Childhood chemotherapy post-platelet recovery evaluation system based on multi-source time-series data

Abstract

The application relates to a child chemotherapy post-platelet recovery evaluation system based on multi-source time sequence data, and relates to the technical fields of medical image processing and multi-modal time sequence data analysis. Medical image sequence and peripheral blood time sequence feature vectors in the plain scan period and the enhanced period are acquired; a microvessel gradient distribution matrix is constructed based on the time difference between images and the voxel intensity difference, and a recovery potential factor is extracted in combination with a physiological weight; an initial activation value and a candidate state are generated by using a time sequence evolution network to map the peripheral blood features, the activation value is reconstructed by using the recovery potential factor, and a real cell state vector is accumulated and output; blood index evolution results are obtained accordingly, and physiological weight coefficients are updated based on the confidence feedback optimization parameters. The application is favorable for breaking through the rigid assumption that static texture determines hematopoietic capacity, and can solve the technical problem that the platelet recovery period of a child with damaged microvessels is seriously overestimated.

Description

Technical Field

[0001] This application relates to the field of medical image processing and multimodal time series data analysis technology, specifically a platelet recovery assessment system for children after chemotherapy based on multi-source time series data. Background Technology

[0002] Thrombocytopenia is one of the most serious complications of chemotherapy for childhood malignant tumors. Accurately predicting the platelet recovery curve after chemotherapy is crucial for the rational use of thrombopoietin, the development of transfusion plans, and the assessment of the chemotherapy window period. Existing prediction systems fall into two categories: the first type is based on peripheral blood routine time-series data and uses regression or recurrent neural network fitting for prediction. Its fundamental limitation is that blood indicators are downstream outputs of hematopoietic activity, exhibiting biological lag and failing to anticipate the true hematopoietic potential of the bone marrow.

[0003] The second type of method incorporates static texture features from bone marrow plain MRI images and performs multimodal fusion by stitching them together with time-series blood data. This approach assumes that the static cell count determines hematopoietic kinetics, an assumption that is seriously flawed in the context of chemotherapy-induced damage. Chemotherapy drugs not only reduce the number of hematopoietic stem cells but also cause endothelial toxicity damage to the bone marrow microcirculatory vascular bed, a latent injury that cannot be detected by existing static imaging.

[0004] Therefore, existing systems may incorrectly apply the recovery rate in a healthy microenvironment, causing the platelet recovery time to be predicted three to five days earlier on average, resulting in false safety warnings and inducing premature discontinuation of medication, which may lead to bleeding risks. This problem of predictive error has been a long-standing bottleneck in the field, and existing improvements have not established a physical constraint between microvascular perfusion and the predicted rate at the mechanism level. Summary of the Invention

[0005] The purpose of this application is to provide a platelet recovery assessment system for children after chemotherapy based on multi-source time-series data, in order to address the technical defects of existing technologies that neglect perfusion gradient characteristics between plain and enhanced scan sequences and fail to break through the rigid assumption that static texture determines hematopoietic capacity, resulting in a serious overestimation of the platelet recovery period in children with microvascular damage.

[0006] The objective of this application can be achieved through the following technical solution: Firstly, a pediatric platelet recovery assessment system based on multi-source time-series data, comprising the following modules:

[0007] The acquisition module is used to acquire the plain scan medical image sequence, the contrast-enhanced medical image sequence, and the peripheral blood time-series feature vector of the target object;

[0008] The extraction module is used to construct a microvascular gradient distribution matrix characterizing the penetration ability of the target tissue area based on the acquisition time difference and spatial voxel intensity difference between the plain scan medical image sequence and the enhanced scan medical image sequence, and to extract recovery potential factors in combination with preset physiological weight coefficients.

[0009] The mapping module is used to map the hidden state features output by the preset temporal evolution network at the previous time step to the peripheral blood temporal feature vector at the current time step to generate the initial activation value and candidate internal state vector at the current time step.

[0010] The vector generation module is used to reconstruct the initial activation value based on the recovery potential factor to obtain the reconstructed target activation value, and to perform information accumulation operation on the candidate internal state vector in combination with a preset temporal evolution network to output the real cell state vector at the current time step.

[0011] The feedback optimization module is used to obtain the evolution results of blood indicators of the target object based on the real cell state vector, obtain feedback optimization parameters that characterize the confidence of the evolution results of blood indicators, and update the preset physiological weight coefficients based on the feedback optimization parameters.

[0012] Secondly, a method for assessing platelet recovery in children after chemotherapy based on multi-source time-series data includes the following steps:

[0013] Obtain the plain scan medical image sequence, the contrast-enhanced medical image sequence, and the peripheral blood temporal feature vector of the target object;

[0014] Based on the acquisition time difference and spatial voxel intensity difference between the plain scan medical image sequence and the enhanced scan medical image sequence, a microvascular gradient distribution matrix characterizing the penetration ability of the target tissue area is constructed, and recovery potential factors are extracted by combining preset physiological weight coefficients.

[0015] Based on the preset temporal evolution network, the hidden state features output by the previous time step are mapped to the peripheral blood temporal feature vector of the current time step to generate the initial activation value and candidate internal state vector of the current time step.

[0016] The initial activation value is reconstructed based on the recovery potential factor to obtain the reconstructed target activation value, and the candidate internal state vector is accumulated by combining the preset temporal evolution network to output the real cell state vector at the current time step.

[0017] Based on the real cell state vector, the evolution results of blood indicators of the target object are obtained, and feedback optimization parameters characterizing the confidence of the evolution results of blood indicators are obtained. The preset physiological weight coefficients are updated based on the feedback optimization parameters.

[0018] Thirdly, a computer storage medium storing computer-executable instructions, which, when executed, implement the pediatric platelet recovery assessment system based on multi-source time-series data described in the first aspect.

