Lung cancer type prediction method and system based on fusion deep learning network

Abstract

The invention provides a lung cancer type prediction method and system based on a fusion deep learning network, and the method comprises the steps: carrying out the dynamic segmentation of a CT image of a lung cancer patient, extracting tumor blood vessel distribution data, calculating a three-dimensional density parameter and a blood vessel fracture direction parameter, and synchronously obtaining a necrotic region volume ratio sequence; circulating tumor cells are separated through blood sample spiral sorting and fluid control, and rigidity change parameters and deformation recovery parameters derived from a deformation track are generated; combining the three-dimensional density, the blood vessel fracture direction and the rigidity parameters, extracting the time sequence offset of the blood vessel fracture direction, and dynamically and synchronously matching the rigidity change response to generate a synchronous set; establishing lag correlation between the blood vessel fracture direction offset and the deformation recovery period, and determining weight matching through cross-modal analysis; and outputting a lung cancer subtype classification result in combination with the necrotic region volume spatial-temporal characteristics. According to the application, the cross-modal correlation model of the blood vessel dynamic migration and the cell mechanical response is fused, and accurate discrimination of the lung cancer subtype is realized.

Description

Technical Field

[0001] This application relates to the technical field of lung cancer type prediction, and particularly to a method and system for predicting lung cancer types based on a fused deep learning network. Background Art

[0002] Lung cancer is one of the malignant tumors with the highest incidence and mortality rates globally. Its early diagnosis and accurate classification are crucial for the treatment and prognosis of patients. Traditional lung cancer type diagnosis mainly relies on pathological examinations, but this method has problems such as strong subjectivity, long time consumption, and limited ability to identify early lesions. With the rapid development of medical imaging technology and artificial intelligence, lung cancer type prediction methods based on deep learning have become a research hotspot. Such technologies can automatically extract features from medical images to assist doctors in making rapid and accurate diagnoses, thus meeting the clinical demand for efficient and intelligent diagnostic tools.

[0003] Currently, lung cancer type prediction methods based on deep learning mainly use single network models, such as convolutional neural networks, recurrent neural networks, or graph convolutional networks, etc. These methods classify lung cancer types by extracting spatial or temporal features from medical images. For example, convolutional neural networks are widely used in the detection and classification of lung cancer in CT images, and they can capture local features of images; recurrent neural networks are used to process dynamic imaging data and extract time series information.

[0004] However, although existing methods have made certain progress in lung cancer type prediction, there are still the following problems: Single network models often can only capture a certain type of features (such as spatial features or temporal features), and it is difficult to comprehensively reflect the multi-dimensional information of lung cancer, resulting in limited classification accuracy. Secondly, simple model combination methods lack effective feature fusion mechanisms, are prone to introducing redundant information or ignoring key features, and affect model performance. Summary of the Invention

[0005] This application provides a method and system for predicting lung cancer types based on a fused deep learning network to solve the problem of low accuracy caused by insufficient fusion of multi-modal dynamic features in the existing lung cancer typing.

[0006] In a first aspect, this application provides a method for predicting lung cancer types based on a fused deep learning network, including: Obtain the CT images and blood samples of lung cancer patients, perform dynamic segmentation on the CT images to extract the vascular distribution data of the tumor region, calculate the three-dimensional density parameter and the vascular fracture direction parameter based on the vascular distribution data, and obtain the volume ratio sequence of the tumor necrosis region; Perform spiral sorting on the blood sample, separate circulating tumor cells through fluid control, collect the deformation trajectory data of the circulating tumor cells, and generate a stiffness change parameter and a deformation recovery parameter based on the deformation trajectory data; Jointly analyze the three-dimensional density parameter, the blood vessel rupture direction parameter, and the stiffness change parameter, extract the time series offset of the blood vessel rupture direction parameter, and synchronously match the time series offset with the dynamic response of the stiffness change parameter to generate a synchronous parameter set; Based on the synchronous parameter set, establish a lag correlation relationship between the time series offset of the blood vessel rupture direction parameter and the recovery period of the deformation recovery parameter, so as to determine the weight matching result of the two through cross-modal analysis; Output the lung cancer subtype classification result according to the weight matching result and the spatio-temporal distribution characteristics of the volume ratio of the tumor necrosis area.

[0007] Optionally, perform dynamic segmentation on the CT image to extract the blood vessel distribution data of the tumor area, calculate the three-dimensional density parameter and the blood vessel rupture direction parameter based on the blood vessel distribution data, and obtain the volume ratio sequence of the tumor necrosis area, including: Perform multi-time point segmentation operations on the CT image, and extract the blood vessel distribution data of the tumor area at each time point. The blood vessel distribution data includes the three-dimensional coordinates of the blood vessel branch nodes and the connection status between adjacent nodes; Based on the three-dimensional coordinates of the blood vessel branch nodes, calculate the density distribution of the blood vessels in three-dimensional space, and define the number of blood vessel branch nodes per unit volume as the three-dimensional density parameter; Statistically analyze the rupture directions of all blood vessel branch nodes, and generate a blood vessel rupture direction parameter according to the angular distribution of the rupture directions in the three-dimensional coordinate system; Generate a volume ratio sequence by measuring the proportion of the volume of the tumor necrosis area to the total volume of the tumor area at each time point.

[0008] Optionally, perform spiral sorting on the blood sample, separate circulating tumor cells through fluid control, collect the deformation trajectory data of the circulating tumor cells, and generate a stiffness change parameter and a deformation recovery parameter, including: Inject the blood sample into a spiral sorting device, and separate circulating tumor cells from non-tumor cells by adjusting the fluid dynamic direction; Collect the movement trajectory data of the separated circulating tumor cells in the fluid environment. The movement trajectory data includes the deformation amplitude and deformation direction of the cells at different time points; According to the change curve of the deformation amplitude over time, calculate the recovery speed after the deformation amplitude reaches the maximum value, and define the reciprocal of the recovery speed as the deformation recovery parameter; Statistically analyze the change frequency of the deformation direction in the movement trajectory, and define the product of the frequency and the deformation amplitude as the stiffness change parameter.

[0009] Optionally, perform a joint analysis on the three-dimensional density parameter, the blood vessel rupture direction parameter, and the stiffness change parameter, extract the time series offset of the blood vessel rupture direction parameter, and synchronously match the time series offset with the dynamic response of the stiffness change parameter to generate a synchronous parameter set, including: Obtain the direction angle values of the blood vessel rupture direction parameter at multiple time points, calculate the angle difference between adjacent time points, and generate a time series offset based on the angle difference; Obtain the parameter values of the stiffness change parameter at the same time points, and extract the dynamic change values of the parameter values over time; Align the time series offset and the dynamic change values on the time axis, calculate the sum of the products at the same time points, and generate a synchronous parameter set.

[0010] Optionally, based on the synchronous parameter set, establish a lag correlation relationship between the time series offset of the blood vessel rupture direction parameter and the recovery period of the deformation recovery parameter, and determine the weight matching result between the two through cross-modal analysis, including: Statistically analyze the time delay between the time point when the time series offset reaches the peak and the end time point of the recovery period to generate a lag time series; Calculate the correlation weight between the time series offset and the recovery period according to the length distribution of the lag time series; Superimpose the correlation weight and the sum of the products in the synchronous parameter set to generate a weight matching result.

[0011] Optionally, output the lung cancer subtype classification result according to the weight matching result and the spatio-temporal distribution characteristics of the tumor necrosis area volume ratio, including: Extract the distribution characteristics of the ratio value changing with the spatial position in the volume ratio sequence, and associate the spatial position with the time point to generate spatio-temporal distribution characteristics; Perform a mapping match between the weight matching result and the spatio-temporal distribution characteristics, and output the corresponding lung cancer subtype classification result according to the preset threshold interval.

[0012] Optionally, perform a mapping match between the weight matching result and the spatio-temporal distribution characteristics, and output the corresponding lung cancer subtype classification result according to the preset threshold interval, including: Obtain the parameter values at each time point in the weight matching result and the spatial position ratio values at the corresponding time points in the spatio-temporal distribution characteristics; Superimpose the parameter values at the same time point with the occupancy ratio to generate a comprehensive parameter for each time point; Statistically calculate the maximum and minimum values of the comprehensive parameters at all time points to generate a dynamic parameter range; Compare the range of the dynamic parameter range with a preset threshold range. If the dynamic parameter range completely falls within any threshold range, output the lung cancer subtype classification result corresponding to the threshold range.

[0013] In a second aspect, the present application provides a lung cancer type prediction system based on a fusion deep learning network, including: An acquisition module for acquiring the CT images and blood samples of lung cancer patients, dynamically segmenting the CT images to extract the vascular distribution data of the tumor region, calculating three-dimensional density parameters and vascular fracture direction parameters based on the vascular distribution data, and obtaining the volume occupancy ratio sequence of the tumor necrosis region; A processing module for performing spiral sorting on the blood samples, separating circulating tumor cells through fluid control, collecting the deformation trajectory data of the circulating tumor cells, and generating stiffness change parameters and deformation recovery parameters according to the deformation trajectory data; An analysis module for jointly analyzing the three-dimensional density parameters, the vascular fracture direction parameters and the stiffness change parameters, extracting the time series offset of the vascular fracture direction parameters, and synchronously matching the time series offset with the dynamic response of the stiffness change parameters to generate a synchronous parameter set; A establishment module for establishing a lag correlation relationship between the time series offset of the vascular fracture direction parameters and the recovery period of the deformation recovery parameters based on the synchronous parameter set, so as to determine the weight matching result of the two through cross-modal analysis; An output module for outputting the lung cancer subtype classification result according to the weight matching result and the spatio-temporal distribution characteristics of the volume occupancy ratio of the tumor necrosis region.

