patient's preoperative enhanced CT / MRI as the registration benchmark, the images at each time phase were sequentially mapped to the same anatomical coordinate system through rigid and non-rigid registration to eliminate positional offsets between the images at different time phases. The rigid registration eliminated translational and rotational offsets caused by positioning differences, while the non-rigid registration used a B-spline free deformation model to correct liver deformation caused by respiratory movements, etc. The two-level registration was performed sequentially, first globally and then locally, to ensure precise alignment of the liver boundaries layer by layer.

[0019] After registration, the gray values ​​of each temporal image are mapped to a standard distribution with a mean of 0 and a standard deviation of 1 using the Z-Score normalization formula, thus unifying the gray dimensions of each temporal image. The layer-by-layer processing calculates the mean and standard deviation on a per-tomographic basis, eliminating differences in scanning dose and making images acquired at different times and with different equipment directly comparable.

[0020] The standardized images of each time phase are sorted and stored according to patient number and acquisition time, and together with the patient's preoperative enhanced CT / MRI, a multi-time phase image library is generated. The sorting and storage uses patient number as the primary index and acquisition time as the secondary index, so that the longitudinal features of any liver segment can be directly accessed according to the time axis, providing complete temporal data support for subsequent liver segment-by-liver modeling and extrapolation of recurrence trends.

[0021] In this embodiment, the operation of the liver segment grid module includes: Preoperative contrast-enhanced CT / MRI images of the liver from a multi-temporal image database were retrieved. Based on the Couinaud eight-segment method, the liver was divided into eight hepatic segments using the hepatic veins and portal vein branches as demarcation markers. The segments were demarcated longitudinally by the right hepatic vein, middle hepatic vein, and left hepatic vein, and laterally by the horizontal course of the left and right portal vein branches. The eight segments were numbered sequentially from S1 to S8. The boundaries were delineated and verified layer by layer, providing a unified spatial zoning benchmark for segmental feature measurements.

[0022] Eight liver segments were traversed. The signal intensity ratio of fat within each segment was measured as fat deposition, the magnetic susceptibility-sensitive signal intensity within each segment as iron deposition, the proportion of portal vein branch blood flow within each segment to the total liver blood flow as portal vein blood supply, the shortest distance between the liver segment edge and the primary tumor edge as tumor margin distance, and the shortest path length between the liver segment center and the main intrahepatic blood supply vessels as blood supply distance. Linear fitting was performed on the characteristic changes of each liver segment at different time points before and after surgery, and the fitting slope was taken as the regeneration slope feature. Specifically, fat deposition and iron deposition were measured within the liver segment using in-phase and out-of-phase sequences and magnetic susceptibility-weighted sequences to reflect the degree of local fatty degeneration and iron overload in the liver segment; portal vein blood supply reflected the difference in the proportion of blood supply sources in the liver segment; tumor margin distance and blood supply distance jointly characterized the spatial risk of liver segment invasion by the primary tumor and hematogenous dissemination; the regeneration slope feature was extracted by linear fitting to show the evolution rate of each liver segment characteristic at different time points, extending the static microenvironment measurement to a time-series dynamic feature.

[0023] The fat deposition, iron deposition, portal vein blood supply, tumor margin distance, blood supply distance, and regeneration slope characteristics of each liver segment are arranged into rows according to the liver segment number. The six characteristics are concatenated into a feature row in a fixed order. The feature rows of all liver segments are stacked vertically to generate a liver segment-level microenvironment feature matrix. The fixed order is preset in the system and is consistent throughout the entire process, so that the column position of the feature matrix corresponds one-to-one with the feature meaning. After vertical stacking, the rows of the matrix correspond to the liver segment numbers, which facilitates subsequent alignment and fusion with the prior vector of recurrence type and expected feature values ​​row by row according to the liver segment.

[0024] The outlines of eight liver segments are drawn layer by layer as binary masks, and the mask values ​​are labeled according to the liver segment numbers to generate a spatial mask map of the liver segments. The binary masks use the liver segment numbers as value labels, so that each pixel in the mask map carries the number information of the liver segment to which it belongs. The recurrence probability score calculated by the subsequent niche determination module can be directly filled into the corresponding mask area according to the number and color assigned to generate a liver segment-level recurrence niche risk heatmap, realizing the spatial positioning and visualization of recurrence risk.

[0025] In this embodiment, the operation of the recurrence type prior module includes: Images of the liver segment containing the primary tumor were extracted from a multi-temporal image database. The maximum diameter, marginal morphology, and arterial phase enhancement of the primary tumor were measured and summarized into a list of imaging features of the primary tumor. The maximum diameter was measured layer by layer in the portal venous phase, and the maximum value was taken. The marginal morphology was quantitatively coded into three categories: smooth, lobulated, and spiculated. The arterial phase enhancement was characterized by the difference in CT values ​​between the arterial and portal venous phases. These three features together characterize the invasive phenotype of the primary tumor, providing imaging evidence for the judgment of intrahepatic metastatic recurrence.