[0019] Compared with the prior art, the beneficial effects of this application are:

[0020] By extracting the voxel-level contrast agent wash-in slope between plain and enhanced images, a microvascular gradient distribution matrix is ​​constructed and mapped as a recovery potential factor. This breaks through the bottleneck of static image texture characterization of hematopoietic capacity, accurately quantifies the degree of damage to the bone marrow microcirculation vascular bed, eliminates the problem of premature prediction of platelet recovery in children with microvascular damage from a mechanistic perspective, and significantly reduces the risk of premature discontinuation of medication caused by false safety warnings.

[0021] Using the residual between the evolution results of blood indicators and the clinically confirmed values ​​as a basis, the gradient is backpropagated to the physiological weight coefficient of the recovery potential factor extraction end, and numerical truncation and normalization constraints are applied to achieve adaptive online calibration of image feature extraction parameters, enabling the system to learn to correctly interpret microvascular perfusion signals and improve its adaptability to different individual differences. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the module of the pediatric chemotherapy post-chemotherapy platelet recovery assessment system based on multi-source time-series data of this application;

[0023] Figure 2 This is a schematic diagram illustrating the steps of the method for assessing platelet recovery in children after chemotherapy based on multi-source time-series data, as described in this application. Detailed Implementation

[0024] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only to illustrate selected embodiments of this application.

[0025] Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item has been defined in one figure, it does not need to be further defined and explained in subsequent figures. The terms first, second, etc. are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] Thrombocytopenia is one of the most frequent and serious hematopoietic system complications during chemotherapy for childhood malignant tumors. While chemotherapy drugs kill tumor cells, they also have an unavoidable toxic effect on bone marrow hematopoietic stem cells and their hematopoietic microenvironment, leading to a sharp drop in peripheral blood platelet counts within one to three weeks after chemotherapy. This exposes children to the life-threatening risk of spontaneous bleeding and even intracranial hemorrhage. Accurately predicting the platelet recovery kinetics curve after chemotherapy is crucial for clinicians to make informed decisions regarding the timing of thrombopoietin administration, develop transfusion support plans, and assess the subsequent chemotherapy window. An early prediction will lead to premature withdrawal of hematopoietic support therapy, discontinuing thrombopoietin before platelets reach a safe level, resulting in severe bleeding events; a late prediction will lead to unnecessary prolongation of medication and wasted hospital stays. Therefore, a high-precision platelet recovery timing prediction system has irreplaceable clinical value for the safe management of pediatric chemotherapy.

[0027] Existing platelet recovery prediction systems can be summarized into two main technical approaches. The first approach focuses on pure time-series modeling, collecting peripheral blood routine test data at different time points after chemotherapy in children, including indicators such as platelet count, white blood cell count, hemoglobin concentration, and reticulum platelet count. Statistical regression models or recurrent neural networks are then used to fit and extrapolate the temporal evolution trajectories of these indicators for prediction. The fundamental limitation of this approach is that peripheral blood indicators are essentially downstream outputs of bone marrow hematopoietic activity, and they have an inherent biological lag effect relative to the actual state of the hematopoietic microenvironment within the bone marrow. Therefore, time-series data of blood indicators alone cannot proactively perceive the true hematopoietic potential of the bone marrow.

[0028] The second type of method incorporates bone marrow medical imaging data to improve prediction accuracy. It supplements peripheral blood time-series data with static texture features from bone marrow plain MRI images, such as gray-level co-occurrence matrices, wavelet energy coefficients, and signal intensity histograms. The image feature vectors are then concatenated with the blood time-series data and fed into a multimodal fusion prediction model. The significance of this approach lies in introducing morphological information from the bone marrow into the prediction system, to some extent filling the gap in the perception of hematopoietic organ status by purely time-series methods. However, the core technical assumptions of these methods contain a serious pathophysiological fallacy—they assume that the absolute static gray-level signal or cell density texture of bone marrow plain MRI images is sufficient to comprehensively characterize the current and future hematopoietic cell recovery capacity, i.e., that cell mass determines hematopoietic kinetic energy. This assumption will lead to systematic and directional prediction biases in specific chemotherapy-induced pathological scenarios.

[0029] From a clinical pathological perspective, the toxic damage of chemotherapy drugs to the bone marrow is not limited to the direct reduction of hematopoietic stem cell numbers; it also causes irreversible endothelial toxicity to the bone marrow microcirculatory vascular bed. The bone marrow microvascular network is a crucial physical foundation for hematopoietic stem cell homing, proliferation, and differentiation. Its structural integrity and perfusion function directly determine whether hematopoietic stem cells can efficiently convert their proliferative potential into detectable actual cell production in the peripheral blood. When a child's bone marrow microcirculatory vascular bed is damaged due to multiple cycles of high-intensity chemotherapy, even if the static cell count in the bone marrow cavity appears relatively normal according to plain CT imaging, the actual hematopoietic kinetic energy will be significantly reduced due to a severe decrease in microenvironment perfusion capacity—a hidden dimension of damage that current technology cannot perceive based on static image texture features.

[0030] Faced with the aforementioned pathological scenarios, existing prediction systems based on static texture features completely ignore bone marrow microvascular perfusion dynamics. When encountering children with conventional chemotherapy-induced damage, where static cell counts are still acceptable but microcirculation is severely impaired, these systems incorrectly apply the hematopoietic recovery rate expected in a healthy microenvironment. This results in the system predicting platelet recovery time three to five days earlier than the actual clinical observations, leading to false safety warnings and potentially inducing clinicians to prematurely discontinue thrombopoietin, withdrawing hematopoietic support before platelets have recovered to safe levels, thus increasing the risk of severe bleeding. This predictive error is a long-standing bottleneck in time-series prediction distortion in this field. Existing technological improvements have not fundamentally established a physical constraint relationship between microvascular perfusion status and time-series prediction rates, thus failing to eliminate this systemic bias at the mechanistic level.

