Construction method of tissue three-dimensional physical property navigation map, control method of double-channel perfusion system and perfusion system

Abstract

The application discloses a method for constructing a three-dimensional physical property navigation map of tissue, a control method of a double-channel perfusion system and a double-channel adaptive perfusion system. The method comprises: obtaining a cell space adjacency graph representing the microstructure relationship of cells in a digitized pathological image; identifying critical response singular points by applying virtual disturbance to the graph and analyzing the structural metric response thereof; and constructing a three-dimensional physical property navigation map containing predicted blood vessel physical parameters based on the singular points. The control method comprises: controlling a perfusion pump to output a tentative fluid pulse signal through a first channel and collecting a pressure response curve; calculating actual blood vessel physical parameters and comparing the actual blood vessel physical parameters with predicted parameters in the navigation map; and when the comparison satisfies a matching condition, controlling a channel switching valve to switch to a second channel to perform liquid delivery, and dynamically adjusting the output flow rate through closed-loop pressure compensation control. The system comprises a tissue modeling module, a double-channel perfusion unit, a sensing unit and an adaptive control unit.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence in medicine, automated control technology for medical devices, and medical image processing technology. In particular, it relates to a precision perfusion system and method for treating liver cancer that combines histopathological microstructure analysis with real-time physical parameter feedback during interventional treatment. Background Technology

[0002] Hepatocellular carcinoma (HCC) is a common type of primary liver cancer worldwide, and its diagnosis and treatment outcomes heavily rely on early and accurate identification of cancerous tissue and highly effective, low-toxicity targeted therapy. At the diagnostic level, histopathological examination is the gold standard, identifying cancer cells by observing stained tissue sections under a microscope. However, this method faces challenges in diagnostic consistency when dealing with early-stage, well-differentiated cancers. With the development of computational pathology, the analysis of digital pathological images using artificial intelligence technologies such as graph neural networks (GNNs) has become a research hotspot. By modeling cells as nodes in a graph, higher-order tissue structure patterns can be learned. However, most existing GNN-based methods remain within a static descriptive paradigm, directly extracting and classifying features from tissue atlases, and their sensitivity is limited by the degree of variation in cell morphology itself. These static analysis methods lack an active detection mechanism to stimulate and amplify the intrinsic differences in structural stability between normal and cancerous tissues, and fail to translate abstract pathological findings into quantitative parameters that can guide physical therapy.

[0003] In terms of treatment, transarterial chemoembolization (TACE) is one of the main interventional treatments for intermediate and advanced liver cancer. Traditional TACE procedures heavily rely on the physician's experience, and the perfusion process often uses manual injection or an electric infusion pump with a fixed flow rate, which is a blind injection mode. The vascular network inside the tumor is usually tortuous, rigid, and highly heterogeneous. During perfusion, the fluid resistance within the blood vessels can change drastically due to minute displacement of the catheter tip, vasospasm, or the gradual filling of the embolic agent. Existing non-feedback devices cannot detect these changes. If the resistance is abnormally high and drug is still forcibly injected, it may lead to vascular rupture or drug reflux into normal organs such as the gastrointestinal tract, causing serious complications. If the resistance is abnormally low (such as the catheter slipping into a large blood vessel), the chemotherapy drug will be rapidly diluted by the blood flow, failing to form an effective concentration in the tumor area. Therefore, how to go beyond static pathological diagnosis and blind physical perfusion, and develop a diagnostic and therapeutic approach that combines microscopic tissue structure analysis with real-time physical feedback of the macroscopic perfusion process to achieve an integrated approach from accurate prediction to adaptive treatment, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that, in the prior art, it is difficult to convert the results of histopathological analysis into quantitative parameters for interventional treatment, and perfusion equipment lacks adaptive control capabilities based on real-time physical feedback. The invention provides a method for constructing a three-dimensional physical property navigation map of tissue, a control method for a dual-channel perfusion system, and a dual-channel adaptive perfusion system.

[0005] In a first aspect, this application provides a method for constructing a three-dimensional physical property navigation map of a tissue, comprising: acquiring a cell spatial adjacency map characterizing cell nuclei and their spatial adjacency relationships in a digital pathological image, wherein the cell spatial adjacency map includes multiple nodes and edges connecting the nodes, each node corresponds to a cell nucleus and is associated with an initial node feature vector characterizing the quantitative features of the cell nucleus; defining a global structural metric characterizing the structural stability of the cell spatial adjacency map; iteratively applying a conformational stress sequence to at least one seed node in the cell spatial adjacency map, the conformational stress sequence being a progressively increasing perturbation vector; during the iterative application process, using a pre-trained graph neural network model, calculating the response value of the global structural metric determined by the node embedding output by the graph neural network model after each application of the perturbation vector, to generate a global structural metric response curve; identifying a critical response singularity in the global structural metric response curve, and constructing a three-dimensional physical property navigation map characterizing a tissue region based on the critical response singularity, the three-dimensional physical property navigation map including a predicted vascular physical parameter at one or more locations within the region.

[0006] Optionally, obtaining the cell spatial adjacency map includes: preprocessing the digitized pathological image and segmenting all cell nucleus instances in the image; defining each cell nucleus instance as a node in the cell spatial adjacency map; for each node, extracting at least one quantized feature of the corresponding cell nucleus to generate the initial node feature vector; and determining the connection relationship between the nodes based on the spatial distance relationship between the cell nucleus instances to generate the edges.

[0007] Optionally, the global structure metric is the global structure entropy, which is calculated based on the Shannon entropy of the cosine similarity distribution between all node embedding vectors output by the graph neural network model and their mean embedding vector.