[0026] The microvascular invasion status recorded in the pathology report was read from clinical laboratory data. The volume change rate between adjacent time phases was calculated from the residual liver volume measurements at each time phase and arranged by time phase to obtain the residual liver regeneration rate sequence. The microvascular invasion status was encoded using positive and negative binary values ​​as a pathological marker of hematogenous tumor dissemination. The residual liver regeneration rate sequence quantifies the postoperative liver regeneration process into time-series data. Abnormal decay of the regeneration rate indicates abnormal intrahepatic microenvironment in the recurrence of neoplastic carcinoma, supplementing the deficiencies of imaging features from both pathological and functional dimensions.

[0027] The primary tumor image feature list is input into the image stream of the dual-stream feature extraction module, while the microvascular invasion status and residual liver regeneration rate sequence are input into the clinical stream of the same module. Features are extracted from each data stream separately, and the dual-stream features are output. The image stream uses a 3D convolutional kernel to extract the spatial morphological features of the primary tumor, while the clinical stream uses a multilayer perceptron to extract pathological and temporal features. The parameters of the two data streams are independent and do not interfere with each other, avoiding feature scale imbalance and gradient coupling contamination caused by direct stitching of heterogeneous data in the early stages.

[0028] The dual-stream features are input into the inverse coupling cross-modal bridge module. The cross-modal resonance fusion mechanism multiplies the image stream features and clinical stream features dimension by dimension and injects them into the opposite stream with inverse weights. After iterative coupling, a fused feature vector is output. The cross-modal resonance fusion mechanism first multiplies the two features dimension by dimension to obtain a resonance component that represents the consistency between the image features and clinical features. The stronger the resonance component, the more the image manifestation and pathology and regeneration status in that dimension are mutually corroborated, and the clearer the recurrence direction. Then, the resonated features are injected into the opposite stream with inverse weights. That is, the image stream features are injected into the clinical stream after inversion, and the clinical stream features are injected into the image stream after inversion. This allows each data stream to retain its own modality-dominant information while preferentially absorbing the discriminative components in the opposite stream that are complementary to the current response of the current stream. The iterative coupling is performed in two to three rounds. After each round of injection, layer normalization and gated residual updates are performed so that the two features are gradually aligned to the same discriminative space in multiple rounds of inversion interaction. Finally, a fused feature vector with deep intermodal coupling is output.

[0029] The fused feature vector is mapped to a three-dimensional vector through a fully connected layer. The first two dimensions are normalized using Softmax to represent the probabilities of intrahepatic metastasis and neoplastic cancer, respectively. The third dimension is output as a prior vector for recurrence type. This third dimension encodes the degree of mixed tendency beyond the two types of recurrence. Together with the probabilities of intrahepatic metastasis and neoplastic cancer, it constitutes the prior vector for recurrence type. This vector is used by the niche determination module to assign differentiated weights to different sources of recurrence when fusion is performed on liver segments. It is also used by the screening scheduling module to select the corresponding liver segment for re-examination based on the dominant probability direction when the warning level label is high-risk, thus achieving direct transmission of recurrence type discrimination results to screening actions.

[0030] In this embodiment, the operation of the longitudinal dynamics module includes: Images of each liver segment at each follow-up time point were read from a multi-temporal image library. Values ​​of fat deposition, tumor margin distance, blood supply distance, and regeneration slope were extracted segment by segment and phase by phase. These values ​​were then arranged chronologically according to follow-up time to obtain a sequence of omics characteristic values ​​for each liver segment. This segment-by-segment, phase-by-phase extraction ensured that each liver segment had an independent temporal observation trajectory. The four features characterized the evolutionary state of the liver segment microenvironment from four dimensions: degree of steatosis, risk of tumor margin invasion, hematogenous dissemination conditions, and liver regeneration process. After being arranged chronologically according to follow-up time, each sequence represents a discrete observation sample of the liver segment microenvironment on the time axis, providing a basis of experimental data for subsequent dynamic modeling.

[0031] The earliest follow-up time point of each liver segment was used as the initial characteristic value. An upper growth limit parameter and a growth rate parameter were set, and a Gompertz growth equation with time as the independent variable and characteristic values ​​as the dependent variable was constructed in the kinetic evolution module. The Gompertz growth equation was chosen instead of nonlinear extrapolation or a common exponential model because the growth of the liver segment microenvironment driven by tumor recurrence is not an infinite exponential expansion, but rather exhibits a rapid initial growth followed by a slower, asymptotic saturation characteristic, which naturally aligns with the mathematical property of the Gompertz equation that "the growth rate decreases proportionally with the increase of the characteristic value." The initial characteristic value was taken as the measured value of the earliest follow-up time point, anchoring the extrapolation starting point to actual observations and avoiding cumulative drift caused by purely fitting the starting point.