[0031] Therefore, such as Figure 1 As shown, this application provides a system for assessing platelet recovery in children after chemotherapy based on multi-source time-series data, including the following modules:

[0032] The acquisition module is used to acquire the plain scan medical image sequence, the contrast-enhanced medical image sequence, and the peripheral blood time-series feature vector of the target object;

[0033] The extraction module is used to construct a microvascular gradient distribution matrix characterizing the penetration ability of the target tissue area based on the acquisition time difference and spatial voxel intensity difference between the plain scan medical image sequence and the enhanced scan medical image sequence, and to extract recovery potential factors in combination with preset physiological weight coefficients.

[0034] The mapping module is used to map the hidden state features output by the preset temporal evolution network at the previous time step to the peripheral blood temporal feature vector at the current time step to generate the initial activation value and candidate internal state vector at the current time step.

[0035] The vector generation module is used to reconstruct the initial activation value based on the recovery potential factor to obtain the reconstructed target activation value, and to perform information accumulation operation on the candidate internal state vector in combination with a preset temporal evolution network to output the real cell state vector at the current time step.

[0036] The feedback optimization module is used to obtain the evolution results of blood indicators of the target object based on the real cell state vector, obtain feedback optimization parameters that characterize the confidence of the evolution results of blood indicators, and update the preset physiological weight coefficients based on the feedback optimization parameters.

[0037] This invention employs a three-layer architecture design: perfusion gradient extraction, microenvironment factor mapping, and temporal gating coupling. Instead of simply concatenating bone marrow imaging features as ordinary static data vectors with peripheral blood time-series data, the system follows a causal logic paradigm that uses the physical constraints of image perfusion dynamics to predict the temporal prediction rate. It decouples the microvascular damage information from the image space into a purely scalar constraint factor, and then directly intervenes in the state transition mechanism within the pre-defined temporal evolution network of the peripheral blood time-series data using multiplicative coupling, forcibly limiting its state accumulation rate to be consistent with physiological reality.

[0038] The entire data flow strictly follows the causal closed-loop principle: the acquisition module obtains multi-source benchmark data; the extraction module extracts recovery potential factors from the spatiotemporal gradient of images to quantify the hematopoietic potential of the bone marrow microenvironment; the mapping module and the vector generation module complete the internal state update of the peripheral blood temporal feature vector under the rigid constraint of the recovery potential factors; and the feedback optimization module uses clinical blood test results as a supervision signal to update the preset physiological weight coefficients of the image feature extraction end in a closed loop, continuously improving the system's individualized adaptability.

[0039] I. Data Acquisition Module – Multi-Source Data Acquisition;

[0040] The acquisition module is responsible for obtaining multi-source baseline data of the target group (i.e., pediatric patients receiving chemotherapy), covering three core data categories:

[0041] The first is the plain scan medical imaging sequence, which is a three-dimensional magnetic resonance imaging sequence of the bone marrow region acquired without the injection of contrast agent. This sequence reflects the basic signal intensity distribution of the bone marrow cavity and serves as the intensity baseline for subsequent calculation of the contrast agent washout slope. The second is the enhancement-phase medical imaging sequence, which is a three-dimensional magnetic resonance imaging sequence of the bone marrow acquired after intravenous injection of magnetic resonance contrast agent (such as gadolinium contrast agent) and waiting for a specific delay time. The degree to which the contrast agent penetrates into the bone marrow interstitium through microvessels directly reflects the local microvascular density and permeability. The third is the peripheral blood temporal feature vector, which is a sequence of peripheral blood routine test indicators sampled discretely at preset time steps along the chemotherapy cycle. It includes multi-dimensional blood indicators such as platelet count, white blood cell count, and hemoglobin concentration at each time point. After standardization and feature alignment processing, it is constructed as the peripheral blood temporal feature vector of the current time step.

[0042] Plain CT and contrast-enhanced CT medical imaging sequences must be acquired within the same chemotherapy cycle and from the same imaging position to ensure that they have a consistent anatomical reference coordinate system. The acquisition module extracts the precise scan timestamps from the medical digital imaging (DICOM) header files of the two sets of sequences and calculates the contrast agent washout delay time of the contrast-enhanced CT medical imaging sequence relative to the plain CT medical imaging sequence. This time difference will be used as the physical time denominator when the subsequent extraction module constructs the microvascular gradient distribution matrix.

[0043] II. Extraction Module – Acquisition of Target Tissue Regions;

[0044] The extraction module first needs to define the effective spatial range for subsequent microvascular gradient calculations. Because the bone marrow cavity contains a mixed distribution of functional red bone marrow and non-hematopoietic yellow bone marrow (fatty bone marrow), if perfusion gradient statistics are performed on all bone marrow voxels indiscriminately, the extremely low perfusion signal in the yellow bone marrow region will severely dilute and mask the effective perfusion characteristics of active red bone marrow areas, leading to distorted microenvironment assessment. Therefore, the system must first accurately locate the target tissue region, i.e., the range of active red bone marrow with actual hematopoietic activity.

[0045] The acquisition of the target tissue region is performed in two steps. The first step is spatial coordinate alignment: using the plain CT scan medical image sequence as the spatial reference coordinate benchmark, a three-dimensional elastic deformation registration algorithm based on the mutual information maximization criterion is used to perform non-rigid spatial correction on the contrast-enhanced medical image sequence to eliminate spatial misalignment deviations caused by respiratory movements and slight changes in body position between the two scans. After alignment, the plain CT scan medical image sequence generates a first three-dimensional data matrix, and the contrast-enhanced medical image sequence generates a second three-dimensional data matrix. The two achieve precise correspondence of spatial coordinates at the voxel (three-dimensional pixel) level, meaning that the voxel intensity values ​​at corresponding spatial coordinate points can be directly subtracted.