[0008] Optionally, constructing the three-dimensional physical property navigation map based on the critical response singularity includes: calculating the deviation of the embedding vector of each node from the system mean embedding vector at the iteration number corresponding to the critical response singularity; and generating the predicted vascular physical parameters of the region where each node is located based on the mapping relationship between the deviation and vascular physical properties.

[0009] Optionally, the predicted vascular physical parameters include parameters selected from at least one of the following: predicted dynamic fluid impedance, predicted vascular wall compliance time constant.

[0010] Secondly, this application also provides a control method for a dual-channel perfusion system, applied to a perfusion system including a perfusion pump, a channel switching valve, and a pressure sensor, comprising: controlling the perfusion pump to output an exploratory fluid pulse signal through a first channel, and acquiring a pressure response curve triggered by the exploratory fluid pulse signal in real time through the pressure sensor; calculating an actual vascular physical parameter based on the pressure response curve; comparing the actual vascular physical parameter with a predicted vascular physical parameter at the corresponding position in a three-dimensional physical property navigation map constructed by structural perturbation analysis of a cell spatial adjacency map based on a digital pathological image; when the comparison result meets a preset matching condition, controlling the channel switching valve to switch from the first channel to the second channel, and controlling the perfusion pump to perform drug delivery through the second channel; during the drug delivery process, performing closed-loop pressure compensation control based on the real-time feedback of the pressure sensor to dynamically adjust the output flow rate of the perfusion pump.

[0011] Optionally, the actual vascular physical parameters and the predicted vascular physical parameters include parameters selected from at least one of the following groups: dynamic fluid impedance and vascular wall compliance time constant; wherein, the dynamic fluid impedance is calculated based on a peak pressure, a reference pressure, and a preset flow rate of the trial fluid pulse signal of the pressure response curve; and the vascular wall compliance time constant is determined based on a function fit of the pressure response curve to a pressure decay segment after the peak pressure.

[0012] Optionally, the preset matching condition is: the difference between the actual vascular physical parameters and the predicted vascular physical parameters is less than a preset matching threshold.

[0013] Optionally, the closed-loop pressure compensation control includes: acquiring the actual pressure value in the flow path in real time through the pressure sensor; and dynamically adjusting the output flow rate of the injection pump through the second channel based on the error between the actual pressure value and a preset target pressure upper limit using a proportional-integral-derivative control algorithm; wherein the preset target pressure upper limit is determined based on the peak pressure measured in the pressure response curve of the exploratory fluid pulse signal.

[0014] Optionally, after the drug delivery is completed, the control method further includes: controlling the channel switching valve to switch from the second channel back to the first channel; controlling the infusion pump to deliver a preset volume of flushing fluid through the first channel, the preset volume matching the dead cavity volume of the infusion system.

[0015] Secondly, this application also provides a dual-channel adaptive perfusion system, comprising: a tissue modeling module for acquiring a cell spatial adjacency map characterizing a digital pathological image, and constructing a three-dimensional physical property navigation map containing predicted vascular physical parameters by applying virtual perturbation to the map and analyzing its structural response; a dual-channel perfusion unit, including a perfusion pump and a channel switching valve connected to a first channel and a second channel flow path; a sensing unit for acquiring a pressure response curve in real time when the perfusion pump outputs an exploratory fluid pulse signal through the first channel; and an adaptive control unit configured to: calculate an actual vascular physical parameter based on the pressure response curve; compare the actual vascular physical parameter with the predicted vascular physical parameter at the corresponding position in the three-dimensional physical property navigation map; control the channel switching valve to switch from the first channel to the second channel to perform drug delivery based on the comparison result; and, during drug delivery, perform a closed-loop pressure compensation control based on the real-time feedback of the sensing unit to adaptively adjust the output flow rate of the perfusion pump.

[0016] Optionally, the tissue modeling module includes: an image preprocessing submodule for color normalization and cell nucleus instance segmentation of the digitized pathological image; an atlas construction submodule for constructing the cell spatial adjacency atlas based on the spatial distance relationships of the segmented cell nucleus instances; and a graph neural network inference submodule for identifying critical response singularities through a structural perturbation titration process and generating the three-dimensional physical property navigation map based on the critical response singularities.

[0017] Optionally, the sensing unit includes a pressure sensor with a sampling rate of not less than 1000Hz, used to acquire the pressure transient response triggered by the exploratory fluid pulse signal.

[0018] Optionally, the adaptive control unit has a built-in proportional-integral-derivative (PI-DI) control algorithm module, which is configured to dynamically adjust the drive signal of the infusion pump based on the error between the real-time pressure value collected by the sensing unit and a preset target pressure upper limit.

[0019] Optionally, the adaptive control unit is further configured to: after the drug delivery is completed, control the channel switching valve to switch back to the first channel; and control the infusion pump to deliver a preset volume of flushing fluid through the first channel to perform a tube sealing operation.

[0020] The construction method of this invention combines graph neural networks with structural perturbation titration, which can extract quantitative features reflecting the structural stability of the tissue microenvironment from static two-dimensional pathological images and transform them into three-dimensional navigation parameters that can be used to guide physical intervention operations, thus realizing a technological leap from abstract pathological analysis to operable physical parameter prediction.

[0021] The control method of this invention establishes a position verification mechanism based on physical characteristic matching by comparing the predicted and actual measured vascular physical parameters in the perfusion flow path, and adopts millisecond-level PID closed-loop pressure compensation control, enabling the perfusion system to dynamically adjust the output flow rate according to the real-time pressure feedback in the flow path, thereby improving the safety and accuracy of the perfusion process.