[0032] The probability of intrahepatic metastasis and the probability of neoplastic carcinogenesis in the prior vector of recurrence type are input into the Logistic chaotic mapping. The generated chaotic sequence modulates the growth rate parameter phase by phase, and the modulated parameter is injected into the Gompertz growth equation to obtain the modulated evolution equation. This is the core innovation of this module: the recurrence process itself has nonlinear evolutionary characteristics and is sensitive to initial conditions. A fixed growth rate parameter cannot characterize the phase fluctuations of the recurrence driving intensity. The Logistic chaotic mapping uses two types of recurrence probabilities as input drivers to generate a non-periodic sequence in the chaotic interval. The growth rate parameter is transformed from a constant to a dynamic quantity that fluctuates with time phases. This allows the evolution equation to follow the Gompertz saturated growth law as a whole, while superimposing local perturbations determined by the recurrence type probability. When the probability of intrahepatic metastasis is dominant, the perturbation emphasizes the accelerated growth of tumor margin and blood supply-related characteristics. When the probability of neoplastic carcinogenesis is dominant, the perturbation emphasizes the abnormal rise of fat deposition and regeneration slope-related characteristics. Thus, the recurrence type prior is truly integrated into every step of the extrapolation process, rather than just weighted at the end.

[0033] The modulated evolution equation was solved point-by-point along the follow-up time axis. The solutions were used as the expected characteristic values ​​at each follow-up time point, and the time required for the expected characteristic value to double to the initial characteristic value was defined as the characteristic doubling time. This characteristic doubling time is an equivalent time constant characterizing the relapse rate, compressing the entire evolution curve into a single comparable indicator. A shorter doubling time indicates a faster rate of evolution of the liver segment's microenvironment towards a relapse state, providing a direct and usable quantitative basis for the screening scheduling module to set the follow-up cycle.

[0034] The difference between the expected feature value and the measured feature value extracted from the actual image at each follow-up time point is calculated and arranged in chronological order of follow-up time to generate a sequence of expected and measured deviations. This deviation sequence serves as a direct bridge between the model's predictive ability and the measured data: if the deviation of a certain liver segment continuously increases over time, it indicates that the measured microenvironment evolution has deviated from the conventional recurrence trajectory, suggesting that there is abnormal proliferative activity in that liver segment that is underestimated by the model, representing the earliest and most sensitive warning signal of recurrence. The deviation, along with the expected feature value and the feature doubling time, is output to the niche determination module, where it is fused with the liver segment-level microenvironment feature matrix and the recurrence type prior vector in a three-way manner according to the liver segment. This achieves temporal quantitative perception of recurrence trends and segment-by-segment risk scoring, upgrading recurrence warning from static threshold judgment to a prospective discrimination at the dynamic evolution level.

[0035] In this embodiment, the working process of the niche determination module includes: The feature rows corresponding to each liver segment in the liver segment-level microenvironment feature matrix are read, along with the prior vector of recurrence type and the expected feature values ​​and expected / measured deviation sequences for each liver segment. The three data streams for the same liver segment are then concatenated end-to-end to obtain the fused input vector for each segment. These three data streams respectively carry the static phenotype of the liver segment microenvironment, the prior indication of the recurrence type, and the temporal prediction of the evolutionary trend—the feature rows depict the current microenvironment state of the liver segment, the prior vector of the recurrence type indicates whether the risk originates from intrahepatic metastasis or neoplastic carcinogenesis, and the expected feature values ​​and deviation sequences provide the expected path of future evolution and the measured degree of deviation. The end-to-end concatenation allows the three information streams to be fully aggregated at the same liver segment level, ensuring that the risk assessment input simultaneously covers the "present, source, and future" levels, avoiding the one-sidedness of a single-dimensional scoring method.

[0036] The fused input vectors of each liver segment are input into a fully connected layer. After weighted summation and nonlinear activation, a single numerical value is output. The Sigmoid function is used to map this value to the interval between 0 and 1, obtaining the recurrence probability score for each liver segment. The weights of the fully connected layer are obtained during the training phase using the random access questionnaire labels for postoperative recurrence as supervision, automatically allocating the discriminative contributions of three types of heterogeneous information—static microenvironment, prior recurrence type, and dynamic bias—in a unified weight space. The Sigmoid function mapping gives the output a clear probabilistic semantic; the closer the score is to 1, the closer the microenvironment combination of that liver segment is to the known pre-recurrence state. Furthermore, the scores of each liver segment can be directly compared horizontally under the same dimension, providing a unified benchmark for subsequent threshold determination and heatmap colorization.

[0037] The recurrence probability score of each liver segment is compared with preset high-risk and intermediate-risk thresholds. Liver segments with a recurrence probability score not lower than the high-risk threshold are labeled as high-risk, those not lower than the intermediate-risk threshold but lower than the high-risk threshold are labeled as intermediate-risk, and those lower than the intermediate-risk threshold are labeled as low-risk. This threshold comparison converts continuous probability into discrete levels. The high-risk, intermediate-risk, and low-risk classification directly corresponds to clinical follow-up decisions—high-risk liver segments trigger targeted re-examination and shorten the follow-up period, intermediate-risk liver segments are included in routine follow-up observation, and low-risk liver segments maintain the standard period. This transforms continuous risk measurement into an actionable tiered treatment basis, avoiding the subjectivity and omissions of manual segment-by-segment assessment.