[0046] The second step is the segmentation and extraction of active hematopoietic regions: The first three-dimensional data matrix is ​​input into a preset tissue segmentation network. This network learns the differences in signal intensity and spatial distribution between red and yellow bone marrow in bone marrow plain scan images and outputs a binary mask matrix with the same spatial resolution as the original image. In this binary mask matrix, the three-dimensional spatial coordinates of active hematopoietic red bone marrow are marked with a first preset value (value 1), while the coordinates of non-hematopoietic yellow bone marrow, cortical bone, and other non-target tissues are marked with non-first preset values ​​(value 0). The system marks all three-dimensional spatial coordinate sets in the binary mask matrix with element values ​​of the first preset value as target tissue regions, and extracts all spatial voxel points corresponding to the first and second three-dimensional data matrices within these target tissue regions.

[0047] III. Extraction Module – Construction of Microvascular Gradient Distribution Matrix;

[0048] The microvascular gradient distribution matrix is ​​the core intermediate data structure of this invention. Its physical essence is a three-dimensional distribution map of the contrast agent wash-in rate per unit time at each spatial voxel point in the bone marrow target tissue region, which directly quantifies the perfusion activity of the bone marrow microcirculation vascular bed at each spatial location.

[0049] The initial washout slope of contrast agent entering local bone marrow tissue is strictly limited by the capillary network density and microvascular endothelial permeability at that spatial location. When chemotherapy drugs cause damage to the bone marrow microvascular endothelium, the contrast agent penetration capacity of the damaged area decreases significantly, and its washout slope is significantly lower than that of normal bone marrow areas. Therefore, voxel-level washout slope is the most direct and sensitive imaging proxy feature for revealing latent damage to bone marrow microcirculation, and this is precisely the information dimension that current methods using only static texture features on flat scans cannot capture at all.

[0050] The construction process of the microvascular gradient distribution matrix is ​​as follows: First, the physiological delay time from contrast agent injection to enhanced scan acquisition is extracted from the DICOM header file of the enhanced medical image sequence and determined as the acquisition time difference. The time difference is used to convert the intensity difference between two images into the physical denominator of the wash-in rate per unit time. Subsequently, for each target voxel point within the target tissue region, the voxel intensity values ​​at the spatial coordinates of that voxel point are extracted from the second 3D data matrix and the first 3D data matrix. The difference between the two values ​​is calculated to obtain the spatial voxel intensity difference for that target voxel point, characterizing the signal gain caused by the contrast agent at that location. Finally, the spatial voxel intensity difference is divided by the acquisition time difference, and a ratio calculation is performed to obtain the wash-in slope value corresponding to that target voxel point.

[0051] ;

[0052] in, Represents the target voxel point The corresponding wash-in slope value (i.e., microvascular perfusion gradient). This represents the voxel intensity value at the target voxel point in the second three-dimensional data matrix (enhancement phase); This represents the voxel intensity value at the target voxel point in the first three-dimensional data matrix (flat scan period); This indicates the time difference in data collection.

[0053] After sequentially performing the above-mentioned wash-in slope calculation on all target voxel points within the target tissue region, the wash-in slope values ​​of all voxel points are assembled according to their corresponding three-dimensional spatial coordinates to generate a microvascular gradient distribution matrix consistent with the spatial resolution of the target tissue region. Each element in this matrix precisely quantifies the microvascular perfusion activity of the corresponding spatial micro-unit within the bone marrow, constituting the core physical characterization basis for revealing latent bone marrow damage.

[0054] IV. Extraction Module – Extraction of Recovery Potential Factors;

[0055] The recovery potential factor is a key step in this invention to reduce the dimensionality of the three-dimensional microvascular gradient distribution matrix to a single scalar physical constraint coefficient. It serves as a bridge data connecting image perfusion information and the internal state evolution mechanism of the temporal prediction network. Its value is in the (0,1) range. The closer the recovery potential factor is to 1, the more complete the overall perfusion capacity of the microvascular network within the target tissue region, and the more robust the physical foundation for rapid platelet recovery in the child's bone marrow. Conversely, the closer the recovery potential factor is to 0, the more severe the damage to bone marrow microcirculation, and the more severely the hematopoietic rate is forcibly inhibited by physiological mechanisms.

[0056] The extraction of recovery potential factors is carried out in three steps. The first step is the removal of large vessel interference values: all element values ​​in the microvascular gradient distribution matrix are extracted, and elements exceeding a preset quantile threshold are removed to obtain the effective gradient value set. The physical basis for this removal is that there are a small number of macroscopic large veins in the bone marrow cavity, with extremely high contrast agent perfusion volumes, which will produce extremely high washout slopes far exceeding those of the microvascular capillary bed. If these extreme values ​​are not removed, they will seriously interfere with the representativeness of subsequent statistics, leading to misjudgments of the overall microcirculation status. The preset quantile threshold can be set to the 95th percentile, that is, removing the top 5% of large vessel interference data points in terms of washout slope.

[0057] The second step is statistical dimensionality reduction and extraction: Two key statistics are calculated from the effective gradient numerical set: the median, which serves as a central tendency parameter, characterizing the concentration of overall microvascular perfusion intensity in the bone marrow within the target tissue region; and the absolute median difference, which serves as a divergence parameter, quantifying the spatial heterogeneity of microvascular damage within the target tissue region, i.e., the unevenness of damage levels in different areas of the bone marrow. Both the median and the absolute median difference are nonparametric statistics robust to extreme values, exhibiting higher stability and representativeness compared to the mean and standard deviation when strong local signal interference exists in bone marrow imaging.