[0022] The perfusion system of this invention integrates a tissue modeling module, a dual-channel perfusion unit, a sensing unit, and an adaptive control unit, achieving hardware integration from navigation map construction to adaptive perfusion control, and providing a complete equipment solution for clinical applications. Attached Figure Description

[0023] Figure 1 This is a flowchart of a liver cancer tissue structure modeling method based on a graph neural network according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the process of constructing a cell spatial adjacency map according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a structure perturbation titration process according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the global structural metric response curve and critical response singularity identification according to an embodiment of the present invention.

[0027] Figure 5 This is a schematic diagram of the overall structure of a dual-channel adaptive perfusion system for liver cancer according to an embodiment of the present invention.

[0028] Figure 6 This is a control flowchart of an adaptive perfusion method according to an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0032] Example 1 This embodiment details a perfusion method for liver cancer treatment that combines histopathological microstructure analysis with real-time intraoperative physical parameter feedback. The core concept of this method lies in constructing a closed-loop technology from microstructure prediction to macroscopic physical verification, and then to adaptive perfusion execution. The execution process is divided into two main stages: the first stage involves constructing a preoperative three-dimensional physical property navigation map, which utilizes a graph neural network model to perform in-depth analysis of digital pathological sections to predict the vascular physical properties of the tumor region; the second stage involves the execution of real-time feedback adaptive perfusion during surgery, where, guided by the navigation map, target points are verified and intelligent drug perfusion is performed by measuring intravascular physical parameters in real time.

[0033] Phase 1: Construction of the 3D Physical Property Navigation Map (S100-S550) The goal of this stage is to transform a static two-dimensional digital pathological image into a three-dimensional navigation model that can be used to guide physical interventional surgery. The core information of this model is the quantitative prediction of the physical characteristics of the vascular network in different regions inside the tumor.

[0034] S100: Obtain a cell spatial adjacency map that characterizes cell nuclei and their spatial adjacency relationships in a digital pathological image.

[0035] In one specific implementation, this method uses a digital whole-slide scanner to acquire a whole-slide digital image (WSI) of liver cancer tissue, achieving a resolution of 0.25 micrometers / pixel at 40x magnification. The acquired raw image data is used to construct a structured graph capable of capturing cellular entities and their spatial topological relationships, namely a Cellular Spatial Adjacency Graph (CSAG). This graph is denoted as... ,in It is a set of nodes representing the cell nucleus. It is the set of edges representing the adjacency relationships between cells, and It is a matrix that stores the initial features of all nodes.

[0036] This step can be further subdivided into the following sub-steps: S110: Preprocess the digitized pathological images.

[0037] The input WSI images are huge in size and often exhibit color variations due to differences in staining batches and scanning instruments. To ensure the stability and reproducibility of the analysis, image standardization is necessary. A specific processing workflow involves first dividing the WSI into multiple image patches that are easier to process.

[0038] For example, from a WSI image with a resolution of 100,000 × 80,000 pixels, a Region of Interest (ROI) of 2048 × 2048 pixels is randomly selected or selected based on the tissue region. For this ROI, a color normalization algorithm is applied, such as the Macenko method proposed in the reference "A Method for Normalizing Histology Slides for Quantitative Analysis". This method separates the staining vectors of hematoxylin and eosin through singular value decomposition and maps the color space of the source image to a predefined target color space. This process eliminates the variability caused by variations in staining depth. Subsequently, to suppress high-frequency noise while preserving the edge details of cell nuclei, a Gaussian filter is used to smooth the image. The kernel size of this Gaussian filter can be set to 5 × 5 pixels, with a standard deviation of [missing value]. The value was set to 1.5 (dimensionless). After preprocessing, the color reference of the image was unified, creating conditions for subsequent accurate cell nucleus segmentation.

[0039] S120: Segment all cell nucleus instances in the image and define each cell nucleus instance as a node in the cell space adjacency graph.

[0040] This sub-step aims to identify and separate each individual cell nucleus from the preprocessed image. Given that cell nuclei in liver cancer tissue are often densely packed, overlapping, and morphologically diverse, this application employs a deep learning-based instance segmentation model.

[0041] For example, a 2048×2048 pixel image patch processed by S110 is input into a pre-trained StarDist model. StarDist is an instance segmentation model designed specifically for star-shaped objects (such as cell nuclei), which performs segmentation by predicting the star-shaped polygon distance and object probability for each pixel. The model outputs two parts of information: a binary segmentation mask, which is a matrix of the same size as the input image, where pixels belonging to any cell nucleus have a value of 1 and background pixels have a value of 0; and the contour coordinates of each identified cell nucleus instance. By performing post-processing operations such as non-maximum suppression on the model output, all independent cell nucleus instances in the image patch can finally be resolved. Suppose that a total of 1048×2048 pixel images are identified in this image patch. The cell nucleus. For the first For each cell nucleus, its centroid coordinates can be obtained by calculating the geometric center of its masked region. This coordinate will serve as the spatial identifier for this cell nucleus. For example, the first... The centroid coordinates of each cell nucleus were calculated as follows: (Unit: pixels). This Each cell nucleus instance together constitutes a set of nodes in the atlas. .

[0042] S130: For each node, extract at least one quantitative feature of the corresponding cell nucleus to generate the initial node feature vector.

[0043] In order for a graph neural network to understand the intrinsic properties of each node, a multi-dimensional numerical feature vector needs to be constructed for each node. This vector aims to comprehensively quantify the intrinsic properties of the cell nucleus. In a preferred embodiment, the feature vector is composed of three types of features: morphological features, texture features, and color intensity features.