[0038] The recurrence probability scores for each liver segment are filled into the corresponding mask areas in the liver segment spatial mask image according to the liver segment number. The mask areas are then colored according to the recurrence probability scores to generate a liver segment-level recurrence niche risk heatmap. The filling process utilizes the liver segment number information carried by each pixel in the liver segment spatial mask image to restore the abstract score values ​​to their anatomical spatial locations. This allows the recurrence risk of each liver segment to be directly presented on the liver image with a color gradient. The viewer can intuitively locate the liver segment with the highest risk and its spatial relationship with the primary tumor and major blood vessels by following the heatmap. This provides a spatial reference for delineating the scope of follow-up examinations and selecting target areas for subsequent punctures, interventions, and other treatments, realizing a visual closed loop from numerical determination of recurrence risk to anatomical location.

[0039] In this embodiment, the working process of the screening and scheduling module includes: The module sequentially reads the liver segment level labels output by the niche determination module. When a liver segment with a warning level label of "high risk" is found, a targeted re-examination process is triggered. This sequential reading makes this module the scheduling center of the entire process, triggering targeted re-examination only when a high-risk liver segment is present. This avoids the waste of resources caused by uniform and dense follow-up of the entire liver, and ensures that re-examination resources are accurately directed to the liver segment region with the highest risk.

[0040] The liver segment recurrence niche risk heatmap is read, and the liver segment with the highest recurrence probability score is selected. The segment number, follow-up examination phase, and imaging modality are packaged into a targeted imaging follow-up instruction. The follow-up examination phase and imaging modality are set according to the risk source of the liver segment. When blood supply-related features are dominant, the arterial and portal venous phases are selected. When fat and iron deposition features are dominant, in-phase and out-of-phase and magnetic susceptibility-weighted sequences are selected. This ensures that the images obtained from the follow-up examination are directly matched with the risk source, improving the detection efficiency of a single follow-up examination.

[0041] The module compares the probabilities of intrahepatic metastasis and neoplastic malignancy in the prior vector of recurrence type. When the probability of intrahepatic metastasis is dominant, the liver segment adjacent to the high-risk segment is prioritized for re-examination. When the probability of neoplastic malignancy is dominant, the liver segment with the same blood supply as the high-risk segment is prioritized for re-examination. This is the core innovation of this module: intrahepatic metastasis spreads to adjacent liver segments via portal vein tumor thrombus or direct invasion, and the spread follows the principle of spatial adjacency; neoplastic malignancy occurs along a multicentric distribution of abnormal microenvironment, with lesions sharing the same portal vein blood supply, and the spread follows the principle of homologous blood supply. The comparison mechanism transforms the prior of recurrence type into a differentiated segment expansion strategy, making the re-examination scope closely match the actual dissemination path, ensuring that the most likely affected liver segment is not missed, while controlling the re-examination scope to the minimum necessary set.

[0042] The characteristic doubling time of high-risk liver segments is read, and the follow-up period is adjusted to no more than one-third of the characteristic doubling time. The shorter the characteristic doubling time, the shorter the follow-up period. A dynamic follow-up plan is generated, recording the re-examination time points and re-examination contents for each liver segment. The one-third ratio ensures that at least two observation opportunities are obtained before the tumor volume doubles. The follow-up period dynamically expands and contracts with the doubling time, replacing the shortcomings of traditional fixed-interval follow-up, which "cannot detect early recurrence and detects too frequently for slow recurrence," and achieving automatic matching between follow-up intensity and recurrence speed.

[0043] The dynamic follow-up plan is written back to the multi-temporal image database, triggering the next round of screening based on the most recent follow-up examination time point in the plan. This write-back process ensures that the images generated during follow-up examinations are categorized into the image database by patient number and acquisition time, and then re-enter the complete judgment process, forming a closed loop of "screening—judgment—targeted distribution—follow-up write-back—rescreening." This allows the early warning sensitivity to cumulatively increase with each follow-up round, achieving full-cycle dynamic adaptive monitoring of recurrence risk.

[0044] In this embodiment, the working process of the state initialization module and the extrapolation deviation calculation module in the longitudinal dynamics module includes: The state initialization module multiplies the maximum value of the omics feature values ​​at all follow-up time points of each liver segment by a preset multiple to obtain the upper limit of growth parameter; it reads the omics feature values ​​at the first two follow-up time points and calculates the natural logarithm of the ratio between the two to obtain the growth rate parameter; the preset multiple is pre-set in the system, with a value range of 1.2 to 1.5, which both reserves extrapolation space for feature growth and prevents the upper limit of saturation from deviating from the reasonable magnitude of anatomy and physiology; the growth rate parameter substitutes the instantaneous growth rate between the earliest two points into the equation so that the slope of the tangent at the starting point of the evolution curve is consistent with the evolution rate of the measured sequence; once the three parameters are initialized, they are written into the Gompertz growth equation and locked, and subsequent modulation only acts on the temporal perturbation term of the growth rate parameter without changing the initial anchor value, thus ensuring the stability of the extrapolation benchmark.