[0014] In a third aspect, an embodiment of the present application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a multi-dimensional data processing method for intelligent scoring of emergency patients as described in the first aspect above.

[0015] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program, and when the computer program is executed by a computer, it implements a multi-dimensional data processing method for intelligent scoring of emergency patients as described in the first aspect.

[0016] This application obtains the CT images and blood samples of lung cancer patients, dynamically segments the CT images to extract the vascular distribution data of the tumor region, calculates the three-dimensional density parameter and the vascular fracture direction parameter based on the vascular distribution data, and obtains the volume ratio sequence of the tumor necrosis region, which can fuse radiomics and microvascular spatial heterogeneity features, realize the quantitative characterization of the core necrosis volume of the tumor and the multi-dimensional analysis of the mechanical properties of the three-dimensional vascular network.

[0017] By performing spiral sorting on the blood samples, separating circulating tumor cells through fluid control, collecting the deformation trajectory data of the circulating tumor cells, and generating the stiffness change parameter and the deformation recovery parameter according to the deformation trajectory data, it is possible to accurately capture the dynamic characteristics of the mechanical response of circulating tumor cells, eliminate the interference of non-tumor cells, and establish the dynamic mapping between the cell stiffness gradient and the tumor invasion behavior.

[0018] By jointly analyzing the three-dimensional density parameter, the vascular fracture direction parameter and the stiffness change parameter, extracting the time series offset of the vascular fracture direction parameter, and synchronously matching the dynamic response of the time series offset with the stiffness change parameter to generate a synchronous parameter set, it is possible to achieve the precise alignment of the time axis of the image dynamic feature and the cell mechanical behavior, solve the problem of multi-modal data time series tomography, and enhance the spatio-temporal correlation consistency of cross-scale features.

[0019] By establishing the lag correlation relationship between the time series offset of the vascular fracture direction parameter and the recovery period of the deformation recovery parameter based on the synchronous parameter set, and determining the weight matching result of the two through cross-modal analysis, it is possible to reveal the biomechanical coupling mechanism between the dynamic evolution of the microvascular fracture direction and the mechanical recovery behavior of tumor cells, and provide a quantitative discrimination basis for subtype classification with cross-modal association.

[0020] By outputting the lung cancer subtype classification result according to the weight matching result and the spatio-temporal distribution characteristics of the tumor necrosis region volume ratio, it is possible to fuse the multi-dimensional features of vascular network heterogeneity, cell mechanical dynamic response and necrosis region spatio-temporal distribution, construct a lung cancer subtype discrimination model based on cross-modal deep association and spatio-temporal weight matching, and significantly improve the pathological interpretability and prediction accuracy of the classification result.

[0021] Furthermore, by dynamically segmenting CT images at multiple time points and extracting the three-dimensional coordinates and connection status of vascular branch nodes, the spatio-temporal evolution characteristics of the vascular network in the tumor region can be accurately captured, providing high-precision spatial topology data for subsequent quantitative analysis; calculating the node density per unit volume based on the three-dimensional coordinates of vascular branch nodes can effectively characterize the spatial heterogeneity distribution characteristics of blood vessels in the tumor microenvironment and reveal the correlation law between angiogenesis and tumor invasiveness; by statistically analyzing the three-dimensional angular distribution of blood vessel rupture directions, the mechanical properties of blood vessel structure damage can be quantified, providing important parameters for evaluating the mechanical properties of the tumor microenvironment; combining the time series analysis of the volume ratio of the tumor necrosis region to achieve dynamic monitoring of the tumor necrosis process, providing a quantitative discrimination basis with both spatial distribution characteristics and time evolution laws for lung cancer subtype identification, and significantly improving the accuracy and reliability of the diagnosis results.

[0022] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0024] Figure 1 Shows a flowchart of a lung cancer type prediction method based on a fusion deep learning network provided by the present application; Figure 2 Shows a scenario diagram of a circulating tumor cell microfluidic analysis process provided by the present application; Figure 3 Shows a flowchart of a dynamic image evaluation and processing process of lung adenocarcinoma provided by the present application; Figure 4 Shows a scenario diagram of a multi-modal intelligent diagnosis process of lung adenocarcinoma provided by the present application; Figure 5 Shows a schematic structural diagram of a lung cancer type prediction system based on a fusion deep learning network provided by the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] In order to enable those skilled in the art to better understand the solution of the present application, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application.

[0026] In some of the processes described in the specification, claims, and the above-mentioned drawings of the present application, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first" and "second" herein are used to distinguish different messages, devices, modules, etc., do not represent a sequence, and do not limit that "first" and "second" are of different types.

[0027] Researchers found that the existing lung cancer subtype classification methods lack the cross-modal collaborative analysis of CT image dynamic segmentation and blood sample mechanical response, resulting in insufficient synchronization between vascular heterogeneity parameters and the dynamic behavior of circulating tumor cells, and the lag correlation relationship between parameters has not been effectively analyzed. Based on this, a lung cancer subtype classification method based on spatio-temporal dynamic fusion is provided. This method can achieve accurate modeling of subtype classification through cross-modal weight matching between the vascular fracture direction and cell mechanical parameters. The technical solution of this application is applicable to scenarios such as early lung cancer screening, dynamic assessment of tumor heterogeneity, and prediction of targeted therapy response.

[0028] The entire R & D process embodies the fusion mechanism of multi-modal dynamic parameter synchronous modeling and spatio-temporal lag correlation analysis, aiming to overcome the limitations of image and blood parameter fragmentation, dynamic response mismatch, and fuzzy cross-modal weight allocation in existing solutions. Through the collaboration of CT image dynamic segmentation and spiral sorting fluid control, the spatio-temporal matching problem between tumor vascular heterogeneity parameters and the mechanical behavior of circulating tumor cells is solved; based on the synchronous analysis of time series offset and dynamic response, the one-sided dependence of single-modal parameters on subtype classification is eliminated; combined with lag correlation relationship modeling and weight matching optimization, the traditional method's neglect of the dynamic phase difference between parameters is broken through; finally, through the joint decision of the spatio-temporal distribution of tumor necrosis volume and cross-modal weights, the systematic improvement of the robustness, dynamics, and clinical interpretability of lung cancer subtype classification is achieved.

[0029] Next, the technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0030] Figure 1 The flowchart of a lung cancer type prediction method based on a fusion deep learning network is provided for the embodiments of the present application, as Figure 1 shown, the method includes: 101. Obtain the CT images and blood samples of lung cancer patients, perform dynamic segmentation on the CT images to extract the vascular distribution data of the tumor region, calculate the three-dimensional density parameters and vascular fracture direction parameters based on the vascular distribution data, and obtain the volume ratio sequence of the tumor necrosis region; In this step, the CT image refers to a medical image obtained by computer tomography technology. Dynamic segmentation refers to the process of continuously segmenting an image over time. Vascular distribution data refers to the quantitative data describing the vascular network structure of the tumor region. The three-dimensional density parameter refers to the distribution density characteristics of blood vessels in three-dimensional space. The vascular fracture direction parameter refers to the spatial direction characteristics at the vascular fracture. The volume ratio sequence refers to the volume ratio data of the tumor necrosis region changing over time.

[0031] In the embodiment of the present application, first, obtain the CT images of lung cancer patients from the hospital imaging system, and at the same time collect the blood samples of the patients for subsequent analysis. Secondly, use a deep learning segmentation algorithm (such as 3D U-Net) to perform dynamic segmentation on the CT images, layer by layer identify the tumor region and extract the vascular distribution data, including vascular density, orientation and branching characteristics. Then, calculate the three-dimensional density parameters in three dimensions (such as the density change rate of the X / Y / Z axes) and the vascular fracture direction parameters (such as the angular distribution at the fracture) based on the segmentation results. Finally, use the voxel counting method to statistically obtain the volume ratio sequence of the tumor necrosis region, record the changes in the necrosis region at different time points, and establish an imaging basis for subsequent multimodal analysis.

[0032] For example, in a hospital tumor diagnosis and treatment center, a 62-year-old female patient was admitted to the hospital due to a pulmonary space-occupying lesion. Enhanced CT scan showed an irregularly enhanced lesion in the upper lobe of the left lung. After the medical team obtained her chest CT images and peripheral blood samples, they used a dynamic segmentation algorithm to analyze the tumor region layer by layer and extracted the three-dimensional distribution data of the microvessels inside the lesion. Through calculation, it was found that the three-dimensional density parameters of the blood vessels in the anterior edge region of the tumor showed a gradient increasing feature, while in the dorsal region, multiple vascular fracture direction parameters were detected, showing an acute intersection between the vascular network and the pleural plane. The simultaneously generated volume ratio sequence of the tumor necrosis region showed that the necrosis volume in the central area of the lesion expanded step by step over time, which was highly correlated with the patient's progressive dyspnea symptoms. The three-dimensional reconstruction model clearly presented the spatial correspondence between the vascular fracture direction parameters and the boundary of the necrosis region.