[0044] Morphological features describe the geometry and size of the cell nucleus, which is crucial for distinguishing between regularly shaped normal cell nuclei and cancer cell nuclei of varying sizes and irregular shapes. Specific features calculated include: nuclear area, obtained by calculating the number of pixels within its masked region and converting it to square micrometers (µm²) according to image resolution; nuclear perimeter, i.e., the length of the mask outline, in micrometers (µm); the ratio of the major and minor axes, obtained by elliptical fitting of the masked region, and is a dimensionless value; and circularity, defined as... An ideal circle has a circularity of 1 (dimensionless); and a set of seven Hu invariant moments that are invariant to translation, rotation, and scaling.

[0045] Texture features describe the distribution pattern of chromatin within the cell nucleus. The chromatin in cancer cell nuclei often appears coarse and uneven. This application uses a Gray-Level Co-occurrence Matrix (GLCM) to extract texture features. A GLCM is a statistical matrix that describes the frequency of pixel pairs with specific gray values ​​appearing in an image under specific spatial relationships. Based on GLCM, a series of statistics can be calculated, such as contrast, dissimilarity, homogeneity, energy, and correlation. These features are all dimensionless values.

[0046] Color intensity characteristics reflect the staining properties of the cell nucleus. In standard H&E staining, due to increased DNA content, cancer cell nuclei typically exhibit deeper basophilia (i.e., bluer). This application calculates statistical measures, such as mean and standard deviation, of pixel values ​​in each color channel within the cell nucleus mask region in multiple color spaces, including RGB, HSV, and Lab, as color intensity characteristics.

[0047] For example, for the first Each cell nucleus, after calculation, has an initial node feature vector. Let it be a 24-dimensional vector. Assume its morphological feature vector is... (Units are µm², µm, dimensionless, dimensionless, etc.); its texture feature vector is (All are dimensionless); its color intensity feature vector is (These are all pixel intensity values, ranging from 0 to 255). Concatenate these three vectors and normalize them using the L2 norm to obtain the final 24-dimensional initial node feature vector. .all The feature vectors of each node together constitute the initial feature matrix. .

[0048] S140: Based on the spatial distance relationship between the cell nucleus instances, determine the connection relationship between the nodes to generate the edges.

[0049] Edges of the graph Encoding intercellular interactions and spatial adjacency relationships is central to tissue structure. This application employs a hybrid strategy combining fixed radius and K-Nearest Neighbors (KNN) to construct edges. This strategy ensures graph connectivity and adapts to varying cell densities in different regions.

[0050] For example, setting a radius threshold Micrometer, nearest neighbor number For any node in the graph First, find all the points whose spatial distance from its centroid is within a certain range. Other nodes within the range And establish connections (edges) with them. This radius can be biologically interpreted as the range within which a cell can directly influence its surrounding environment. Assume nodes... If a node has 6 neighbors within a 50-micrometer range, then an edge is established with these 6 neighbors. Additionally, to prevent nodes in sparsely populated regions from becoming isolated, extra calculations are performed on the nodes. Find the nearest distance to all other nodes in the graph. Each node has 8 neighbors. If any of these 8 neighbors are new neighbors that are not within the aforementioned 6 radii, then edges are also established with these new neighbors. This ensures that the degree (number of connected edges) of each node is at least 8. Optionally, each edge can be... Assign a weight The strength of the connection is represented by a weight function, which can be a distance-based Gaussian kernel. ,in It is a node and Euclidean distance between the centroids It is a hyperparameter that controls the rate of distance decay; for example, it can be set to... .

[0051] At this point, the cellular spatial adjacency map (CSAG) has been constructed.

[0052] S200: Define a global structural metric to characterize the structural stability of the cell spatial adjacency map.

[0053] This step introduces a quantifiable scalar metric to assess the orderliness or stability of the entire organizational graph. This metric is highly sensitive to changes in the distribution of node embeddings in the feature space after graph neural network processing. In this embodiment, this metric is specifically implemented as Global Structural Entropy (GSE).

[0054] Entropy is a measure of system uncertainty in information theory. Here, global structural entropy is used to quantify the consistency of the behavior of all node embedding vectors in a graph relative to their collective center. The core idea is that in a structurally stable and highly homogeneous organization (such as normal liver tissue), the node embeddings formed after information aggregation by a GNN should exhibit strong clustering and regularity, with a relatively concentrated distribution, thus corresponding to a lower entropy value. Conversely, in a structurally unstable and highly heterogeneous organization (such as tissue containing cancerous regions), the node embeddings will be more dispersed and disordered, thus corresponding to a higher entropy value.

[0055] The calculation of global structural entropy is based on the set of all node embedding vectors output by the GNN model under a specific perturbation state. The calculation process includes three core steps: Centrality metric: Calculates a center position for all node embeddings, i.e., the mean embedding vector. This vector represents the average state or centroid of the entire tissue map in the feature space under the current state.

[0056] Similarity distribution construction: Calculate the embedding of each node With mean embedding vector The similarity between them. Cosine similarity is used here:

[0057] Cosine similarity focuses on the consistency of vector directions, and its value ranges from [value missing]. It can effectively eliminate the impact of differences in node embedding module length.

[0058] Entropy calculation: similarity values ​​of all nodes The distribution is discretized, and then its Shannon entropy is calculated. Specifically, the distribution is discretized. The range is evenly divided into There are several bins. The statistics are based on whether a value falls within each bin. Number of similarity values And calculate its probability. Ultimately, the global structural entropy... Through formula Calculated. When When, define .

[0059] For example, suppose a map contains After a certain perturbation calculation, 1000 node embedding vectors were obtained. First, their mean vectors were calculated. Then, calculate the embedding of each node with... The cosine similarity was used to obtain 1000 ranges within the range. The similarity value within. The interval is divided into Each bin is of equal width. After analysis, it was found that 600 similarity values ​​fell into [a specific category / area]. Within this range, 200 values ​​fell into... The remaining 200 values ​​are scattered across other intervals. The probabilities of these two main intervals are respectively... and Because the distribution is highly concentrated, the calculated global structural entropy... It will be a relatively small value (in bits), which reflects the high degree of orderliness of the system in this state.

