An artificial intelligence-fused individualized treatment method for tumor patients

Abstract

This invention relates to the field of artificial intelligence technology and discloses a personalized treatment method for cancer patients that integrates artificial intelligence. The method includes: acquiring two adjacent phases of contrast-enhanced tumor images; generating a spatially weighted oligoprogression margin after registration and lesion segmentation; constructing a graph structure containing one image node and three treatment text nodes corresponding to continuing the original systemic therapy, adding local therapy, and changing systemic therapy, respectively; and writing the margin into directed edges from the image node to each treatment node. Based on the margin, a gating mapping is performed to generate directional weights, which are combined with the initial representation of the image node to form edge messages; attention coefficients of each directed edge are calculated and normalized using attention-temperature mapping. The representations of the three treatment text nodes are updated by weighting the corresponding edge messages with the attention coefficients, outputting the probability distribution of each node. Supervised training is performed using historical treatment text labels, and shape alignment constraints are used to ensure that the gating mapping and the margin change in the same direction, resulting in a target graph neural network.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for individualized treatment of cancer patients that integrates artificial intelligence. Background Technology

[0002] In the field of oncology treatment, medical imaging has become the basis for assessing disease progression and treatment efficacy. Doctors typically decide whether to continue the current treatment plan, combine it with local therapy, or switch to systemic therapy based on continuous imaging changes and the patient's past treatment records. With the development of artificial intelligence in medical data analysis, how to combine imaging information with real treatment records to build intelligent models capable of predicting or generating treatment recommendations has become an important research direction in medical AI. In recent years, graph neural networks, due to their ability to represent multi-entity, multi-relationship medical knowledge structures, have been explored for joint modeling of multi-source data such as imaging, pathology, and drugs, providing a theoretical foundation for automated assisted medical recommendations.

[0003] Existing image- and text-based intelligent recommendation systems are mostly trained based on a consistent positive correlation between image changes and treatment adjustments. That is, when an image shows an increase in lesion size, the treatment plan should be changed. However, in clinical practice, this training basis does not always hold true. Some patients experience localized progression of only a few lesions during systemic treatment, while the majority of lesions remain under control. In such cases, doctors often use local radiotherapy or ablation to treat the progressive lesions while continuing the original systemic treatment. This means that the same image change may correspond to completely different treatment decisions in different contexts: for some patients, it means continuing the current plan, while for others it means changing the plan. Traditional learning models struggle to accurately represent the complex nonlinear relationship between image changes and treatment choices in such situations, easily forming misleading associations during the training phase and thus outputting results that deviate from the actual decision-making logic during the inference phase. Summary of the Invention

[0004] This invention provides a personalized treatment method for cancer patients that integrates artificial intelligence, solving the technical problem of how to establish a mechanism that can adaptively describe the direction and intensity of the influence of image changes on treatment behavior during the construction of graph neural networks, so that the model can still maintain the expressive ability and learning stability consistent with actual clinical decisions when dealing with cases of local progression or partial control.

[0005] This invention provides a method for personalized treatment of cancer patients that integrates artificial intelligence, comprising:

[0006] After acquiring two adjacent tumor imaging images, registering and segmenting the lesions, a spatially weighted oligoprogression mixture margin is generated based on preset parameters.

[0007] Construct a graph containing one image node and three treatment text nodes. The three treatment text nodes correspond to continuing the original system treatment, adding local treatment, and changing the system treatment, respectively. Establish directed edges from the image node to each treatment text node, and write the spatial weighted oligoprogression mixing margin as an edge attribute into each directed edge.

[0008] Gating mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain orientation weights, and edge messages of each directed edge are generated by combining the initial representation of the image node; attention temperature mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain attention coefficients of each directed edge and normalize them;

[0009] The representations of the three treatment text nodes are updated by weighting the edge messages with attention coefficients, and the probabilities corresponding to the three treatment text nodes are output respectively.

[0010] Supervised training is performed using labels extracted from historical treatment texts corresponding to three treatment text nodes. Meanwhile, shape alignment constraints are used to ensure that the gating mapping and spatially weighted oligoprogression mixing margin change in the same direction, resulting in a target graph neural network.