[0033] 102. Perform spiral sorting on the blood sample, separate circulating tumor cells through fluid control, collect the deformation trajectory data of the circulating tumor cells, and generate the stiffness change parameter and the deformation recovery parameter according to the deformation trajectory data; In this step, spiral sorting process refers to the technology of separating cells using a spiral microfluidic channel. Fluid control refers to the method of regulating cell movement through hydrodynamic parameters. Circulating tumor cells refer to tumor cells that enter the peripheral blood circulation from the primary tumor. Deformation trajectory data refers to the chronological record of cell movement and deformation in fluid. Stiffness change parameter refers to the quantitative index of cell hardness change with external force. Deformation recovery parameter refers to the kinetic characteristics of a cell returning to its original state after deformation.

[0034] In the embodiment of the present application, as Figure 2 shown, first, inject a blood sample into the microfluidic chip, and separate circulating tumor cells through spiral sorting process using the centrifugal force field and fluid control technology. Secondly, use a high-speed microscopic imaging system (1000 frames per second) to capture the movement process of circulating tumor cells in the spiral channel, and collect deformation trajectory data of cell compression deformation and rebound. Then, analyze the deformation trajectory data through a mechanical model (such as the Hertz contact model), calculate the change in elastic modulus when the cell is compressed, and generate a stiffness change parameter reflecting the cell hardness characteristics. Finally, fit the cell rebound curve to extract the deformation recovery parameter (such as the time required for 80% deformation recovery) to fully characterize the dynamic mechanical response characteristics of the cell.

[0035] For example, the peripheral blood sample of this patient is processed by a spiral sorting device, and blood cells are introduced into the spiral microfluidic channel through fluid control technology. Utilizing the inertial difference between tumor cells and blood cells, the system successfully separates five circulating tumor cells with cytokeratin markers. The high-speed microscopic imaging system records the deformation trajectory data of these cells when passing through the narrow section of the spiral path: two of the cells show continuous deformation under the flow field pressure, and the deformation recovery parameter is significantly lower than the threshold; the other three show rapid elastic rebound characteristics. Further analysis shows that the fluctuation range of the stiffness change parameter of the abnormally deformed cells is more than twice the normal value, and the curve of its stiffness change parameter has a similar oscillation frequency to the blood vessel fracture direction parameter in the CT image, suggesting a potential association between the mechanical characteristics and the tumor microenvironment.

[0036] 103. Jointly analyze the three-dimensional density parameter, the blood vessel fracture direction parameter, and the stiffness change parameter, extract the time series offset of the blood vessel fracture direction parameter, and synchronously match the time series offset with the dynamic response of the stiffness change parameter to generate a synchronous parameter set; In this step, joint analysis refers to the method of integrating multi-source data for comprehensive analysis. Time series offset refers to the phase difference of the blood vessel fracture direction parameter changing with time. Dynamic response refers to the real-time reaction of the stiffness change parameter to external force stimulation. Synchronous matching refers to the process of aligning data in different time dimensions. Synchronous parameter set refers to the multi-modal feature set after time alignment.

[0037] In the embodiments of the present application, first, the three-dimensional density parameter and the blood vessel rupture direction parameter in step 101 are spatio-temporally aligned with the stiffness change parameter in step 102 to construct a multi-modal joint analysis matrix. Secondly, the time series offset of the blood vessel rupture direction parameter (such as the time delay of the blood vessel rupture event relative to the baseline) is detected by the dynamic time warping (DTW) algorithm. Then, the offset is matched with the dynamic response of the stiffness change parameter, and the phase synchronization algorithm is used to eliminate the time scale difference, generating a set of synchronization parameters including the blood vessel rupture time, the stiffness response intensity, and the delay time, and establishing a connection bridge between the imaging features and the mechanical properties.

[0038] For example, the time series of the blood vessel rupture direction parameter (collected every four hours) extracted from the CT image is jointly analyzed with the stiffness change parameter of the circulating tumor cells. Through the time-phase alignment algorithm, it is found that when the blood vessel rupture direction parameter deflects southward, the stiffness change parameter of the circulating tumor cells at the corresponding time node will synchronously show a pulsed increase. The set of synchronization parameters shows that this dynamic response becomes particularly significant when the volume ratio of the tumor necrosis area exceeds 40%. It is particularly noteworthy that the deformation recovery parameter of the circulating tumor cells corresponding to the area with a higher three-dimensional density parameter of the blood vessels always remains at a low level, revealing that the vascular density inside the tumor may affect the mechanical properties of the tumor cells in the peripheral blood through the mechanical conduction mechanism.

[0039] 104. Based on the set of synchronization parameters, establish a lag correlation relationship between the time series offset of the blood vessel rupture direction parameter and the recovery period of the deformation recovery parameter, so as to determine the weight matching result of the two through cross-modal analysis; In this step, the lag correlation relationship refers to the corresponding relationship with a time delay between different parameters. The recovery period refers to the time length for the deformation recovery parameter to complete a recovery process. Cross-modal analysis refers to an analysis method that integrates data from different detection modes. The weight matching result refers to the quantitative corresponding relationship of the importance between different parameters.

[0040] In the embodiments of the present application, first, the time series offset of the blood vessel rupture direction parameter and the recovery period of the deformation recovery parameter are extracted from the set of synchronization parameters, and the lag degree between the two is calculated through cross-correlation analysis. Secondly, a cross-modal correlation network is constructed, and the graph neural network is used to analyze the importance of different lag modes to determine the weight matching result of the blood vessel rupture event and the cell mechanical response. Then, the contribution weights of each feature are optimized and adjusted through backpropagation to highlight the lag relationships with diagnostic significance (such as the sudden drop in stiffness occurring 0.5 - 1 second after the blood vessel rupture), providing a quantitative basis for subtype classification.

[0041] For example, the time series model constructed based on cross-modal analysis shows that there is a lag correlation between the deviation of the blood vessel rupture direction parameter and the recovery period of the deformation recovery parameter. When the blood vessel rupture direction parameter deflects southward three times consecutively, the recovery period of the deformation recovery parameter will be delayed by about six hours and then shorten. Through deep learning network training, it is determined that the three-dimensional density parameter of blood vessels occupies the main weight in cross-modal association, while the time series offset of the blood vessel rupture direction parameter regulates the dynamic response of the stiffness change parameter. The weight matching result shows that the correlation strength between the acute angle blood vessel rupture direction parameter on the dorsal side of the tumor of this patient and the abnormal deformation recovery characteristics of circulating tumor cells reaches the clinical significance threshold, providing a key basis for subtype identification.

[0042] 105. Output the lung cancer subtype classification result according to the weight matching result and the spatio-temporal distribution characteristics of the volume ratio of the tumor necrosis area.

[0043] In this step, the spatio-temporal distribution characteristics refer to the variation law of the volume ratio in time and space. The lung cancer subtype classification result refers to the diagnostic conclusion of the lung cancer type obtained based on multi-parameter analysis.

[0044] In the embodiment of the present application, first, integrate the weight matching result in step 104 and the spatio-temporal distribution characteristics of the volume ratio of the tumor necrosis area in step 101 to construct a three-dimensional classification feature space. Secondly, use the random forest algorithm to train the classification model, and jointly model the blood vessel rupture lag pattern, cell mechanical response characteristics and necrosis area distribution. Then, screen the most discriminative combined features (such as the decrease in stiffness after blood vessel rupture accompanied by central necrosis) through feature importance analysis, and finally output the lung cancer subtype classification result including specific types such as adenocarcinoma and squamous cell carcinoma, providing an objective basis for clinical treatment decision-making.

[0045] For example, by synthesizing the weight matching result and the spatio-temporal distribution characteristics of the volume ratio of the tumor necrosis area, the system outputs the lung cancer subtype classification result as "vascular invasion type adenocarcinoma of the lung". The characteristics of this subtype are as follows: the acute angle distribution of the blood vessel rupture direction parameter coexists with the stepped expansion of the necrosis area volume, and the circulating tumor cells in peripheral blood show dual mechanical abnormalities of high-frequency oscillation of the stiffness change parameter and delayed deformation recovery. Comparing with the historical case database, it is found that the sensitivity of patients with this subtype to anti-angiogenic drugs is three times higher than that of traditional adenocarcinoma. Based on this, the clinical team formulates an individualized regimen of bevacizumab combined with chemotherapy. The reexamination two weeks later shows that the degree of lesion enhancement decreases, and the number of circulating tumor cells decreases by more than 50%, verifying the guiding value of the subtype classification model for precision treatment.

[0046] In summary, steps 101 to 105 achieve the innovation of an intelligent prediction method for lung cancer subtypes based on multimodal data fusion. By integrating the vascular distribution characteristics of CT images and the mechanical properties of circulating tumor cells in blood samples, a cross-modal joint analysis framework is constructed, solving the problem of incomplete tumor feature representation in traditional single-modal diagnosis. By dynamically segmenting to extract the three-dimensional density parameters and fracture direction parameters of tumor blood vessels, and combining the stiffness change and deformation recovery characteristics of circulating tumor cells, a synchronous matching mechanism between the dynamic changes of blood vessels and the mechanical response of cells is established, significantly improving the accuracy and reliability of lung cancer subtype classification and providing a multi-dimensional decision-making basis for precision medicine.