[0060] S300: Apply a conformational stress sequence iteratively to at least one seed node in the cell space adjacency map.

[0061] The core idea of ​​this step is to simulate a virtual stress on the tissue microenvironment by injecting a small, controllable, and continuously increasing external signal (i.e., conformational stress) into the spectral model, and then observe the response of the entire system. This process is defined as structural perturbation titration (SPT).

[0062] The conformational stress sequence is a sequence that varies with the number of iterations. Evolutionary perturbation vector sequence Each perturbation vector The dimension of the perturbation vector is the same as the dimension of the initial feature vector of the node, so that it can be directly applied to the node features. In a specific embodiment, the perturbation vector... The generation method is designed to be linear and unidirectional, and its formula is: .

[0063] This formula contains three key components: Step size factor : A small positive constant that controls the increment of the perturbation strength in each iteration. Its value determines the precision of the titration process; a smaller value means finer perturbation growth. For example, It can be set to 0.001 (dimensionless).

[0064] Number of iterations : An integer starting at 0 and incrementing, representing the time of the titration process or the degree of accumulation of disturbances. Maximum number of iterations. It needs to be set large enough to ensure that the disturbance accumulates to a point where it triggers a critical response in the system. For example, It can be set to 1000.

[0065] Unit direction vector The vector determines the direction of the perturbation and has a definite physical meaning in the feature space. To effectively detect carcinogenic potential, the vector... It is designed to point in a direction related to cancerous features. One specific implementation involves using a labeled dataset to calculate the average initial feature vectors for all cancer cell nodes and normal cell nodes, respectively. and Then, the unit direction vector. Defined as a direction vector pointing from the normal state to the cancerous state, and then normalized: .

[0066] After defining the conformational stress sequence, it is necessary to select the object to which the perturbation is applied, i.e., at least one seed node. Choosing a seed node instead of applying the perturbation globally is to simulate the occurrence and development of a local lesion.

[0067] For example, in a containing In the graph of 10 nodes, 10 nodes located in the central region of the image are selected as the seed node set. .

[0068] Next, the iterative application process is executed. This process starts from... Start, repeat until In each iteration step First, the perturbation vector for the current step is calculated. Then, update the valid initial feature vector of the seed node. For each seed node... Its effective initial feature vector in the current step Updated to the sum of the original eigenvector and the perturbation vector: For all non-seed nodes, their feature vectors remain unchanged.

[0069] S400: During the iterative application process, a pre-trained graph neural network model is used to calculate the response value of the global structure metric determined by the node embedding output by the graph neural network model after each application of the perturbation vector, so as to generate a global structure metric response curve.

[0070] In S300 for each iteration step Both generated a perturbed initial feature matrix. Next, this step uses a pre-trained GNN model as a physics simulator to calculate how this perturbation propagates throughout the graph and affects all nodes, and ultimately quantifies the change in global structural metric.

[0071] In one specific embodiment, the pre-trained model is a model containing A graph convolutional network (GCN) with [number] layers. The update rule for each layer is as follows: ,in It is an adjacency matrix with self-loops added. yes The degree matrix, It is the first Layer node embedding matrix, It is the first The trainable weight matrix of the layer. The weight matrix of this model. The GNN has been trained on a large amount of labeled data in the offline phase, and its parameters are frozen during the S400 titration process. This ensures that the behavior of the GNN is consistent and deterministic, and the change in its output is entirely due to the perturbation of the input. Driven by.

[0072] At each step of the iterative process (from arrive The following calculation process will be executed: The perturbation feature matrix of the current step As input to GNN.

[0073] The forward propagation process of the GNN is performed. Information (including initial features and applied perturbations) is transmitted between nodes through the edges of the graph. The passing and convergence of wheels. Through After layer propagation, the final node embedding matrix is ​​obtained. In the embedding vector of each node All of them contain their own initial characteristics and their position in the map up to The structure and characteristics of hop neighbors.

[0074] Based on the obtained final node embedding matrix Calculate the global structural entropy defined in S200. .

[0075] The calculated global structural entropy value With the current iteration step (or the corresponding disturbance intensity) As a data point Write it down.

[0076] This process from (Starting from a undisturbed state) Repeat execution. This process will eventually result in a sequence of data points. Plot these data points on a graph with the number of iterations... The horizontal axis represents the global structural entropy. In a coordinate system with the vertical axis as the ordinate, a global structural response curve is formed.

[0077] S500: Identify a critical response singularity in the global structure metric response curve.

[0078] The response curves visually demonstrate how the structural stability of a tissue pattern evolves with the continuous enhancement of external perturbations. A healthy, stable tissue structure may exhibit a curve that grows gradually over a long range of perturbations. However, a tissue containing inherently unstable regions (such as cancerous areas) may experience a sharp, non-linear jump in its curve at a certain perturbation intensity. This point is the phase transition point where the system structure undergoes a qualitative change, i.e., the Critical Response Singularity (CRS) defined in this invention.

[0079] To accurately locate this singularity from discrete data points, this application employs the method of analyzing the derivative of the curve. Since the curve is composed of discrete data points, the derivative is approximated using differences.

[0080] The first step is to calculate the first difference (approximate first derivative) of the response curve: ,for This sequence represents the slope or rate of increase of the curve. Near the singularity, It will reach a local maximum.

[0081] The second step, in order to more robustly capture the points where the slope changes most drastically, is to calculate the second difference (approximately the second derivative): ,for The second-order difference describes the concavity or convexity of the curve. When the curve changes from a gradual increase (concave) to a rapid increase (convex), the second-order difference reaches a significant peak. The number of iterations corresponding to this peak point... This point, where the system undergoes its most drastic change, is thus identified as the critical response singularity. Therefore, the number of iteration steps corresponding to the critical response singularity... It was identified as: .