[0045] The extrapolation deviation calculation module divides the difference at each follow-up time point by the measured feature value at the corresponding time point to obtain a dimensionless relative deviation, which is then arranged chronologically to generate a sequence of expected and measured deviations. The normalization eliminates the differences in dimensions and numerical scales of heterogeneous features such as fat deposition, tumor margin distance, and blood supply distance, ensuring that the deviations of each feature fall within the same comparable range. This allows for direct row-by-row splicing and fusion with the liver segment-level microenvironment feature matrix and the prior vector of recurrence type. The absolute value of the relative deviation is used in the risk scoring of the niche determination module, while the original sequence with the sign retained is output separately to trace the direction of the liver segment's evolutionary deviation—positive deviation indicates that the measured deviation is higher than expected, suggesting accelerated abnormal proliferation, while negative deviation indicates that the measured deviation is lower than expected, suggesting a stable recovery process. The distinction of direction makes the warning signal interpretable in addition to its sensitivity.

[0046] Example 1: To verify the early warning capability of this invention in the scenario of monitoring postoperative recurrence of liver cancer, the liver cancer postoperative recurrence monitoring system based on multi-temporal image dynamics modeling proposed in this invention was deployed in the hepatobiliary surgery center of a tertiary hospital in East my country. This center performs approximately 400 radical resections of liver cancer annually, with over 1500 patients under follow-up management. Follow-up imaging includes both enhanced CT and enhanced MRI modalities, with follow-up intervals of 1, 3, and 6 months postoperatively, and then every 3 to 6 months thereafter. In actual monitoring, because the recurrence rate of liver cancer within two years postoperatively is as high as 40% to 60%, and the dissemination pathways of intrahepatic metastasis and neoplastic carcinoma are drastically different, the traditional fixed-interval follow-up model struggles to balance detection sensitivity and medical resource consumption, resulting in generally late recurrence detection and missed opportunities for radical re-resection. Previously, the center mainly relied on radiologists to compare images before and after each case according to a fixed follow-up cycle, supplemented by screening with tumor markers such as alpha-fetoprotein. However, when faced with a large number of follow-up cases and numerous imaging phases, problems such as missed detection of small recurrences, mismatch between follow-up intervals and tumor doubling rate, and indiscriminate application of re-examination scope often occurred, which seriously restricted the early diagnosis rate of recurrence and the conversion rate of retreatment.

[0047] The center generates a massive amount of enhanced CT / MRI images, pathology reports, and tumor marker test data at various time points daily, with an average of over 60 new follow-up imaging examinations per day and over 300 slices per individual imaging session. Traditional follow-up models often rely solely on visually visible nodules or significant increases in marker values ​​on imaging, lacking quantitative extraction of liver segment-level microenvironment characteristics such as fat deposition, iron deposition, portal vein blood supply, and tumor margin distance. Furthermore, they cannot characterize the longitudinal evolution trend of each liver segment's characteristics across follow-up phases, leading to the underestimation of many early recurrent liver segments that are still in a "pre-imaging state." Especially when the recurrence type is unclear, the current model lacks the ability to differentiate between intrahepatic metastasis and new cancerous changes, forcing the follow-up scope to be expanded to the entire liver based on experience. This increases radiation exposure and contrast agent burden, and lengthens the waiting period for each round of follow-up examinations.

[0048] In practical application of this invention, the system first reads preoperative enhanced CT / MRI and follow-up phase images and clinical test data from the central image archiving system. After rigid registration, non-rigid registration, and layer-by-layer Z-Score standardization, a multi-temporal image library is constructed. Subsequently, the liver segments are divided according to the Couinaud eight-segment method, and features such as fat deposition, iron deposition, portal vein blood supply, tumor margin distance, blood supply distance, and regeneration slope are extracted to generate a liver segment-level microenvironment feature matrix. The recurrence type prior module fuses image flow and clinical flow features through dual-flow feature extraction and inverse coupling cross-modal bridge to output the probability discrimination between intrahepatic metastatic and neoplastic carcinoma types. The longitudinal dynamics module constructs the Gompertz growth equation and modulates the growth rate parameter with a Logistic chaotic sequence to extrapolate the expected feature values ​​of each liver segment and generate a sequence of expected and measured deviations. The niche determination module fuses and scores the three data streams and classifies them into high-risk, intermediate-risk, and low-risk categories to generate a liver segment-level recurrence niche risk heatmap. The screening and scheduling module encapsulates targeted imaging re-examination instructions for high-risk liver segments, selects adjacent liver segments or liver segments with the same blood supply as the dominant recurrence type, and issues re-examination instructions. The follow-up cycle is dynamically compressed by one-third of the feature doubling time, and the re-examination results are written back to the multi-temporal image database to enter the next closed loop.