[0047] In order to quantify the spatio-temporal evolution characteristics of the vascular network in the tumor microenvironment through multi-time point dynamic segmentation and three-dimensional spatial analysis techniques, calculate the spatial density distribution based on the three-dimensional coordinates of vascular branch nodes to evaluate vascular heterogeneity, analyze the spatial law of vascular structure damage in combination with fracture direction parameters, and monitor the dynamic trend of tumor tissue degenerative changes using the sequence of the proportion of necrotic area volume, so as to construct an association model between the morphology of the vascular network and the pathological progression of tumors, providing an accurate quantitative evaluation basis for the dynamic monitoring of the tumor microenvironment and anti-angiogenic therapy.

[0048] In some embodiments, as described in step 101, performing dynamic segmentation on the CT image to extract the vascular distribution data of the tumor region, calculating the three-dimensional density parameter and the vascular fracture direction parameter based on the vascular distribution data, and obtaining the sequence of the proportion of the volume of the tumor necrosis region, includes: 201. Perform multi-time point segmentation operations on the CT image, and extract the vascular distribution data of the tumor region at each time point, where the vascular distribution data includes the three-dimensional coordinates of vascular branch nodes and the connection status between adjacent nodes; In step 201, the multi-time point segmentation operation refers to performing segmentation processing on the CT image at different time points. The vascular branch node refers to the bifurcation connection point in the vascular network. The connection status refers to the description of the connectivity between vascular nodes.

[0049] In the embodiments of the present application, first, a time series registration algorithm is used to preprocess the CT image to ensure the spatial alignment of images at different time points. Secondly, at each time point, a three-dimensional vascular segmentation algorithm (such as VesselNet based on deep learning) is used to accurately segment the tumor region and extract the vascular network structure. Then, the vascular network is transformed into a topological graph composed of vascular branch nodes and connection edges through a skeletonization algorithm, recording the three-dimensional coordinates (such as the positions on the X / Y / Z axes) of each node and the connection status (such as the connectivity between adjacent nodes). Finally, the vascular topological graphs at each time point are arranged along the time axis to construct a complete vascular evolution time series database.

[0050] 202. Calculate the density distribution of blood vessels in three-dimensional space based on the three-dimensional coordinates of the blood vessel branch nodes, and define the number of blood vessel branch nodes per unit volume as the three-dimensional density parameter; In step 202, the density distribution refers to the density degree characteristic of blood vessels in space. The unit volume refers to the set standard space measurement unit. The number of blood vessel branch nodes refers to the count of blood vessel bifurcation points within a specific spatial range.

[0051] In the embodiment of the present application, first, starting from the three-dimensional coordinates of the blood vessel branch nodes obtained in step 201, divide the tumor region into 1 mm³ cubic voxel grids. Secondly, count the number of blood vessel nodes contained in each voxel, and define this value as the three-dimensional density parameter of this voxel. Then, perform smoothing processing on the density distribution through three-dimensional Gaussian filtering to eliminate the noise interference caused by small blood vessel fluctuations. Finally, integrate the density parameters at each time point according to the spatial position to generate a density field sequence that can reflect the spatial distribution characteristics of blood vessels.

[0052] 203. Statistically analyze the fracture directions of all blood vessel branch nodes, and generate blood vessel fracture direction parameters according to the angular distribution of the fracture directions in the three-dimensional coordinate system; In step 203, the fracture direction refers to the spatial orientation characteristic at the fracture of the blood vessel. The three-dimensional coordinate system refers to the spatial reference system composed of the X / Y / Z axes. The angular distribution refers to the azimuth statistical characteristic of the fracture direction in three-dimensional space.

[0053] In the embodiment of the present application, first, identify all fracture nodes (i.e., the end nodes with only one connection point) in the blood vessel topology graph, and extract their fracture directions (such as the vector from the last connection point to the fracture point). Secondly, project each fracture direction vector onto the three-dimensional coordinate system and calculate its angles with the X / Y / Z axes (such as in the range of 0 - 180 degrees). Then, use the spherical statistical method to analyze the angular distribution characteristics of the fracture directions, and generate blood vessel fracture direction parameters including the main fracture direction (such as the 45° direction with the highest proportion) and the direction dispersion degree. Finally, arrange the fracture direction parameters at each time point in a time series for subsequent dynamic analysis.

[0054] 204. Generate a volume ratio sequence by measuring the ratio of the volume of the tumor necrosis region to the total volume of the tumor region at each time point.

[0055] In step 204, the volume measurement refers to the calculation of the spatial proportion of the tumor necrosis region. The total volume refers to the complete spatial volume of the tumor region. The ratio calculation refers to the volume ratio of the necrosis region to the whole tumor.

[0056] In the embodiments of the present application, first, based on the tumor region segmentation result of step 201, the threshold segmentation method (such as the Otsu algorithm) is used to distinguish the tumor necrosis region and the active tissue region. Secondly, the volume of the necrosis region is calculated through three-dimensional connected component analysis and divided by the total tumor volume to obtain the volume ratio. Then, the volume ratio values at each time point are recorded in time series to generate a volume ratio sequence reflecting the evolution of the necrosis region. Finally, the time series curve is smoothed by the moving average method to eliminate the interference of short-term fluctuations on the analysis result.

[0057] The following is a specific example, such as Figure 3 shown as In Figure 3 the lung adenocarcinoma dynamic imaging evaluation system, the multi-phase CT angiography analysis technology realizes the accurate monitoring of the evolution of the tumor microenvironment. For a mixed ground-glass nodule in the lower lobe of the right lung of a certain patient, the system first performs time point segmentation on the three-phase enhanced CT images (step 201), extracts the three-dimensional coordinates and connection status of the vascular branch nodes in the tumor region in the arterial phase, venous phase and delayed phase respectively, and finds that there is a characteristic vascular "pruning" phenomenon in the edge region of the lesion. Based on the spatial distribution of the vascular nodes (step 202), the system calculates and shows that the three-dimensional density parameter in the core area of the lesion is significantly higher than that in the surrounding area, presenting a typical "central enrichment" distribution pattern. By statistically analyzing the vascular rupture direction (step 203), a rupture direction parameter showing an angle aggregation of 90 - 120 degrees in the upper pole region of the lesion is generated, indicating that there is vascular rupture caused by mechanical stress in this region. At the same time, the system measures the proportion of the necrosis region in each phase (step 204), generates a proportion sequence showing that the necrosis volume gradually expands over time, and its expansion direction highly coincides with the vascular rupture aggregation area. When the CT of a new case presents similar central enrichment density parameters and the vascular rupture pattern in the upper pole region, the system automatically predicts it as an invasive lung adenocarcinoma subtype, guides the clinical priority for EGFR gene detection, and finally the pathology confirms it as invasive adenocarcinoma with micropapillary components. This case optimizes the multi-phase vascular analysis algorithm, improves the early recognition ability of the invasive characteristics of lung adenocarcinoma, and forms an intelligent evaluation system from dynamic vascular analysis to molecular subtype prediction.

[0058] In summary, steps 201 to 204 achieve the dynamic quantification and three-dimensional modeling of the tumor vascular distribution characteristics. By extracting the three-dimensional coordinates and connection status of the vascular branch nodes through the multi-time point segmentation technology, the accurate quantification of the spatial distribution of the vascular network is realized. Based on the number of vascular nodes per unit volume, the three-dimensional density parameter is calculated, and combined with the angular distribution of the rupture direction, the vascular rupture direction parameter is generated, constructing a dynamic evaluation system for tumor vascular heterogeneity. Through the volume ratio sequence of the necrosis region at consecutive time points, it provides a vascular morphological index with spatio-temporal evolution characteristics for lung cancer subtype classification, enhancing the characterization ability of tumor microenvironment characteristics.

[0059] In some embodiments, as described in step 102, the blood sample is subjected to spiral sorting to separate circulating tumor cells through fluid control, and the deformation trajectory data of the circulating tumor cells is collected, and a stiffness change parameter and a deformation recovery parameter are generated based on the deformation trajectory data, including: 301. Inject the blood sample into a spiral sorting device, and separate circulating tumor cells from non-tumor cells by adjusting the fluid dynamic direction; In step 301, the spiral sorting device refers to an experimental device that uses a spiral microfluidic channel to separate cells. The fluid dynamic direction refers to the mechanical action direction generated by the fluid in the spiral channel. Circulating tumor cells refer to tumor cells that enter the peripheral blood from the primary tumor. Non-tumor cells refer to other cell components in the blood except tumor cells. Adjustment refers to the process of optimizing the separation effect by changing experimental parameters.