[0082] For example, suppose the obtained response curve data shows that, arrive between, The value rapidly increased from 3.5 bits to 5.0 bits. By calculating the second-order difference, it was found that... hour, The value exhibited a sharp peak significantly higher than all other points. Therefore, the critical response singularity was determined to occur at iteration step number [number missing]. .

[0083] S550: Based on the critical response singularity, construct a three-dimensional physical property navigation map characterizing the liver cancer tissue region, the navigation map containing a predicted vascular physical parameter at one or more locations within the region.

[0084] The core task of this step is to transform the abstract critical response singularity information identified in the S500 into quantitative physical parameter predictions that can be used to guide actual surgery.

[0085] First, at the critical moment At the moment of system collapse, those inherently unstable nodes that cause system instability (i.e., the cancerous core cells and their microenvironment) will exhibit the most aberrant behavior. This behavior can be calculated in... At time t, the embedding vector of each node Relative to the system mean embedding vector deviation To quantify: . The larger the value of a node, the greater its contribution to system instability; that is, the more unstable its structure.

[0086] Next, this invention establishes a mapping model to account for the microstructural instability of nodes (caused by...). The metric is correlated with the macroscopic vascular physical properties of the region where the node is located. Pathological knowledge indicates that tumor angiogenesis lacks a smooth muscle layer, resulting in rigid vessel walls, poor resilience, and a chaotic network structure, which manifests as high resistance and low compliance in fluid dynamics. Low compliance implies rapid pressure decay, i.e., a small time constant. Therefore, the following mapping relationship can be constructed: Predicting dynamic fluid resistance ( ): Structural instability of nodes It is positively correlated with the degree of vascular disorder in its region, and thus positively correlated with fluid resistance. The mapping formula can be set as follows: .

[0087] Predicted compliance time constant ( ): Structural instability of nodes It is positively correlated with the vascular stiffness of its region, and thus negatively correlated with the compliance time constant. The mapping formula can be set as follows: .

[0088] in, and These are the baseline physical parameters for blood vessels in normal liver tissue. and These are calibrable mapping coefficients.

[0089] Finally, all deviations Greater than a preset threshold (e.g., all) Nodes with the mean plus twice the standard deviation of the values ​​were identified as core cancerous nodes. The three-dimensional spatial coordinates of these nodes (obtained through registration with CT / MRI and other imaging data) and their corresponding predicted physical parameters were then analyzed. , By integrating these elements, a three-dimensional physical property navigation map is constructed.

[0090] For example, assuming the reference impedance of a normal hepatic artery time constant Calibration coefficient , (Dimensionless). For a given quantity in... Calculate high deviation at all times The core nodes of cancer The predicted vascular physical parameters for its location are:

[0091]

[0092] And for a deviation that is very low ( stable nodes Its predicted parameters are close to normal values. This navigation map provides clear and quantitative targets for intraoperative validation.

[0093] Phase Two: Real-time Feedback Adaptive Injection Execution (S600-S900) This stage is performed in the interventional operating room, using a dual-channel adaptive perfusion system to complete precise drug perfusion of the tumor under the guidance of the navigation map generated in the first stage.

[0094] S600: System self-test and reference flow resistance calibration.

[0095] This step is performed before inserting the perfusion catheter into the patient and is intended to initialize and calibrate the perfusion system.

[0096] S610: Tubing pre-filling. After the system is powered on, the microcontroller unit (MCU) controls the three-way solenoid valve to first connect the first channel (channel A, loaded with saline) and drives a peristaltic pump to run at a constant low speed (e.g., 2 ml / min) to fill the entire tubing with saline until the tip of the microcatheter, expelling all air.

[0097] S620: Static pressure acquisition. When the pump stops rotating and the microcatheter tip is outside the body or placed in a large blood vessel such as the abdominal aorta, the pressure sensor acquires the current hydrostatic reference value. This step is used to eliminate inherent pressure errors caused by differences in drug bag height, atmospheric pressure fluctuations, and the length of the catheter itself. For example, data is collected... .

[0098] S630: Dead space volume setting. Depending on the type of microcatheter connected (e.g., 2.7F, 150cm), the dead space volume of the tubing can be preset in the system or manually entered. This parameter is crucial for subsequent accurate cannulation. For example, setting... .

[0099] S700: Performs exploratory fluid pulse detection and calculates actual vascular physical parameters.

[0100] This step is initiated when the physician places the tip of the microcatheter near the target feeding artery shown on the navigation map using image guidance, in order to detect the actual physical characteristics of that location in real time.

[0101] S710: Detection command transmission and transient response acquisition. The operator issues a detection command, the MCU controls the valve to remain in channel A (physiological saline), and instructs the pump to execute a short square wave pulse. Specific parameters are: flow rate set to... (180 ml / min), pulse duration 200 ms. During this period, the pressure sensor, with a high sampling rate of 1000 Hz, completely records the entire process of pressure rise, peak hold, and decline caused by the pulse within the blood vessel, generating a pressure-time response curve. .

[0102] S720: Calculation of Dynamic Fluid Impedance ( The MCU's internal algorithm module utilizes formulas. Calculate dynamic fluid impedance. For example, from the acquired data... The peak pressure was read from the curve. Using the S620's reference pressure and known pulse flow rate The calculation yields: .

[0103] S730: Calculation of the time constant of vascular wall compliance ( The algorithm module supports... The curve was analyzed during the pressure decay segment after the pulse ended, and the exponential decay formula was fitted. Determine the time constant For example, by performing least-squares fitting on multiple points on the decay curve, the following can be calculated: .