[0049] The table below details the comparison data of the monitoring system proposed in this invention and the fixed-interval follow-up mode originally used by the center during a 12-month control follow-up period, focusing on key performance indicators. Data collection covered 1200 patients under management and approximately 6800 follow-up imaging sessions. Comparison dimensions included early recurrence detection rate, recurrence detection advance time, small lesion missed detection rate, range of single-round follow-up, average number of follow-up sessions per person per year, radical treatment conversion rate, and time spent interpreting a single case. Table 1. Performance comparison data of the present invention and the traditional fixed-interval follow-up mode in postoperative recurrence monitoring of liver cancer. Based on the comparative data shown in Table 1, it can be seen that the monitoring system proposed in this invention has significant performance advantages over the traditional fixed-interval follow-up mode in monitoring postoperative recurrence of liver cancer, especially in key indicators such as early recurrence detection, early detection time, accuracy of follow-up examination, and conversion of relapse into treatment.

[0050] In terms of early recurrence detection and control of missed lesions, this invention expands the criteria for judgment from visible nodules to abnormal evolution of microenvironment indicators such as fat deposition and portal vein blood supply by using a liver segment-level microenvironment feature matrix and longitudinal dynamic extrapolation. This increases the detection rate of early lesions (less than 1 cm) within one year post-surgery from 41.6% to 82.3%, and reduces the missed detection rate of small lesions from 18.7% to 5.2%. For example, in a case where no clear nodule was seen on imaging during a 6-month follow-up, the expected-to-measured deviation of the tumor margin distance in segment S8 continuously increased for two consecutive time phases. The risk heatmap identified this segment as high-risk, and a follow-up directional MRI detected a 0.6 cm small recurrence lesion near the portal vein in segment S8, confirming the ability of the dynamic deviation signal to capture the "pre-imaging state."

[0051] Regarding the early detection time of recurrence, this invention relies on Gompertz equation extrapolation and characteristic doubling time quantification to achieve a prospective assessment of the recurrence evolution rate. The early detection time for intrahepatic metastatic cases was extended from 2.8 months to 5.6 months, and for neo-cancer cases from 3.5 months to 6.9 months. For example, in a case where the recurrence type was predominantly neo-cancer, the abnormal elevation of multicentric hepatic segmental steatosis and regeneration slope was captured earlier by the chaotically modulated evolution equation, allowing for targeted follow-up nearly 7 months earlier than the significant elevation of biomarkers, thus providing a window for radical radiofrequency ablation.

[0052] Regarding the accuracy of follow-up examinations, this invention differentiates the coverage of liver segments based on the dominant recurrence type—for intrahepatic metastases, along adjacent liver segments; for neoplastic cancers, along segments with the same blood supply—reducing the average number of liver segments covered per round of follow-up for high-risk cases from 8 segments of the entire liver to 3.4 segments. This reduces radiation exposure and contrast agent burden by more than half while ensuring no missed detections. Simultaneously, the follow-up period dynamically adjusts according to one-third of the characteristic doubling time, increasing the annual follow-up frequency for high-risk cases to 5.5 times, while only increasing the average number of follow-ups per case by 1.2 times. This achieves automatic matching of monitoring intensity and recurrence rate, rather than uniform densification across the entire population.

[0053] In terms of interpretation efficiency and treatment conversion, this invention uses a risk heatmap to directly direct the attention of the patient to the high-risk liver segment, reducing the interpretation time per case from 26.4 minutes to 7.8 minutes; the interval from detection to treatment initiation is reduced from 21.3 days to 9.6 days; and the conversion rate to radical treatment is increased from 33.8% to 61.5%, significantly expanding the proportion of cases that have the opportunity for radical re-resection or ablation.

[0054] Overall, this invention successfully addresses the problems of missing microenvironmental features, indiscriminate recurrence type differentiation, one-size-fits-all follow-up intensity, and blindly expanding the scope of follow-up examinations in existing technologies by constructing a multi-temporal image library, extracting liver segment-level microenvironment features, identifying recurrence types through dual-stream inversion coupling, extrapolating Gompertz chaotic modulation dynamics, and establishing a closed loop for niche risk scoring and targeted distribution. Particularly in terms of early lesion capture and precise allocation of follow-up resources, it demonstrates a significant improvement in sensitivity and timeliness, filling the technical gaps in the traditional fixed-interval follow-up model for dynamic monitoring of postoperative recurrence in liver cancer.