[0060] In the embodiments of the present application, first, the collected blood sample is injected into the inlet end of the spiral sorting device through a precision syringe pump, and the initial flow rate is set to 1 ml / min. Secondly, the fluid dynamic direction in the device is adjusted through a microfluidic control system to generate a stable laminar flow and centrifugal force field in the spiral channel. Then, taking advantage of the differences in mass and deformation ability between tumor cells and blood cells, the circulating tumor cells move along the outer side of the spiral channel due to greater inertia, while the non-tumor cells are thrown to the inner side due to strong deformation ability. Finally, a sorting valve is set at the end of the channel, and highly pure circulating tumor cells are collected at the outer outlet to complete the cell separation process.

[0061] 302. Collect the movement trajectory data of the separated circulating tumor cells in the fluid environment, where the movement trajectory data includes the deformation amplitude and deformation direction of the cells at different time points; In step 302, the movement trajectory data refers to the spatio-temporal record data of the cell movement in the fluid. The deformation amplitude refers to the degree of deformation generated by the cell under the action of an external force. The deformation direction refers to the main spatial orientation of the cell deformation. Collection refers to the process of obtaining experimental data through detection equipment.

[0062] In the embodiments of the present application, first, the separated circulating tumor cells are suspended in a buffer solution and re-injected into the observation channel at a constant flow rate. Secondly, a high-speed microscopic imaging system (such as 1000 frames / second) is used to continuously capture the cell movement process to obtain a time-series image of the cell contour. Then, the cell contour at each time point is extracted through an edge detection algorithm, and the deformation amplitude (such as the change amount of the ratio of the long axis to the short axis of the cell) and the deformation direction (such as the angle between the deformation main axis and the flow direction) are calculated. Finally, the measurement results at each time point are integrated into a movement trajectory data table including a time stamp, a deformation amplitude, and a deformation direction.

[0063] 303. Calculate the recovery speed after the deformation amplitude reaches the maximum value according to the change curve of the deformation amplitude over time, and define the reciprocal of the recovery speed as the deformation recovery parameter; In step 303, the change curve refers to the function image of the deformation amplitude changing over time. The recovery speed refers to the speed at which the cell returns to its original state from the maximum deformation state. The reciprocal refers to the reciprocal operation relationship in mathematics. The deformation recovery parameter refers to the characteristic parameter that quantifies the elastic recovery ability of the cell.

[0064] In the embodiments of the present application, first, extract the change curve of the deformation amplitude over time from the movement trajectory data, and determine the time point when the deformation reaches the maximum value through the peak detection algorithm. Secondly, select the deformation recovery data within the 100 ms time window after the maximum value, and use linear regression to fit the slope of the recovery curve. Then, define the reciprocal of the slope of the recovery curve (i.e., the time required to recover a unit deformation) as the deformation recovery parameter. The smaller the value of this parameter, the stronger the elastic recovery ability of the cell. Finally, take the average of multiple deformation-recovery cycles of the same batch of cells to obtain a representative deformation recovery parameter.

[0065] 304. Statistically analyze the change frequency of the deformation direction in the movement trajectory, and define the product of the frequency and the deformation amplitude as the stiffness change parameter.

[0066] In step 304, the change frequency refers to the frequency of change of the deformation direction in the trajectory. The product refers to the result of the multiplication operation in mathematics. The stiffness change parameter refers to the quantitative index that reflects the dynamic change of the cell hardness. Statistics refers to the method of summarizing and analyzing data.

[0067] In the embodiments of the present application, first, analyze the change of the deformation direction in the movement trajectory data, and calculate the change frequency of the deformation direction in the frequency band of 0.1 - 10 Hz through Fourier transform. Secondly, multiply the frequency value of each frequency band by the deformation amplitude at the corresponding moment to obtain the frequency-amplitude product spectrum. Then, select the frequency band corresponding to the maximum value in the product spectrum as the characteristic frequency band, and define the product value of this frequency band after normalization as the stiffness change parameter, which reflects the dynamic response characteristics of the cell stiffness to the external fluid force. Finally, perform correlation analysis on the stiffness change parameter and the deformation recovery parameter to establish a complete description system of the cell mechanical properties.

[0068] The following is a specific example: In the intelligent analysis system for liquid biopsy of lung cancer, microfluidic chip technology enables the precise assessment of the mechanical properties of circulating tumor cells. For peripheral blood samples from suspected small cell lung cancer patients, the system first injects the blood into a spiral sorting device (step 301), and successfully separates the population of circulating tumor cells with characteristic rapid movement by optimizing the fluid dynamic direction. The high-speed microscopic imaging system collects the movement trajectory data of these cells in the spiral channel (step 302), and records the "slingshot" deformation mode and its unique deformation direction presented when the cells pass through the narrow area. The system analyzes the change curve of the deformation amplitude over time (step 303), and calculates that small cell lung cancer cells have significantly rapid deformation recovery parameters, which is consistent with the high elasticity characteristics of the cell membranes of this type of tumor. Further, it statistically analyzes the change frequency of the deformation direction (step 304) to generate a pattern map showing the characteristic high-frequency stiffness change parameters of small cell lung cancer. When the new sample test shows similar rapid recovery parameters and high-frequency stiffness changes, the system automatically associates with the small cell lung cancer prediction model, guides the clinical detection of neuroendocrine markers, and finally the pathology confirms small cell lung cancer with the characteristic "oat-like" morphology. This case optimizes the fluid parameter settings of spiral sorting, improves the specific recognition ability of circulating tumor cells in small cell lung cancer, and forms an intelligent diagnosis closed loop from cell mechanical property analysis to pathological subtype determination.

[0069] In summary, steps 301 to 304 achieve the precise detection and parametric characterization of the mechanical properties of circulating tumor cells. Through the fluid dynamic regulation of the spiral sorting device, the efficient separation and purification of circulating tumor cells are realized. Based on the high-speed trajectory acquisition technology to obtain the dynamic process of cell deformation, the deformation recovery parameter is defined by the reciprocal of the deformation recovery speed, and the stiffness change parameter is defined by the product of the change frequency and amplitude of the deformation direction, thus establishing a quantitative index system for cell mechanical properties. This technology breaks through the static limitations of traditional cell detection methods and provides characteristic parameters reflecting the dynamic mechanical behavior of cells for lung cancer subtype classification.

[0070] In some embodiments, as described in step 103, the three-dimensional density parameter, the blood vessel rupture direction parameter, and the stiffness change parameter are jointly analyzed, the time series offset of the blood vessel rupture direction parameter is extracted, and the time series offset is synchronously matched with the dynamic response of the stiffness change parameter to generate a synchronous parameter set, including: 401. Obtain the direction angle values of the blood vessel rupture direction parameter at multiple time points, calculate the angle difference between adjacent time points, and generate a time series offset according to the angle difference; In step 401, the direction angle value refers to the specific angle value of the blood vessel rupture direction in three-dimensional space. The angle difference refers to the change amount of the direction angle at adjacent time points. The time series offset refers to the phase difference of the blood vessel rupture direction changing with time. Calculation refers to the process of obtaining characteristic parameters through mathematical operations.

[0071] In the embodiment of the present application, first, extract the measurement data of the blood vessel rupture direction parameters at consecutive time points from the blood vessel image analysis system, and obtain the direction angle values (such as 45°, 60°, etc.) recorded at each time point. Secondly, adopt the differential calculation method, subtract the angle value of the previous time point from the angle value of the next time point to obtain the angle difference between adjacent time points (such as 60° - 45° = 15°). Then, arrange these angle differences in chronological order to form a time series offset reflecting the change trend of the blood vessel rupture direction. Finally, perform smoothing processing on the offset sequence by the moving average method to eliminate the random fluctuations caused by measurement noise.

[0072] 402. Obtain the parameter values of the stiffness change parameter at the same time point, and extract the dynamic change values of the parameter values changing with time; In step 402, the parameter value refers to the specific value of the stiffness change parameter at a specific time point. The dynamic change value refers to the fluctuation characteristics of the parameter changing with time. Extraction refers to the analysis method of obtaining specific information from data. The same time point refers to the synchronous time node of the blood vessel rupture direction and the stiffness change parameter.

[0073] In the embodiment of the present application, first, obtain the stiffness change parameter values at the same time point as the blood vessel rupture direction parameter from the cell mechanics test data. Secondly, adopt the time series analysis method to calculate the difference between the stiffness change parameters at adjacent time points to obtain the dynamic change values reflecting the parameter fluctuation situation. Then, convert the dynamic change values at different time points to the same dimension through normalization processing for subsequent comparative analysis. Finally, draw a curve graph of the dynamic change values changing with time to visually display the change law of the stiffness parameter.

[0074] 403. Align the time series offset and the dynamic change value on the time axis, calculate the sum of the products at the same time point of the two, and generate a synchronous parameter set.

[0075] In step 403, time axis alignment refers to synchronously processing the time dimensions of different data. The sum of products refers to the accumulated result after multiplying the parameters at the corresponding time points. The synchronous parameter set refers to the characteristic set reflecting the time synchronization of multiple parameters. Generation refers to the process of obtaining the final result through calculation.

[0076] In the embodiments of the present application, first, the time series offset generated in step 401 and the dynamic change value obtained in step 402 are matched and aligned at the same time point to ensure that their time axes are completely synchronized. Secondly, at each time point, the corresponding offset and dynamic change value are multiplied to obtain the co-variation amount at this time point. Then, the product results of all time points are accumulated to generate a set of synchronization parameters reflecting the degree of coordination between the change in the blood vessel rupture direction and the change in cell stiffness. Finally, through normalization processing, the values in the set of synchronization parameters are converted to the range of 0-1 for subsequent weight matching analysis.