[0104] S800: Decision-making and channel switching based on prediction-actual comparison.

[0105] This step is the decision-making stage for verifying the prediction guidance.

[0106] S810: Navigation and Comparison. The MCU converts the actual physical parameters calculated by the S700 ( , In the 3D physical property navigation map generated by S550, the predicted physical parameters corresponding to the current duct position are ( , ) for comparison.

[0107] For example, the navigation map displays the predicted parameters for the current target location as follows: , Calculate the difference: , Assume the preset matching thresholds are as follows: and .because and The system determined that the prediction matched the actual situation very well, confirming that the catheter had accurately reached the target point.

[0108] S820: Safety logic judgment. Simultaneously, the system will also perform safety judgments based on actual measured values ​​to identify dangerous states. For example: if... extremely high and Extremely high pressure (without pressure decay) indicates catheter tip adhesion to the wall or blockage; if Extremely low, close to the baseline value, indicates catheter slippage into a major artery. In these dangerous situations, the system will lock in channel A, prohibit drug administration, and issue an audible and visual alarm.

[0109] S830: Channel Switching. After the matching verification of S810 and the safety judgment of S820 are both passed, the MCU automatically controls the three-way solenoid valve to switch the flow path from channel A (physiological saline) to the second channel (channel B, loaded with chemotherapy drugs) in preparation for drug perfusion.

[0110] S900: Millisecond-level PID closed-loop pressure-compensated injection and precise sealing.

[0111] This step performs intelligent drug infusion and ensures accurate dosage delivery.

[0112] S910: Target parameter setting and real-time monitoring. The MCU uses data measured by S720. Calculate a safe target pressure limit ,For example Throughout the perfusion process, the pressure sensor continuously monitors the actual pressure within the blood vessel. .

[0113] S920: PID algorithm adjustment. The system does not inject at a constant rate, but performs dynamic pressure following. The PID (proportional-integral-derivative) control algorithm within the MCU adjusts based on the error. Adjust the pump drive signal in real time This changes the infusion flow rate.

[0114] For example, when a patient's heart enters systole, the pressure in the blood vessels themselves increases. It then rises and approaches At this point, the error When the pressure decreases, the PID controller reduces the pump speed to give way to blood flow; when the heart enters diastole, vascular pressure decreases. Decrease, error If the pressure is increased, the PID controller will increase the pump speed, utilizing this low-pressure window to quickly and efficiently infuse drugs into the tumor.

[0115] S930: Dual-channel cleaning and precise tubing sealing (Bolus Chasing). Once the pump's motor encoder displays the preset chemotherapy dose (e.g., 50ml) has been infused, the system pauses pumping. The MCU instructs the three-way valve to immediately switch back to channel A (normal saline), and then drives the pump to continue infusing a volume equal to... (For example, 1.2 ml) of normal saline. This process completely pushes the high concentration of chemotherapy drug remaining in the microcatheter lumen (approximately 1-2 ml) into the tumor, ensuring that the prescribed dose reaches the lesion 100%, while simultaneously filling the catheter with non-toxic normal saline to ensure the safety of medical staff during catheter removal. This completes one precise adaptive perfusion procedure.

[0116] Example 2 This embodiment describes a specific hardware architecture for a dual-channel adaptive perfusion system. For example... Figure 5 As shown, the system consists of the following four core functional modules.

[0117] The tissue modeling module is deployed on a GPU-accelerated workstation or embedded computing platform. It comprises three functional sub-modules: an image preprocessing sub-module, responsible for color normalization and cell nucleus instance segmentation of digitized pathological images; an atlas construction sub-module, responsible for constructing a cell spatial adjacency map based on the segmentation results; and a graph neural network inference sub-module, responsible for performing structural perturbation titration, identifying critical response singularities, and generating a 3D physical property navigation map containing predicted vascular physical parameters. The output of this module, the 3D physical property navigation map, is transmitted to the adaptive control unit via a wired or wireless interface.

[0118] The dual-channel perfusion unit's core hardware includes a precision peristaltic pump and a three-way solenoid valve. The peristaltic pump is driven by a stepper motor, its speed precisely controlled by a PWM signal from an adaptive control unit, enabling millisecond-level adjustment of the output flow rate. The three-way solenoid valve connects to the pump's output, the first channel (channel A, connected to the saline reservoir bag), and the second channel (channel B, connected to the medication reservoir bag). Valve switching is driven by digital I / O signals from the adaptive control unit, with a switching time of less than 50 milliseconds. The tubing from the two channels merges through the valve and connects to a microcatheter.

[0119] The sensing unit's core component is a high-sensitivity piezoresistive pressure sensor, installed in the proximal end of the microcatheter (near the valve outlet). This sensor has a measurement range of 0 to 300 mmHg, an accuracy of ±1 mmHg, and a sampling rate of at least 1000 Hz. The sensor's analog output signal, after passing through a signal conditioning circuit (including amplification and low-pass filtering), is input to the analog-to-digital converter (ADC) of the adaptive control unit.

[0120] The adaptive control unit is centered around a microcontroller unit (MCU), such as a processor based on the ARM Cortex-M4 architecture. The MCU's firmware implements the following control logic: receiving real-time pressure data from the sensing unit; calculating actual vascular physical parameters (dynamic fluid impedance and compliance time constant) based on the pressure response curve; comparing the actual parameters with predicted parameters from the navigation map received from the tissue modeling module; controlling the on / off state of the three-way solenoid valve based on the comparison result to achieve channel switching; during the drug delivery phase, its built-in proportional-integral-derivative (PID) control algorithm module operates at millisecond intervals, dynamically calculating and outputting the pump's drive PWM signal based on the error between the real-time pressure and the preset target pressure upper limit, achieving closed-loop pressure compensation control; after drug delivery is completed, it automatically performs channel reversal and tubing sealing operations for quantitative flushing fluid delivery.