[0058] The third step is interval mapping: An interval mapping operation is performed on the weighted difference between the central trend parameter and the divergence parameter. Specifically, the central trend parameter is weighted by a first preset physiological weight coefficient, then the weighted divergence parameter by a second preset physiological weight coefficient is subtracted. The resulting linear combination value is then input into a preset smoothing activation function, mapping it to the (0,1) interval to generate a recovery potential factor. The preset smoothing activation function can be the Sigmoid function, whose range is (0,1) and is continuously differentiable everywhere, satisfying the mathematical requirements of subsequent gradient backpropagation.

[0059] ;

[0060] in, This represents the recovery potential factor, with values ​​ranging from (0,1). Indicates the preset smooth activation function; This represents the first preset physiological weight coefficient, used to apply a positive contribution weight to the central trend parameter; This represents the second preset physiological weight coefficient, which is used to apply a negative penalty weight to the divergence parameter. Both of these are preset physiological weight coefficients. This represents the central trend parameter (i.e., the median value of the effective gradient set). This represents the divergence parameter (i.e., the absolute median difference of the set of effective gradient values).

[0061] V. Mapping Module – Generation of Initial Activation Values ​​and Candidate Internal State Vectors;

[0062] The mapping module uses a preset temporal evolution network as the computational carrier to perform nonlinear mapping between the peripheral blood temporal feature vector of the current time step and the hidden state features of the previous time step, thereby generating the initial activation value and candidate internal state vector of the current time step.

[0063] The initial activation value physically represents the proportion of surface recovery kinetic energy tendency estimated by the preset temporal evolution network based on the fluctuation of peripheral blood indicators. The candidate internal state vector represents the new information content that may need to be written into long-term memory obtained by mapping the peripheral blood temporal feature vector in the high-dimensional feature space. Together, they constitute the basic raw materials for the subsequent vector generation module to perform information accumulation operation.

[0064] The pre-defined temporal evolution network employs a Long Short-Term Memory (LSTM) architecture, internally containing two independent sets of weight matrices, one for the initial activation value and the other for the candidate internal state vectors. The first set of weight matrices corresponds to input gating calculations, including the mapping weight matrix of the peripheral blood temporal feature vector at the current time step, the mapping weight matrix of the hidden state features at the previous time step, and corresponding bias parameters. The second set of weight matrices corresponds to candidate state calculations, also containing an independent set of mapping weight matrices for the current peripheral blood features, the mapping weight matrix for historical hidden states, and corresponding bias parameters. These two sets of weight matrices are optimized independently during model training and are not shared, ensuring the decoupling of gating ratio and candidate content calculations.

[0065] The process of generating the initial activation value is as follows: Based on the first set of weight matrices, matrix multiplication and addition are performed on the peripheral blood time series feature vector of the current time step and the hidden state feature of the previous time step, respectively. The results of the two matrix operations are added together to obtain the first superposition result. The first superposition result is input into the first preset activation function (usually the Sigmoid function) for nonlinear transformation, and the initial activation value with a value in the range of (0,1) is output. Its physical meaning is the proportion of information to be written into long-term memory expressed by the peripheral blood time series feature vector of the current time step.

[0066] The process of generating candidate internal state vectors is as follows: Based on the second set of weight matrices, independent matrix multiplication and addition are performed on the peripheral blood time series feature vector of the current time step and the hidden state feature of the previous time step, respectively. The results of the two matrix operations are added together to obtain the second superposition result. The second superposition result is input into the second preset activation function (using the hyperbolic tangent function) for nonlinear transformation, and the candidate internal state vector with a value in the interval (-1,1) is output. Its physical meaning is the candidate content of new information that may be written into long-term memory expressed by the peripheral blood time series feature vector of the current time step in the feature space.

[0067] It should be noted that the initial activation values ​​generated at this stage only represent the surface recovery kinetic tendency derived by the preset temporal evolution network based on the fluctuation of surface blood indicators. They have not been verified by the physical state of the bone marrow microenvironment and do not have real clinical decision reference value. They must be transferred to the vector generation module to accept the physical constraint reconstruction of the recovery potential factors.

[0068] VI. Vector Generation Module – Reconstruction of Target Activation Values;

[0069] The vector generation module is the core innovative algorithm of this invention. Its first task is to perform multiplicative reconstruction on the initial activation value output by the mapping module based on the recovery potential factor as a dynamic constraint, to obtain the reconstructed target activation value. The target activation value is the proportion of the true recovery kinetic tendency after being forcibly verified by the physical state of bone marrow microenvironment perfusion, and it serves as an effective write control switch for subsequent information accumulation operations.

[0070] The reconstruction process of the target activation value consists of two steps. The first step is the generation of the dynamic penalty multiplier: the recovery potential factor is multiplied by a preset control constant, and the product is used as the adjustment base. This adjustment base is then input into the hyperbolic tangent function, and the resulting output value is the dynamic penalty multiplier. (Preset control constant) This is a positive curvature control parameter used to adjust the mapping curvature and response sensitivity of the hyperbolic tangent function within the range of the recovery potential factor. The larger the value, the steeper the growth of the dynamic penalty multiplier during the process of the recovery potential factor increasing from 0 to 1, and the more sensitive the system is to microcirculatory damage.

[0071] The hyperbolic tangent function has a range of (0,1) and exhibits monotonically increasing, bounded smooth saturation characteristics, ensuring the numerical stability of the dynamic penalty multiplier. More importantly, when the recovery potential factor γ approaches 0, the adjustment base... Approaching 0, the hyperbolic tangent function has a value of 0 at 0, the dynamic penalty multiplier approaches 0, achieving complete suppression; when When it approaches 1, the adjustment base approaches 1. The hyperbolic tangent function approaches The dynamic penalty multiplier approaches the preset upper limit, resulting in minimal suppression. The second step involves multiplying the initial activation value by the corresponding element of the dynamic penalty multiplier: the initial activation value output by the mapping module is multiplied element-wise by the dynamic penalty multiplier to obtain the reconstructed target activation value.