[0077] The following is a specific example: In a multi-modal dynamic monitoring system for lung adenocarcinoma, the spatio-temporal feature fusion technology realizes the accurate assessment of the evolution of tumor blood vessels. For a patient with a mixed ground-glass nodule in the upper lobe of the left lung, the system first obtains the time series data of the blood vessel rupture direction parameters in the three-phase enhanced CT images (step 401), calculates the offset of the direction angle value from the arterial phase to the venous phase, and finds a characteristic 45-degree direction mutation in the anterior edge area of the lesion. Synchronously analyze the dynamic monitoring results of the circulating tumor cell stiffness change parameters (step 402), extract the stiffness parameter fluctuation characteristics matching the CT scan time points, and show an obvious decrease in stiffness in the venous phase. Through precise time axis alignment (step 403), the system synchronously correlates the blood vessel rupture direction offset with the cell stiffness change value to generate a set of synchronization parameters showing a strong correlation between the two in the venous phase. When a new case presents a similar synchronous pattern of direction mutation and stiffness decrease, the system automatically predicts it as a high-risk lung adenocarcinoma subtype with micropapillary components, guiding clinical targeted treatment decisions. Postoperative pathology confirms the existence of characteristic blood vessel co-infiltration in the predicted area. This case optimizes the multi-modal time alignment algorithm, improves the early warning ability for the invasive biological behavior of lung adenocarcinoma, and forms an intelligent evaluation system from dynamic analysis of imaging features to prediction of treatment response.

[0078] In summary, steps 401 to 403 achieve spatio-temporal synchronization and correlation analysis of cross-modal feature parameters. By extracting the time series offset of the blood vessel rupture direction parameters and aligning the time axis with the dynamic response of the stiffness change parameters, a synchronous matching mechanism for multi-modal data is constructed. By calculating the sum of the products of the time series offset and the dynamic change value to generate a set of synchronization parameters, a dynamic correlation model between the blood vessel morphology change and the cell mechanical properties is realized, significantly improving the spatio-temporal consistency of cross-modal feature fusion and providing an accurate data basis for subsequent lag relationship analysis.

[0079] In some embodiments, as described in step 104, based on the set of synchronization parameters, establishing a lag correlation relationship between the time series offset of the blood vessel rupture direction parameters and the recovery period of the deformation recovery parameters to determine the weight matching result between the two through cross-modal analysis includes: 501. Statistically analyze the time delay between the time point when the time series offset reaches the peak and the end time point of the recovery period, and generate a lag time series; In step 501, the lag time series refers to the time difference series between the time series offset and the recovery period. The peak time point refers to the moment when the time series offset reaches the maximum. The end time point of the recovery period refers to the time node when the deformation recovery process is completed. The time delay refers to the time interval between the occurrence times of two events. Statistics refers to the process of summarizing and analyzing data.

[0080] In the embodiments of the present application, first, peak detection is performed on the time series offset generated in step 401, and the peak time point corresponding to each offset peak is determined through the sliding window extreme value analysis method. Secondly, the end time point when each recovery period is completely ended is extracted from the deformation recovery parameters in step 303. Then, the time difference between each offset peak and the end point of the nearest subsequent recovery period is calculated to generate a lag time series reflecting the delay relationship between the blood vessel rupture event and the cell mechanical response. Finally, abnormal lag values are removed through box plot analysis, and the delay data with biological significance is retained.

[0081] 502. Calculate the correlation weight between the time series offset and the recovery period according to the length distribution of the lag time series; In step 502, the length distribution refers to the statistical characteristics of the duration of the lag time series. The correlation weight refers to a quantitative index reflecting the correlation between the time series offset and the recovery period. Calculation refers to the process of obtaining characteristic parameters through mathematical operations. The recovery period refers to the time length of a recovery process of the deformation recovery parameters.

[0082] In the embodiments of the present application, first, the lag time series is grouped and statistically analyzed according to the delay duration, and a length distribution histogram of the lag time is drawn. Secondly, the kernel density estimation method is used to calculate the probability density of different lag durations, and the reciprocal of the probability density is used as the correlation weight of this lag duration (for example, the shorter the delay, the higher the weight). Then, the weight value is normalized to the range of 0-1 through the sigmoid function to ensure that the weight difference between long delays and short delays conforms to biological laws. Finally, a mapping relationship table between the lag time and the weight is established, and the corresponding weight value is assigned to each observed lag time.

[0083] 503. Superimpose the sum of the products of the correlation weight and the synchronization parameter set to generate a weight matching result.

[0084] In step 503, superposition refers to the operation of mathematically accumulating different parameters. The weight matching result refers to a comprehensive parameter that reflects the matching degree between the blood vessel rupture direction and the cell mechanical properties. Generation refers to the process of obtaining the final result through calculation. The sum of products refers to the accumulated result after multiplying the parameters at the corresponding time points in the synchronous parameter set.

[0085] In the embodiments of the present application, first, the sum of products values at each time point are extracted from the synchronous parameter set in step 403. Secondly, according to the mapping relation table established in step 502, the corresponding associated weights are matched for each sum of products. Then, the sum of products and the associated weights are multiplied corresponding to each time point to implement weighted superposition calculation. Finally, the weighted results at all time points are accumulated and summed to generate a weight matching result that comprehensively considers the time delay and the synchronization strength, and this result quantifies the overall association degree between the blood vessel rupture event and the cell mechanical response.

[0086] The following is a specific example: In the multi-modal dynamic evaluation system for lung squamous cell carcinoma, the spatio-temporal feature fusion technology realizes the accurate quantification of the tumor blood vessel-mechanics coupling characteristics. For the continuous monitoring data of an irregular mass in the right middle lobe of a patient's lung, the system first statistically calculates the time delay between the peak point of the time series offset of the blood vessel rupture direction and the end point of the mechanical recovery cycle of circulating tumor cells (step 501), and generates a time series showing a characteristic 0.8-second lag in the upper pole region of the lesion. By analyzing the distribution characteristics of the lag time series (step 502), the associated weights between the blood vessel direction mutation and the cell mechanical recovery are calculated, and it is found that a high-weight coupling mode appears in the tumor core region. The system superimposes the associated weights with the previously generated blood vessel-mechanics synchronous parameter set (step 503), and finally generates a quantitative map showing a significant weight matching result in the leading edge region of the lesion. When a new case presents similar upper pole region lag characteristics and high-weight coupling in the core region, the system automatically predicts it as a subtype of lung squamous cell carcinoma with a tendency to form keratin pearls, guiding the clinical detection of PD-L1 expression, and pathological examination confirms the existence of characteristic dyskeratotic cell clusters in the predicted region. This case optimizes the weight algorithm for spatio-temporal coupling, improves the recognition accuracy of the specific blood vessel-mechanics characteristics of squamous cell carcinoma, and forms an intelligent decision-making system from multi-modal dynamic analysis to immunotherapy response prediction.

[0087] In summary, steps 501 to 503 realize the lag correlation modeling between the dynamic changes of blood vessels and the cell mechanical recovery. By statistically calculating the time delay between the peak of the time series offset and the end point of the recovery cycle, a lag correlation model between the blood vessel rupture direction and the cell deformation recovery is established. The associated weights are calculated based on the length distribution of the lag time series and superimposed and optimized with the synchronous parameter set to generate a weight result reflecting the dynamic matching relationship of cross-modal features. This technology breaks through the limitations of traditional static correlation analysis and realizes the quantitative characterization of the dynamic interaction mechanism between tumor blood vessel abnormalities and cell mechanical responses.

[0088] In some embodiments, as described in step 105, according to the spatio-temporal distribution characteristics of the weight matching result and the volume ratio of the tumor necrosis region, the lung cancer subtype classification result is output, including: 601. Extract the distribution characteristics of the ratio values in the volume ratio sequence that change with the spatial position, and associate the spatial position with the time point to generate spatio-temporal distribution characteristics; In step 601, the ratio value refers to the ratio of the volume of the tumor necrosis region to the total volume. The spatial position refers to the specific position coordinates in a three-dimensional coordinate system. The spatio-temporal distribution characteristics refer to the variation law of the volume ratio in the time and space dimensions. Association refers to establishing a correspondence relationship between data in different dimensions. Extraction refers to an analysis method for obtaining specific information from data.

[0089] In the embodiments of the present application, first, a volume ratio sequence is extracted from the tumor image analysis system, and the ratio values of the tumor necrosis region at different spatial positions (such as front / back / left / right / up / down) at each time point are obtained. Secondly, a spatial interpolation algorithm is used to convert the discrete spatial position ratio values into a continuous three-dimensional distribution map. Then, the three-dimensional distribution maps at each time point are arranged in chronological order, and the spatial position is associated with the time point through spatio-temporal coding technology to generate spatio-temporal distribution characteristics that can simultaneously reflect the spatial distribution and time evolution of the necrosis region. Finally, principal component analysis is performed on the spatio-temporal distribution characteristics to extract the most representative spatio-temporal change pattern.