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A radiomics-based early screening system for liver cancer recurrence, characterized in that, Includes the following modules: The data acquisition module is used to collect multimodal medical data from patients and standardize the processing to generate a multi-temporal image library; The liver segment grid module is used to divide the liver according to the Couinaud eight-segment method, extract fat deposition, tumor margin distance, blood supply distance and regeneration slope features segment by segment, generate liver segment-level microenvironment feature matrix, and save the outline of each liver segment to obtain liver segment spatial mask map. The recurrence type prior module is used to extract the primary tumor image feature list, microvascular invasion status and residual liver regeneration rate sequence and input them into the dual-stream attention fusion network to obtain the recurrence type prior vector, and output the probability of intrahepatic metastasis and the probability of new cancer. The dual-stream attention fusion network includes a dual-stream feature extraction module and an inverse coupling cross-modal bridge module. The inverse coupling cross-modal bridge module embeds a cross-modal resonance fusion mechanism. The longitudinal dynamics module is used to extract feature values ​​segment by segment of the liver and input them into the improved Gompertz model to obtain feature doubling time and expected feature values. The improved Gompertz model includes a state initialization module, a dynamic evolution module, an inverse coupling prior injection module, and an extrapolation deviation calculation module. The inverse coupling prior injection module embeds a chaotic modulation prior mechanism and generates a sequence of measured deviations based on the measured images. The niche determination module is used to fuse the liver segment-level microenvironment feature matrix, recurrence type prior vector and expected feature value with the measured deviation sequence, calculate the recurrence probability score, map it to the liver segment spatial mask map to generate a niche risk heat map, and determine the warning level label for each liver segment. The screening scheduling module is used to generate re-examination instructions using risk heatmaps when the warning level label is high-risk. It selects the corresponding liver segment for re-examination based on the probability of intrahepatic metastasis and the probability of new cancerous transformation, adjusts the follow-up cycle based on the feature doubling time, generates a dynamic follow-up plan, and drives the next round of screening.

2. The radiomics-based early screening system for liver cancer recurrence according to claim 1, characterized in that, The working process of the data acquisition module includes: Preoperative enhanced CT / MRI and multimodal medical data were collected from the patient, including postoperative follow-up images at various time points and clinical laboratory data. Using the patient’s preoperative enhanced CT / MRI as the registration benchmark, the images of each time phase are sequentially mapped to the same anatomical coordinate system through rigid registration and non-rigid registration to eliminate the positional offset between the images of each time phase. After registration, the gray values ​​of each temporal image are mapped to a standard distribution with a mean of 0 and a standard deviation of 1 by applying the Z-Score normalization formula to each layer, thereby unifying the gray dimension of each temporal image. The standardized images of each time phase are sorted and stored according to the patient number and the acquisition time, and together with the patient's preoperative enhanced CT / MRI, a multi-time phase image library is generated.

3. The radiomics-based early screening system for liver cancer recurrence according to claim 1, characterized in that, The working process of the liver segment grid module includes: Read the liver images of the patient’s preoperative enhanced CT / MRI from the multi-temporal image library, and divide the liver into eight liver segments according to the Couinaud eight-segment method, with the hepatic vein and portal vein branches as the boundary markers; Eight liver segments were traversed. The signal intensity ratio of fat within each segment was measured as fat deposition, the magnetic susceptibility-sensitive signal intensity within each segment was measured as iron deposition, the proportion of portal vein branch blood flow within each segment to the total blood flow of the liver was measured as portal vein blood supply, the shortest distance between the edge of the liver segment and the edge of the primary tumor was measured as the tumor margin distance, and the shortest path length between the center of the liver segment and the main blood supply vessels within the liver was measured as the blood supply distance. Linear fitting was performed on the characteristic changes of each liver segment at different time points before and after surgery, and the fitting slope was taken as the regeneration slope characteristic. The fat deposition, iron deposition, portal vein blood supply, tumor margin distance, blood supply distance and regeneration slope characteristics of each liver segment are arranged into rows according to the liver segment number. The six characteristics are spliced ​​into a feature row in a fixed order. The feature rows of all liver segments are stacked vertically to generate a liver segment-level microenvironment feature matrix. The outlines of the eight liver segments are drawn layer by layer as binary masks, and the mask values ​​are labeled according to the liver segment number to generate a spatial mask map of the liver segments.

4. The radiomics-based early screening system for liver cancer recurrence according to claim 1, characterized in that, The working process of the recurrence type prior module includes: Images of the liver segment containing the primary tumor were extracted from a multi-temporal image database. The maximum diameter, marginal morphology, and arterial phase enhancement of the primary tumor were measured and summarized into a list of imaging features of the primary tumor. The microvascular invasion status was read from the pathology report recorded in the clinical test data. The volume change rate between adjacent time phases was calculated from the residual liver volume measurement values ​​at each time phase and arranged by time phase to obtain the residual liver regeneration rate sequence. Input the list of primary tumor imaging features into the image stream of the dual-stream feature extraction module, and input the microvascular invasion status and residual liver regeneration rate sequence into the clinical stream of the dual-stream feature extraction module. After extracting features from the two data streams, output the dual-stream features. The dual-flow features are input into the inverse coupling cross-modal bridge module. The cross-modal resonance fusion mechanism multiplies the image flow features and clinical flow features dimension by dimension and injects them into the contralateral flow with inverse weights. After iterative coupling, the fused feature vector is output. The fused feature vector is mapped to a three-dimensional vector through a fully connected layer. After being normalized by Softmax, it is used as the probability of intrahepatic metastasis and the probability of new cancerous transformation. The third dimension is output as the prior vector of recurrence type.