[0072] ;

[0073] in, This represents the target activation value after reconstruction; Indicates the initial activation value; This represents the preset control constant; Indicates a factor of recovery potential; This represents the dynamic penalty multiplier, whose range is [0,1].

[0074] The causal effectiveness of the above multiplicative coupling mechanism is demonstrated in the following scenarios: when bone marrow microcirculation is severely impaired ( When the value is →0, even if the platelet-related indicators in the peripheral blood time-series feature vector show a temporary false increase due to recent blood transfusions or intravenous fluid resuscitation, the initial activation value calculated by the mapping module may be large. However, after the multiplicative suppression of the dynamic penalty multiplier, the reconstructed target activation value will be mathematically forced down to a level close to 0. This means that no matter how much the peripheral blood indicators show a false increase, the effective write control switch inside the preset time-series evolution network is turned off, fundamentally blocking the optimistic prediction bias based on false blood signals.

[0075] VII. Vector Generation Module – Output of Real Cell State Vectors;

[0076] The second task of the vector generation module is to combine the historical memory retention mechanism of the pre-set temporal evolution network to perform a physically constrained information accumulation operation on the candidate internal state vectors, and output the real cell state vector at the current time step. The real cell state vector is the long-term memory carrier of the pre-set temporal evolution network. In a clinical physical sense, it corresponds to the actual accumulated hematopoietic kinetic energy in the hematopoietic system of the child from the start of chemotherapy to the current time point, and is the direct source for subsequent decoding and generation of blood index evolution results.

[0077] The output process of the real cell state vector is divided into two computational branches: basic retention and restricted addition. The results of the two branches are then added and merged. In the basic retention branch, the system obtains the forgetting activation parameter, which represents the tendency to retain historical information at the current time step, from the output of the preset temporal evolution network, as well as the real cell state vector from the previous time step. The corresponding elements of the two are multiplied to obtain the basic retention information vector, which represents the part of historical hematopoietic kinetic energy that is considered still effective and retained and passed on to the current time step after considering the changes in the current blood state.

[0078] In the restricted new branch, the target activation value output by the first task of the vector generation module is multiplied element-wise with the candidate internal state vector to obtain the current new information vector, representing the incremental hematopoietic energy that can be written into long-term memory at the current time step. Since the target activation value has been forcibly suppressed by the recovery potential factor, the magnitude of this incremental amount is rigidly constrained by the bone marrow microenvironment perfusion state. Finally, the basic retained information vector is added to the current new information vector to generate the true cell state vector at the current time step.

[0079] ;

[0080] in, Indicates the current time step The true cell state vector; Indicates the previous time step The true cell state vector; The forgetting activation parameter (representing the tendency to retain historical information at the current time step, generated by calculation within the preset temporal evolution network) represents the forgetting activation parameter. This indicates the target activation value (which has been restored through multiplicative reconstruction of the potential factor, and its value is affected by...). (Constraint and suppression) Represents the candidate's internal state vector;

[0081] In the above formula, the second term This refers to the newly added information vector in the current period, the core of which lies in the target activation value. Factors that have been restored Suppression: When →0 (severely impaired microcirculation), Approaching 0, the newly added information vector in the current period approaches the zero vector, and the real cell state vector inside the preset temporal evolution network can hardly accumulate new positive values, thus suppressing the platelet recovery slope; when →1 (when microcirculation is healthy and intact), Approaching The newly added information vectors are written according to the normal evolution rate of peripheral blood indicators, and the platelet recovery prediction curve can rise normally.

[0082] In addition, the pre-defined temporal evolution network also uses an output gating mechanism at each time step to generate hidden state features of the current time step by nonlinear filtering the real cell state vector, and temporarily stores them in a temporary cache as the hidden state features of the previous time step when the mapping module of the next time step performs calculations, thus ensuring the logical closed loop of the temporal recursion link.

[0083] VIII. Feedback Optimization Module – Obtaining Blood Index Evolution Results and Feedback Optimization Parameters;

[0084] The first task of the feedback optimization module is to decode and output the evolution results of blood indicators of the target object from the real cell state vector. The real cell state vector of the current time step is input into a preset linear decoding network, which performs a mapping process from a high-dimensional feature space to a one-dimensional blood concentration numerical space, outputting a continuous sequence of predicted blood indicator concentration values ​​as the blood indicator evolution result. The blood indicator evolution result is a sequence of predicted concentration values ​​arranged along the time axis, where each time point corresponds to a predicted platelet concentration value. The entire sequence together constitutes a dynamic evolution prediction curve of the child's platelet count over the next few days.

[0085] After obtaining the evolution results of blood indicators, clinically confirmed values ​​obtained from actual laboratory blood tests at each time point corresponding to the target subject are retrieved from the clinical database. The predicted results are compared with the clinically measured values, and the residual function value representing the prediction deviation is calculated and determined as the feedback optimization parameter. The residual function adopts the mean squared error method, and the arithmetic mean of the sum of squares of the differences between the predicted values ​​and the clinically confirmed values ​​at all time points is used as the quantification result.

[0086] ;

[0087] in, This represents the residual function value used as a feedback optimization parameter; This represents the total number of time points in the evolution of blood indicators; The result of the evolution of blood indicators is the first Blood index concentration values ​​at each time point; Indicates the first Clinically confirmed values ​​at each time point.

[0088] IX. Feedback Optimization Module – Update of Preset Physiological Weight Coefficients;

[0089] The second task of the feedback optimization module is to use the prediction error information carried by the residual function value when the feedback optimization parameter is not zero, and to backpropagate the gradient signal to the first and second preset physiological weight coefficients of the recovery potential factor calculation formula in the extraction module through the chain rule, so as to perform parameter update and adaptively optimize the extraction strategy of bone marrow microvascular image features.