[0090] 602. Perform mapping matching between the weight matching result and the spatio-temporal distribution characteristics, and output the corresponding lung cancer subtype classification result according to a preset threshold interval.

[0091] In step 602, mapping matching refers to the process of corresponding and associating different characteristic parameters. The preset threshold interval refers to the range of classification judgment criteria set in advance. The lung cancer subtype classification result refers to the lung cancer type diagnosis conclusion based on multi-parameter analysis. Output refers to the operation of displaying or returning the final result.

[0092] In the embodiments of the present application, first, the weight matching result obtained in step 503 is aligned with the spatio-temporal distribution characteristics generated in step 601 on the time axis to ensure that the time points of both are completely corresponding. Secondly, the weight matching result at each time point is matched with the spatio-temporal distribution characteristics at the corresponding moment through a feature mapping algorithm to establish a mapping relationship table of parameter-characteristics. Then, according to the threshold interval set based on clinical experience (such as 0.3 - 0.5 corresponding to adenocarcinoma, 0.6 - 0.8 corresponding to squamous cell carcinoma), the mapping matching result is classified into the corresponding interval. Finally, the majority voting method is used to count the matching times in each interval, and the lung cancer subtype classification result corresponding to the interval with the highest frequency is output.

[0093] The following is a specific example: In the intelligent classification system for lung adenocarcinoma, the multi-modal spatio-temporal feature fusion technology realizes the accurate assessment of tumor heterogeneity. For a patient with a mixed-density nodule in the upper lobe of the left lung, the system first extracts the spatial distribution features of the sequence of the proportion of necrotic volume in the three-phase enhanced CT scans (step 601), establishes a "fan-shaped" expansion pattern showing the progression of the posterior basal area of the lesion over time, and generates distribution features with spatio-temporal evolution rules. By performing three-dimensional mapping of the previously obtained blood vessel-mechanics weight matching results with this spatio-temporal feature (step 602), the system discovers that the central area of the lesion presents a synergistic pattern of high-weight coupling and rapid spatial expansion, and automatically outputs a characteristic spectrum conforming to micropapillary lung adenocarcinoma according to the preset classification threshold. When the CT of a new case shows a similar central area synergistic pattern, the system triggers the prediction of the probability of EGFR sensitive mutation, guiding clinical priority for targeted therapy, and the postoperative pathology confirms invasive adenocarcinoma with micropapillary structure. This case optimizes the spatio-temporal collaborative analysis algorithm, improves the early recognition ability of high-risk subtypes, and forms an intelligent decision-making closed loop from image heterogeneity analysis to individualized treatment.

[0094] In summary, steps 601 to 602 achieve the intelligent classification of lung cancer subtypes based on spatio-temporal distribution features. By extracting the spatio-temporal evolution rules of the proportion of the volume of the tumor necrosis area and mapping and correlating it with the weight matching results, a classification decision model integrating multi-dimensional features is constructed. By integrating multi-source features such as blood vessel dynamic characteristics, cellular mechanical responses, and necrosis area distribution, the classification of lung cancer subtypes has achieved a leap from a single morphological index to spatio-temporal dynamic features, significantly improving the biological rationality and clinical applicability of the classification results.

[0095] In some embodiments, as described in step 602, mapping and matching the weight matching results with the spatio-temporal distribution features and outputting the corresponding lung cancer subtype classification results according to a preset threshold range includes: 701. Obtain the parameter values at each time point in the weight matching results, and the spatial position occupancy ratios at the corresponding time points in the spatio-temporal distribution features; In step 701, the parameter value refers to the specific value of the weight matching result at a specific time point. The spatial position occupancy ratio refers to the volume occupancy ratio value of a specific spatial position. Obtain refers to the process of reading the required information from the data. The corresponding time point refers to the moment when the two parameters have the same time stamp.

[0096] In the embodiments of the present application, first, extract the parameter values recorded at each time point from the weight matching results of step 503, and at the same time obtain the spatial position occupancy ratios at the same time points from the spatio-temporal distribution features of step 601. Secondly, ensure that the parameter values and the occupancy ratios come from the same observation moment through time stamp matching. Then, perform standardization processing on the data to eliminate the dimensional differences between different parameters. Finally, establish a correspondence table of time-parameter-occupancy ratio to provide structured data for subsequent comprehensive analysis.

[0097] 702. Superimpose the parameter values at the same time point and the occupancy ratio to generate a comprehensive parameter for each time point. In step 702, the comprehensive parameter refers to a composite parameter that integrates weight matching and spatial occupancy ratio. Superimposing refers to the operation of mathematically combining different parameters. The same time point refers to the data points with the same time mark. Generating refers to the process of obtaining a new parameter through calculation.

[0098] In the embodiments of the present application, first, perform a weighting process on the parameter values and occupancy ratios at the same time point obtained in step 701, and assign different weights to the two types of parameters according to clinical importance (such as a parameter value weight of 0.6 and an occupancy ratio weight of 0.4). Secondly, use the linear weighting method to superimpose the two types of parameters according to the weights to generate a comprehensive parameter reflecting the comprehensive characteristics of this time point. Then, perform a smoothing process on the comprehensive parameters of adjacent time points through the moving average method to eliminate the interference of random fluctuations. Finally, draw a curve graph of the comprehensive parameter changing with time to intuitively display the evolution trend of the characteristics.

[0099] 703. Statistically calculate the maximum and minimum values of the comprehensive parameters at all time points to generate a dynamic parameter interval. In step 703, the dynamic parameter interval refers to the range interval of the comprehensive parameter changing with time. The maximum value refers to the highest value in a set of data. The minimum value refers to the lowest value in a set of data. Statistically calculating refers to the method of summarizing and analyzing data.

[0100] In the embodiments of the present application, first, identify the maximum and minimum values from the comprehensive parameters at all time points generated in step 702 through an extreme value detection algorithm. Secondly, calculate the difference between the maximum value and the minimum value, and equally divide this difference into several intervals (such as 5 intervals). Then, statistically calculate the distribution frequency of the comprehensive parameter in each interval to determine the main change range of the parameter. Finally, use the maximum value and the minimum value as the boundaries to generate a dynamic parameter interval that includes the complete fluctuation range of the parameter, providing a quantitative basis for subtype discrimination.

[0101] 704. Compare the range of the dynamic parameter interval with a preset threshold interval. If the dynamic parameter interval completely falls within any one of the threshold intervals, output the lung cancer subtype classification result corresponding to the threshold interval.

[0102] In step 704, range comparison refers to comparing the inclusion relationship between two numerical intervals. The preset threshold interval refers to the pre-set classification judgment standard range. Completely falling within means that one interval is completely included by another interval. The lung cancer subtype classification result refers to the lung cancer type diagnosis conclusion obtained based on the parameter interval matching. Outputting refers to the operation of displaying or returning the final result.

[0103] In the embodiments of the present application, first, the dynamic parameter range determined in step 703 is compared with the threshold range preset by clinical experts, and the overlapping ratio is calculated using the interval inclusion degree algorithm. Secondly, if the dynamic parameter range completely falls within a certain threshold range (such as the 0.3 - 0.5 range), it is determined as the subtype corresponding to this range (such as adenocarcinoma). Then, when the dynamic parameter range spans multiple threshold ranges, the most likely subtype category is determined using the maximum overlap principle. Finally, a lung cancer subtype classification result including the specific subtype name (such as stage T2 lung adenocarcinoma) and confidence level is output, providing decision support for clinical treatment.

[0104] The following is a specific example, such as Figure 4 shown as: In Figure 4 In a multi-modal intelligent diagnosis system for lung adenocarcinoma, the dynamic parameter fusion technology realizes the accurate classification of tumor subtypes. For a patient with a mixed ground-glass nodule in the middle lobe of the right lung, the system first obtains the weight matching result of the vascular-mechanical coupling in the three-phase enhanced CT scan (step 701). At the same time, the spatial occupancy ratio of the necrotic volume at the corresponding time point is extracted, and it is found that the leading edge of the lesion in the arterial phase presents a characteristic combination of high weight parameters and rapid expansion. The coupling parameter and the spatial occupancy ratio at the same time point are dynamically superimposed (step 702) to generate a comprehensive parameter sequence showing a peak value in the venous phase. By analyzing the extreme value distribution of this sequence (step 703), the system establishes a dynamic parameter range reflecting the overall heterogeneity of the lesion. Comparing this range with the preset subtype classification standard (step 704), its range completely falls within the characteristic threshold range of micropapillary lung adenocarcinoma, and the system automatically outputs this classification result and indicates the possibility of EGFR sensitive mutation. When new cases present similar peak characteristics in the venous phase, clinical targeted treatment based on the system's suggestions achieves significant efficacy. This case optimizes the dynamic interval comparison algorithm, improves the recognition specificity of special subtype lung adenocarcinoma, and forms an intelligent decision-making closed loop from multi-parameter dynamic analysis to precise treatment.

[0105] In summary, steps 701 to 704 achieve the intelligent matching decision of the dynamic parameter range and the preset threshold. By superimposing the weight matching parameter and the spatio-temporal occupancy ratio to generate a comprehensive parameter and statistically analyzing its dynamic change range, a classification decision mechanism based on threshold comparison is constructed. By intelligently matching the dynamic parameter range with the preset threshold range, the automatic determination of lung cancer subtype classification is realized, solving the subjectivity problem of traditional manual interpretation. This technology significantly improves the stability and interpretability of the classification system, providing objective and reliable decision support for clinical diagnosis.