5. The radiomics-based early screening system for liver cancer recurrence according to claim 1, characterized in that, The operation of the longitudinal dynamics module includes: Images of each liver segment at each follow-up time point were read from the multi-temporal image library. The values ​​of fat deposition, tumor margin distance, blood supply distance and regeneration slope were extracted for each liver segment and each time phase. The images were arranged in chronological order of follow-up time to obtain the sequence of omics characteristic values ​​for each liver segment. The characteristic values ​​of the earliest follow-up time point of each liver segment were used as the initial characteristic values. The upper limit parameter and the growth rate parameter were set, and the Gompertz growth equation with time as the independent variable and characteristic values ​​as the dependent variable was constructed in the dynamic evolution module. The probability of intrahepatic metastasis and the probability of neonatal carcinogenesis in the prior vector of recurrence type are input into the Logistic chaotic mapping. The growth rate parameter is modulated phase by phase using the generated chaotic sequence. The modulated parameter is then injected into the Gompertz growth equation to obtain the modulated evolution equation. The modulated evolution equation is solved point by point along the follow-up time axis. The solution of the equation is taken as the expected characteristic value at each follow-up time point. The time required for the expected characteristic value to increase to twice the initial characteristic value is taken as the characteristic doubling time. The difference between the expected feature value and the measured feature value extracted from the measured image at each follow-up time point is calculated, and the values ​​are arranged in chronological order of follow-up time to generate a sequence of measured deviations.

6. The radiomics-based early screening system for liver cancer recurrence according to claim 1, characterized in that, The working process of the niche determination module includes: Read the feature rows corresponding to each liver segment in the liver segment-level microenvironment feature matrix, read the recurrence type prior vector and the expected feature value and expected and measured deviation sequence of each liver segment, and concatenate the three data of the same liver segment to obtain the fusion input vector of each liver segment. The fused input vectors of each liver segment are input into a fully connected layer. After weighted summation and nonlinear activation, a single numerical value is output. The Sigmoid function is used to map the numerical value to the interval between 0 and 1 to obtain the recurrence probability score of each liver segment. The recurrence probability score of each liver segment is compared with the preset high-risk threshold and intermediate-risk threshold. The liver segment with a recurrence probability score not lower than the high-risk threshold is labeled as high-risk, the liver segment with a recurrence probability score not lower than the intermediate-risk threshold but lower than the high-risk threshold is labeled as intermediate-risk, and the liver segment with a recurrence probability score lower than the intermediate-risk threshold is labeled as low-risk. The recurrence probability scores of each liver segment are filled into the corresponding mask area in the liver segment spatial mask map according to the liver segment number. The mask area is then colored according to the recurrence probability scores to generate a liver segment-level recurrence niche risk heatmap.

7. The radiomics-based early screening system for liver cancer recurrence according to claim 1, characterized in that, The working process of the screening and scheduling module includes: The liver segment level labels output by the niche determination module are read in round by round. When there is a liver segment with a warning level label of high risk, the targeted re-examination process is triggered. Read the liver segment-level recurrence niche risk heatmap, select the liver segment with the highest recurrence probability score, and encapsulate the liver segment number, re-examination time phase, and imaging modality into a targeted imaging re-examination instruction; Compare the numerical values ​​of the probability of intrahepatic metastasis and the probability of neo-cancer transformation in the prior vector of recurrence type. When the probability of intrahepatic metastasis is dominant, prioritize the liver segment adjacent to the high-risk liver segment for re-examination. When the probability of neo-cancer transformation is dominant, prioritize the liver segment with the same blood supply as the high-risk liver segment for re-examination. Read the characteristic doubling time of high-risk liver segments, adjust the follow-up period to no more than one-third of the characteristic doubling time, the shorter the characteristic doubling time, the shorter the follow-up period, and generate a dynamic follow-up plan table that records the re-examination time points and re-examination contents of each liver segment; The dynamic follow-up plan is written back to the multi-temporal image database, and the next round of screening is triggered according to the most recent re-examination time point in the plan.

8. The radiomics-based early screening system for liver cancer recurrence according to claim 5, characterized in that, The working process of the state initialization module and the extrapolation deviation calculation module in the longitudinal dynamics module includes: The state initialization module reads the omics feature value of the earliest follow-up time point of each liver segment as the initial feature value, reads the maximum value of the omics feature value of all follow-up time points of each liver segment, multiplies the maximum value by a preset multiple as the growth limit parameter, reads the omics feature values ​​of the first two follow-up time points, calculates the natural logarithm of the ratio of the two as the growth rate parameter, and writes the initial feature value, growth limit parameter and growth rate parameter into the Gompertz growth equation to complete the state initialization. The extrapolation deviation calculation module calculates the difference between the expected feature value and the measured feature value extracted from the measured image at each follow-up time point. The difference is divided by the measured feature value at the corresponding follow-up time point to obtain the relative deviation. The relative deviations at each follow-up time point are arranged in chronological order of follow-up time to generate a sequence of expected and measured deviations.

Images