[0090] This adaptive update mechanism has profound clinical and physical significance: the first pre-set physiological weight coefficient determines the positive contribution weight of the overall microvascular perfusion intensity of the bone marrow (central trend parameter) to the recovery potential factor; the second pre-set physiological weight coefficient determines the negative penalty weight of the heterogeneity of bone marrow microcirculation damage (divergence parameter) to the recovery potential factor. By backpropagating the final platelet recovery prediction error to the above coefficients, the model is taught in reverse how to correctly weigh the relative importance of perfusion intensity and spatial heterogeneity in bone marrow images of different children using real clinical blood test results—this is an adaptive imaging parameter calibration mechanism that uses an effect-to-cause approach, overcoming the individualization deficiencies of traditional manual empirical calibration.

[0091] The update process of the preset physiological weight coefficients is as follows: First, based on the feedback optimization parameters, the preset temporal evolution network is traversed using the chain rule to calculate... Regarding the current first preset physiological weight coefficient partial derivative gradient value The gradient value indicates how the current value of the first preset physiological weight coefficient should be adjusted in the direction of reducing prediction bias.

[0092] Then, the partial derivative gradient value is compared with the preset learning factor. The product of these is used as a negative compensation term, and is combined with the current first preset physiological weight coefficient. Addition calculations are performed to obtain candidate update values. Finally, boundary anti-expansion truncation is applied to the candidate update values. Numerical constraints are imposed on the candidate values ​​using preset minimum and maximum first preset physiological weight coefficients, and the updated first preset physiological weight coefficients are output. By complementary normalization constraints The updated second preset physiological weight coefficient is directly calculated. The sum of the two remains constant at 1.

[0093] The engineering significance of introducing boundary truncation constraints lies in the fact that, in training scenarios with anomalous samples, gradient explosion can cause weight coefficients to fly beyond their physically reasonable range. Truncation can completely prevent such numerical inflation failures, ensuring the numerical stability of the system under extreme conditions. The physical significance of complementary constraints lies in the fact that their sum is always 1, ensuring that the total amount of the trade-off between the two types of information remains constant when extracting the recovery potential factor. The system only needs to optimize one degree of freedom, which simplifies the computational overhead and improves the search stability of the parameter space.

[0094] ;

[0095] in, This represents the updated first preset physiological weight coefficient; This represents the current first preset physiological weight coefficient; Indicates the preset learning factor; This represents the minimum allowed first preset physiological weight coefficient; This represents the maximum allowed first preset physiological weight coefficient; This represents the updated second preset physiological weight coefficient.

[0096] In another implementation, such as Figure 2 As shown, this application also provides a method for assessing platelet recovery in children after chemotherapy based on multi-source time-series data, including the following steps:

[0097] Obtain the plain scan medical image sequence, the contrast-enhanced medical image sequence, and the peripheral blood temporal feature vector of the target object;

[0098] Based on the acquisition time difference and spatial voxel intensity difference between the plain scan medical image sequence and the enhanced scan medical image sequence, a microvascular gradient distribution matrix characterizing the penetration ability of the target tissue area is constructed, and recovery potential factors are extracted by combining preset physiological weight coefficients.

[0099] Based on the preset temporal evolution network, the hidden state features output by the previous time step are mapped to the peripheral blood temporal feature vector of the current time step to generate the initial activation value and candidate internal state vector of the current time step.

[0100] The initial activation value is reconstructed based on the recovery potential factor to obtain the reconstructed target activation value, and the candidate internal state vector is accumulated by combining the preset temporal evolution network to output the real cell state vector at the current time step.

[0101] Based on the real cell state vector, the evolution results of blood indicators of the target object are obtained, and feedback optimization parameters characterizing the confidence of the evolution results of blood indicators are obtained. The preset physiological weight coefficients are updated based on the feedback optimization parameters.

[0102] In another embodiment, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the aforementioned pediatric chemotherapy post-chemotherapy platelet recovery assessment system based on multi-source time-series data.

[0103] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A system for assessing platelet recovery in children after chemotherapy based on multi-source time-series data, characterized in that, Includes the following modules: The acquisition module is used to acquire the plain scan medical image sequence, the contrast-enhanced medical image sequence, and the peripheral blood time-series feature vector of the target object; The extraction module is used to construct a microvascular gradient distribution matrix characterizing the penetration ability of the target tissue area based on the acquisition time difference and spatial voxel intensity difference between the plain scan medical image sequence and the enhanced scan medical image sequence, and to extract recovery potential factors in combination with preset physiological weight coefficients. The mapping module is used to map the hidden state features output by the previous time step of the preset temporal evolution network to the peripheral blood temporal feature vector of the current time step to generate the initial activation value and candidate internal state vector of the current time step. The vector generation module is used to reconstruct the initial activation value based on the recovery potential factor to obtain the reconstructed target activation value, and to perform information accumulation operation on the candidate internal state vector in combination with a preset temporal evolution network to output the real cell state vector at the current time step. The feedback optimization module is used to obtain the evolution results of blood indicators of the target object based on the real cell state vector, obtain feedback optimization parameters that characterize the confidence of the evolution results of blood indicators, and update the preset physiological weight coefficients based on the feedback optimization parameters.

2. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy according to claim 1, characterized in that, The process of obtaining the target tissue region includes: The enhanced phase medical image sequence of the target object is spatially aligned with the plain phase medical image sequence to obtain a first three-dimensional data matrix and a second three-dimensional data matrix after spatial coordinate alignment. The first three-dimensional data matrix is ​​input into a preset tissue segmentation network, and a binary mask matrix representing the range of active hematopoietic red bone marrow is output. The set of three-dimensional spatial coordinates in which the element values ​​are the first preset values ​​is marked as the target tissue region. Based on the binary mask matrix, the set of spatial voxel points corresponding to the first three-dimensional data matrix and the second three-dimensional data matrix in the target tissue region is extracted.

3. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy, as described in claim 2, is characterized in that... The process of constructing the microvascular gradient distribution matrix includes: The contrast agent washout delay time of the enhanced phase medical image sequence relative to the plain phase medical image sequence is obtained as the acquisition time difference. For each target voxel in the spatial voxel point set, the spatial voxel intensity difference between the second three-dimensional data matrix and the first three-dimensional data matrix at the corresponding position is obtained. The ratio of the spatial voxel intensity difference to the acquisition time difference is calculated to obtain the wash-in slope value corresponding to each target voxel point. A microvascular gradient distribution matrix is ​​generated based on the wash-in slope values ​​of all target voxel points in the target tissue region. ； This represents the ingress slope value of the target voxel. and These represent the voxel intensity values ​​of the second three-dimensional data matrix and the first three-dimensional data matrix at the corresponding target voxel points, respectively. This indicates the time difference in data collection.

4. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy, as described in claim 3, is characterized in that... The process of extracting recovery potential factors includes: Extract all element values ​​from the microvascular gradient distribution matrix and remove elements that are greater than a preset quantile threshold to obtain an effective gradient value set. Obtain the median value and absolute median difference value of the effective gradient value set and use them as the central tendency parameter and divergence parameter, respectively. For the central trend parameter With the divergence parameter The weighted difference is used to perform an interval mapping operation to generate a recovery potential factor whose value is in the interval (0,1). ； This represents the recovery potential factor. This indicates the preset smooth activation function. and These represent the first preset physiological weight coefficient and the second preset physiological weight coefficient, respectively.

5. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy according to claim 1, characterized in that, The process of generating initial activation values ​​and candidate internal state vectors includes: The preset temporal evolution network has a first set of weight matrices for input gating calculation and a second set of weight matrices for candidate state calculation. Based on the first set of weight matrices, matrix multiplication and addition are performed on the peripheral blood time series feature vector of the current time step and the hidden state feature of the previous time step to obtain the first superposition result. The first superposition result is then input into the first preset activation function to output the initial activation value. Based on the second set of weight matrices, matrix multiplication and addition are performed on the peripheral blood time series feature vector of the current time step and the hidden state feature of the previous time step to obtain the second superposition result. The second superposition result is then input into the second preset activation function to output the candidate internal state vector.

6. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy according to claim 1, characterized in that, The process of obtaining the reconstructed target activation value includes: The recovery potential factor is multiplied by a preset control constant to generate an adjustment base, and the adjustment base is input into the hyperbolic tangent function to generate a dynamic penalty multiplier for limiting the rate of numerical ascent. The initial activation value is multiplied element-wise by the dynamic penalty multiplier to obtain the reconstructed target activation value. ,in, This represents the initial activation value. This represents the preset control constant. This represents the recovery potential factor.

7. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy according to claim 6, characterized in that, The process of outputting the true cell state vector includes: Obtain the forgetting activation parameter, which represents the tendency to retain historical information at the current time step, and the real cell state vector at the previous time step from the output of the preset temporal evolution network. Perform element-wise multiplication on the two to obtain the basic retained information vector. The target activation value is multiplied element-wise with the candidate internal state vector to obtain the newly added information vector in the current period, and the real cell state vector at the current time step is generated based on the basic retained information vector and the newly added information vector in the current period. ； and These represent the true cell state vectors corresponding to the current time step and the previous time step, respectively. This represents the forgetting activation parameter. This represents the target activation value. This represents the candidate internal state vector.

8. The pediatric platelet recovery assessment system based on multi-source time-series data after chemotherapy, as described in claim 4, is characterized in that... The process of obtaining feedback to optimize parameters and updating preset physiological weight coefficients includes: The real cell state vector is input into a preset linear decoding network for mapping processing, and the continuously predicted blood index concentration values ​​are output as the blood index evolution results. The clinically confirmed values ​​corresponding to each time point in the blood index evolution results are obtained, and the residual function value representing the prediction bias is obtained as the feedback optimization parameter. ; ； This represents the total number of time points in the evolution results of the blood indicators. This indicates the first of the blood index evolution results. Blood index concentration values ​​at various time points, Indicates the first Clinically confirmed values ​​at each time point; When the feedback optimization parameter is not zero, the current first preset physiological weight coefficient is adjusted accordingly. The first preset physiological weight coefficient is updated by comparing it with the second preset physiological weight coefficient. With the second preset physiological weighting coefficient ; Indicates the preset learning factor. This represents the gradient value of the partial derivative of the feedback optimization parameters with respect to the current first preset physiological weight coefficient. and These represent the minimum and maximum allowed first preset physiological weight coefficients, respectively.

9. A method for assessing platelet recovery in children after chemotherapy based on multi-source time-series data, characterized in that, Includes the following steps: Obtain the plain scan medical image sequence, the contrast-enhanced medical image sequence, and the peripheral blood temporal feature vector of the target object; Based on the acquisition time difference and spatial voxel intensity difference between the plain scan medical image sequence and the enhanced scan medical image sequence, a microvascular gradient distribution matrix characterizing the penetration ability of the target tissue area is constructed, and recovery potential factors are extracted in combination with preset physiological weight coefficients. Based on the preset temporal evolution network, the hidden state features output by the previous time step are mapped to the peripheral blood temporal feature vector of the current time step to generate the initial activation value and candidate internal state vector of the current time step. The initial activation value is reconstructed based on the recovery potential factor to obtain the reconstructed target activation value, and the candidate internal state vector is accumulated by combining the preset temporal evolution network to output the real cell state vector at the current time step. Based on the real cell state vector, the evolution results of blood indicators of the target object are obtained, and feedback optimization parameters characterizing the confidence of the evolution results of blood indicators are obtained. The preset physiological weight coefficients are updated based on the feedback optimization parameters.

10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement the pediatric platelet recovery assessment system based on multi-source time-series data as described in any one of claims 1-8.

Images