[0121] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing a three-dimensional physical property navigation map of an organization, characterized in that, include: Obtain a cell spatial adjacency map that characterizes cell nuclei and their spatial adjacency relationships in a digital pathological image. The cell spatial adjacency map includes multiple nodes and edges connecting the nodes. Each node corresponds to a cell nucleus and is associated with an initial node feature vector that characterizes the quantitative features of the cell nucleus. Define a global structural metric to characterize the structural stability of the cell spatial adjacency map; For at least one seed node in the cell space adjacency map, a conformational stress sequence is iteratively applied, wherein the conformational stress sequence is a perturbation vector that is progressively increased. During the iterative application process, a pre-trained graph neural network model is used to calculate the response value of the global structure metric determined by the node embedding output by the graph neural network model after each application of the perturbation vector, so as to generate a global structure metric response curve. A critical response singularity is identified in the global structural metric response curve, and a three-dimensional physical property navigation map characterizing the tissue region is constructed based on the critical response singularity. The three-dimensional physical property navigation map contains a predicted vascular physical parameter at one or more locations within the region.

2. The method according to claim 1, characterized in that, The process of obtaining the cell spatial adjacency map includes: The digital pathological image is preprocessed, and all cell nucleus instances in the image are segmented. Each of the cell nuclei instances is defined as a node in the cell space adjacency graph; For each node, at least one quantitative feature of the corresponding cell nucleus is extracted to generate the initial node feature vector; Based on the spatial distance relationship between the cell nucleus instances, the connection relationship between the nodes is determined to generate the edges.

3. The method according to claim 1, characterized in that, The global structure metric is the global structure entropy, which is calculated based on the Shannon entropy of the cosine similarity distribution between the embedding vectors of all nodes output by the graph neural network model and their mean embedding vector.

4. The method according to claim 1, characterized in that, The construction of the three-dimensional physical property navigation map based on the critical response singularity includes: Calculate the deviation of the embedding vector of each node from the system mean embedding vector at the number of iteration steps corresponding to the critical response singularity; Based on the mapping relationship between the deviation and vascular physical properties, the predicted vascular physical parameters of the region where each node is located are generated.

5. A control method for a dual-channel infusion system, applied to an infusion system including an infusion pump, a channel switching valve, and a pressure sensor, characterized in that, include: The injection pump is controlled to output a tentative fluid pulse signal through the first channel, and a pressure response curve triggered by the tentative fluid pulse signal is acquired in real time by the pressure sensor. Based on the pressure response curve, an actual blood vessel physical parameter is calculated. The actual vascular physical parameters are compared with the predicted vascular physical parameters at the corresponding positions in a three-dimensional physical property navigation map constructed by structural perturbation analysis of a cell spatial adjacency map based on digital pathological images. When the comparison result meets the preset matching conditions, the channel switching valve is controlled to switch from the first channel to the second channel, and the infusion pump is controlled to deliver the medicine through the second channel; During the drug delivery process, based on the real-time feedback from the pressure sensor, closed-loop pressure compensation control is performed to dynamically adjust the output flow rate of the infusion pump.

6. The control method according to claim 5, characterized in that, The actual vascular physical parameters and the predicted vascular physical parameters include parameters selected from at least one of the following groups: dynamic fluid impedance, vascular wall compliance time constant; The dynamic fluid impedance is calculated based on a peak pressure, a reference pressure, and a preset flow rate of the trial fluid pulse signal from the pressure response curve. The vessel wall compliance time constant is determined based on a function fit of the pressure response curve to a pressure decay segment after the peak pressure.

7. The control method according to claim 6, characterized in that, The preset matching condition is: the difference between the actual vascular physical parameters and the predicted vascular physical parameters is less than a preset matching threshold.

8. The control method according to claim 7, characterized in that, After the drug delivery is completed, the control method further includes: Control the channel switching valve to switch from the second channel back to the first channel; The infusion pump is controlled to deliver a preset volume of flushing fluid through the first channel, the preset volume being matched with the dead cavity volume of the infusion system's pipeline.

9. A dual-channel adaptive infusion system, characterized in that, include: A tissue modeling module is used to acquire a cell spatial adjacency map characterizing a digital pathological image, and to construct a three-dimensional physical property navigation map containing predicted vascular physical parameters by applying virtual perturbation to the map and analyzing its structural response. A dual-channel infusion unit includes an infusion pump and a channel switching valve connected to the first and second channel flow paths; A sensing unit is used to acquire a pressure response curve in real time when the injection pump outputs an exploratory fluid pulse signal through the first channel; An adaptive control unit is configured as follows: Calculate actual blood vessel physical parameters based on the pressure response curve; The actual vascular physical parameters are compared with the predicted vascular physical parameters at the corresponding positions in the three-dimensional physical property navigation map; Based on the comparison results, the channel switching valve is controlled to switch from the first channel to the second channel to perform drug delivery; and During the drug delivery process, a closed-loop pressure compensation control is executed based on the real-time feedback of the sensing unit to adaptively adjust the output flow rate of the infusion pump.

10. The system according to claim 9, characterized in that, The tissue modeling module includes: An image preprocessing submodule is used to perform color normalization and cell nucleus instance segmentation on the digitized pathological image; A graph construction submodule is used to construct the cell spatial adjacency graph based on the spatial distance relationships of the segmented cell nucleus instances; A neural network inference submodule is used to identify critical response singularities through a structural perturbation titration process and generate the three-dimensional physical property navigation map based on the critical response singularities.

Images