[0011] The beneficial effects of this invention include: by introducing a spatially weighted oligoprogression mixing margin into the graph neural network, the relationship between image changes and treatment text is transformed from a fixed positive correlation propagation into a directional mapping that can be dynamically adjusted according to the degree of local progression. This enables the graph neural network to automatically balance three semantic paths—continuing the original systemic treatment, adding local treatment, and changing systemic treatment—when the patient's systemic treatment is still effective but local lesions have progressed. This invention achieves dynamic alignment between image spatial information and treatment semantic information, significantly improving the model's ability to express nonlinear clinical auxiliary decision-making patterns. Attached Figure Description

[0012] Figure 1 This is a flowchart of a personalized treatment method for cancer patients that integrates artificial intelligence, according to the present invention. Detailed Implementation

[0013] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0014] like Figure 1 As shown, a personalized treatment method for cancer patients that integrates artificial intelligence includes:

[0015] After acquiring two adjacent tumor imaging images, registering and segmenting the lesions, a spatially weighted oligoprogression mixture margin is generated based on preset parameters.

[0016] Construct a graph containing one image node and three treatment text nodes. The three treatment text nodes correspond to continuing the original system treatment, adding local treatment, and changing the system treatment, respectively. Establish directed edges from the image node to each treatment text node, and write the spatial weighted oligoprogression mixing margin as an edge attribute into each directed edge.

[0017] Gating mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain orientation weights, and edge messages of each directed edge are generated by combining the initial representation of the image node; attention temperature mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain attention coefficients of each directed edge and normalize them;

[0018] The representations of the three treatment text nodes are updated by weighting the edge messages with attention coefficients, and the probabilities corresponding to the three treatment text nodes are output respectively.

[0019] Supervised training is performed using labels extracted from historical treatment texts corresponding to three treatment text nodes. Meanwhile, shape alignment constraints are used to ensure that the gating mapping and spatially weighted oligoprogression mixing margin change in the same direction, resulting in a target graph neural network.

[0020] In one embodiment of the present invention, two adjacent tumor imaging images are acquired, registered, and segmented. A spatially weighted oligoprogression mixing margin is then generated based on preset parameters, including:

[0021] The relative volume change ratio is obtained by dividing the difference between the current volume and the previous volume of each lesion by the previous volume.

[0022] When the relative volume change ratio is greater than the preset significant growth threshold, the corresponding lesion is marked as a candidate lesion for growth; otherwise, it is not marked as a candidate lesion for growth.

[0023] Obtain the spatial location of each lesion, calculate the sum of the spatial locations of all candidate lesions for growth, and then divide the sum of spatial locations by the number of candidate lesions for growth to obtain the spatial aggregation center of the set of candidate lesions for growth.

[0024] Set the spatial kernel scale parameter, calculate the squared value of the distance between the spatial location of each lesion and the spatial aggregation center, divide the squared value by twice the square of the spatial kernel scale parameter, take the negative value, and then calculate the exponential function value of the negative value to obtain the spatial distance weight of each lesion.

[0025] Calculate the difference between the relative volume change ratio of each lesion and the preset significant growth threshold. When the difference is positive, the difference is used as the growth amount exceeding the threshold; otherwise, the growth amount exceeding the threshold is zero. Calculate the sum of the growth amounts exceeding the threshold for all lesions. Divide the growth amount exceeding the threshold for each lesion by the sum of the growth amounts exceeding the threshold to obtain the amplitude normalization weight of each lesion.

[0026] Multiply the spatial distance weight of each lesion by the amplitude normalization weight to obtain the joint weight of each lesion; calculate the sum of the products of the joint weights of all lesions and the corresponding growth candidate markers to obtain the effective progression count; calculate the sum of the joint weights of all lesions to obtain the total effective count.

[0027] Set a small-scale boundary parameter, a priori shape parameter one, and a priori shape parameter two. Calculate the first sum of the priori shape parameter one and the effective progress count, and the second sum of the priori shape parameter two and the difference between the effective total count and the effective progress count.