[0106] Figure 5 The following is a schematic structural diagram of a lung cancer type prediction system based on a fusion deep learning network provided by the embodiments of the present application, as Figure 5 shown, the system includes: An acquisition module 21 is configured to acquire CT images and blood samples of lung cancer patients, perform dynamic segmentation on the CT images to extract vascular distribution data of the tumor region, calculate three-dimensional density parameters and vascular fracture direction parameters based on the vascular distribution data, and acquire a volume ratio sequence of the tumor necrosis region; A processing module 22 is configured to perform spiral sorting on the blood samples, separate circulating tumor cells through fluid control, collect deformation trajectory data of the circulating tumor cells, and generate stiffness change parameters and deformation recovery parameters according to the deformation trajectory data; An analysis module 23 is configured to perform joint analysis on the three-dimensional density parameters, the vascular fracture direction parameters, and the stiffness change parameters, extract the time series offset of the vascular fracture direction parameters, and synchronously match the time series offset with the dynamic response of the stiffness change parameters to generate a synchronous parameter set; A establishment module 24 is configured to establish a lag correlation relationship between the time series offset of the vascular fracture direction parameters and the recovery period of the deformation recovery parameters based on the synchronous parameter set, so as to determine the weight matching result of the two through cross-modal analysis; An output module 25 is configured to output a lung cancer subtype classification result according to the weight matching result and the spatio-temporal distribution characteristics of the volume ratio of the tumor necrosis region.

[0107] Figure 5 The described lung cancer type prediction system based on a fusion deep learning network can execute Figure 1 The described lung cancer type prediction method based on a fusion deep learning network in the illustrated embodiment, and its implementation principle and technical effects will not be elaborated. For the lung cancer type prediction system based on a fusion deep learning network in the above embodiments, the specific manners in which each module and unit perform operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A lung cancer type prediction method based on a fusion deep learning network, characterized in that: include: Obtaining CT images and blood samples from lung cancer patients, dynamically segmenting the CT images to extract vascular distribution data of the tumor area, calculating three-dimensional density parameters and vascular fracture direction parameters based on the vascular distribution data, and obtaining a volume percentage sequence of the tumor necrosis area; Performing spiral sorting processing on the blood sample, separating circulating tumor cells by fluid control, collecting deformation trajectory data of the circulating tumor cells, and generating stiffness change parameters and deformation recovery parameters according to the deformation trajectory data; Jointly analyzing the three-dimensional density parameter, the blood vessel fracture direction parameter and the stiffness change parameter, extracting the time series offset of the blood vessel fracture direction parameter, and synchronously matching the time series offset with the dynamic response of the stiffness change parameter to generate a synchronous parameter set; Based on the synchronization parameter set, a hysteresis correlation relationship between the time series offset of the blood vessel fracture direction parameter and the recovery period of the deformation recovery parameter is established to determine a weight matching result of the two through cross-modal analysis; According to the weight matching result and the spatiotemporal distribution characteristics of the volume proportion of the tumor necrosis area, a lung cancer subtype classification result is output.

2. The method according to claim 1, characterized in that Dynamically segmenting the CT image to extract blood vessel distribution data of the tumor area, calculating three-dimensional density parameters and blood vessel fracture direction parameters based on the blood vessel distribution data, and obtaining a volume percentage sequence of the tumor necrosis area, including: Performing a multi-time point segmentation operation on the CT image, and extracting blood vessel distribution data of the tumor area at each time point, wherein the blood vessel distribution data includes the three-dimensional coordinates of blood vessel branch nodes and the connection status between adjacent nodes; Based on the three-dimensional coordinates of the blood vessel branch nodes, the density distribution of the blood vessels in the three-dimensional space is calculated, and the number of blood vessel branch nodes in a unit volume is defined as a three-dimensional density parameter; The fracture directions of all vascular branch nodes are counted, and the vascular fracture direction parameters are generated according to the angle distribution of the fracture directions in the three-dimensional coordinate system; The volume proportion series were generated by measuring the ratio of the volume of the tumor necrotic area to the total volume of the tumor area at each time point.

3. The method according to claim 1, characterized in that The blood sample is subjected to spiral sorting processing, circulating tumor cells are separated by fluid control, deformation trajectory data of the circulating tumor cells are collected, and stiffness change parameters and deformation recovery parameters are generated according to the deformation trajectory data, including: injecting the blood sample into a spiral sorting device to separate circulating tumor cells from non-tumor cells by adjusting the direction of fluid dynamics; Collecting movement trajectory data of the separated circulating tumor cells in the fluid environment, wherein the movement trajectory data includes the deformation amplitude and deformation direction of the cells at different time points; According to the curve of the deformation amplitude changing with time, the recovery speed after the deformation amplitude reaches the maximum value is calculated, and the reciprocal of the recovery speed is defined as the deformation recovery parameter; The frequency of change of the deformation direction in the moving trajectory is counted, and the product of the frequency and the deformation amplitude is defined as the stiffness change parameter.

4. The method according to claim 1, characterized in that: The three-dimensional density parameter, the blood vessel fracture direction parameter and the stiffness change parameter are jointly analyzed to extract the time series offset of the blood vessel fracture direction parameter, and the time series offset is synchronously matched with the dynamic response of the stiffness change parameter to generate a synchronous parameter set, including: Obtaining the direction angle values ​​of the blood vessel fracture direction parameter at multiple time points to calculate the angle difference between adjacent time points, and generating a time series offset according to the angle difference; Obtaining parameter values ​​of the stiffness variation parameters at the same time point, and extracting dynamic variation values ​​of the parameter values ​​over time; The time series offset and the dynamic change value are aligned on a time axis, and the sum of the products of the two at the same time point is calculated to generate a synchronization parameter set.

5. The method according to claim 1, characterized in that Based on the synchronization parameter set, a hysteresis correlation relationship between the time series offset of the blood vessel fracture direction parameter and the recovery period of the deformation recovery parameter is established to determine a weight matching result of the two through cross-modal analysis, including: Counting the time delay between the time point when the time series offset reaches a peak value and the end time point of the recovery period to generate a lag time series; Calculating the association weight between the time series offset and the recovery period according to the length distribution of the lagged time series; The association weight is superimposed with the sum of the products in the synchronization parameter set to generate a weight matching result.

6. The method according to claim 1, characterized in that Outputting lung cancer subtype classification results according to the weight matching result and the spatiotemporal distribution characteristics of the tumor necrosis area volume ratio includes: Extracting the distribution characteristics of the volume proportion values ​​in the volume proportion sequence as they change with spatial positions, and associating the spatial positions with the time points to generate spatiotemporal distribution characteristics; The weight matching result is mapped and matched with the spatiotemporal distribution feature, and the corresponding lung cancer subtype classification result is output according to a preset threshold interval.

7. The method according to claim 6, characterized in that Mapping and matching the weight matching result with the spatiotemporal distribution feature, and outputting the corresponding lung cancer subtype classification result according to a preset threshold interval, including: Obtaining the parameter value of each time point in the weighted matching result and the spatial position proportion value of the corresponding time point in the spatiotemporal distribution feature; Superimposing the parameter value and the proportion value at the same time point to generate a comprehensive parameter at each time point; Count the maximum and minimum values ​​of the comprehensive parameters at all time points to generate dynamic parameter intervals; The dynamic parameter interval is compared with a preset threshold interval. If the dynamic parameter interval completely falls within any threshold interval, the lung cancer subtype classification result corresponding to the threshold interval is output.

8. A lung cancer type prediction system based on a fusion deep learning network, characterized in that: include: an acquisition module, for acquiring CT images and blood samples of lung cancer patients, dynamically segmenting the CT images to extract vascular distribution data of the tumor area, calculating three-dimensional density parameters and vascular fracture direction parameters based on the vascular distribution data, and acquiring a volume percentage sequence of the tumor necrosis area; a processing module, configured to perform spiral sorting processing on the blood sample, separate circulating tumor cells by fluid control, collect deformation trajectory data of the circulating tumor cells, and generate stiffness change parameters and deformation recovery parameters according to the deformation trajectory data; an analysis module, for jointly analyzing the three-dimensional density parameter, the blood vessel fracture direction parameter and the stiffness change parameter, extracting a time series offset of the blood vessel fracture direction parameter, and synchronously matching the time series offset with a dynamic response of the stiffness change parameter to generate a synchronous parameter set; An establishing module, for establishing a hysteresis correlation relationship between the time series offset of the vascular rupture direction parameter and the recovery period of the deformation recovery parameter based on the synchronization parameter set, so as to determine a weight matching result of the two through cross-modal analysis; The output module is used to output the lung cancer subtype classification result according to the weight matching result and the spatiotemporal distribution characteristics of the volume proportion of the tumor necrosis area.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a lung cancer type prediction method based on a fusion deep learning network as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, a lung cancer type prediction method based on a fusion deep learning network as described in any one of claims 1 to 7 is implemented.

Images