[0028] Using a small-scale boundary parameter as the cutoff value, calculate the regularized incomplete beta function values ​​corresponding to the first and second sums; calculate the logarithm of the ratio of the regularized incomplete beta function value to a value minus the regularized incomplete beta function value to obtain the spatially weighted oligoprogression mixing margin.

[0029] The relative volume change ratio is an indicator that measures the degree of change in the volume of a single lesion between two adjacent imaging phases. A positive number indicates that the lesion volume has increased, a negative number indicates that it has decreased, and zero indicates that there has been no change. It is used to preliminarily determine whether there is a growth trend of the lesion. Specifically, the relative volume change ratio is equal to (the current lesion volume minus the previous lesion volume) divided by the previous lesion volume corresponding to the current lesion volume.

[0030] The preset significant growth threshold is a critical value that distinguishes between normal measurement fluctuations and actual significant growth of lesions. It needs to be set in conjunction with the accuracy of the imaging equipment (such as 1 mm for CT and 0.5 mm for MRI) and the RECIST 1.1 standard to avoid misjudging measurement errors as lesion growth. Specifically, if the imaging equipment is CT (measurement accuracy 1 mm), the preset significant growth threshold is set to 20% (corresponding to the volume change threshold, since a 20% change in diameter is defined as progression in RECIST 1.1); if it is MRI (measurement accuracy 0.5 mm), the threshold is set to 15%.

[0031] The candidate lesion marker for growth is a binary identifier used to distinguish between lesions that may progress and those that are under control. Marking it as "yes" indicates that the lesion has a significant growth trend, while marking it as "no" indicates that the lesion volume change is within the normal range of fluctuation.

[0032] The spatial location of a lesion is its three-dimensional coordinates in a medical imaging coordinate system (such as the Cartesian coordinate system of CT, where the X / Y / Z axes are in millimeters), used to quantify the spatial distribution of the lesion.

[0033] The spatial clustering center is the average coordinate of the spatial location of all candidate lesions, reflecting the degree of spatial concentration of the lesions. The more concentrated the clustering center, the stronger the spatial clustering of the lesions. Specifically, the X-axis coordinate of the spatial clustering center is equal to the sum of the X-axis coordinates of all candidate lesions divided by the number of candidate lesions. The Y-axis and Z-axis coordinates are calculated in the same way as the X-axis, and finally the three-dimensional spatial clustering center coordinates are obtained.

[0034] The spatial kernel scale parameter is a parameter that controls the range of influence of spatial distance on weight. The smaller the value, the greater the weight of lesions that are close to the aggregation center (sensitive to clustering). The larger the value, the more dispersed the weight distribution (insensitive to clustering). Specifically, the spatial kernel scale parameter is equal to one-third of the maximum distance from all lesions (including non-growing candidate lesions) to the spatial aggregation center. For example, if the maximum distance is 30 mm, the spatial kernel scale parameter is set to 10 mm.

[0035] Spatial distance weight is an indicator that quantifies the spatial correlation between lesions and the clustering center of growing candidate lesions. The larger the value, the closer the lesion is to the clustering center, and the stronger the spatial clustering. Specifically, the steps are as follows: First, calculate the three-dimensional Euclidean distance between the spatial location of a single lesion and the spatial clustering center (i.e., the square root of the sum of the squares of the differences in the X-axis, Y-axis, and Z-axis coordinates); second, calculate the square of this distance; third, calculate the square of twice the spatial kernel scale parameter; fourth, divide the square of the distance by the result of the second step, and take the negative value of the quotient; fifth, calculate the natural exponential function value of this negative value (i.e., e raised to the power of the negative value), and the result is the spatial distance weight.

[0036] The over-threshold growth is an indicator that measures the portion of lesion growth that exceeds a significant threshold. It is only non-zero when there is significant growth in the lesion and is used to distinguish the magnitude of the growth. Specifically, the over-threshold growth is equal to the relative volume change ratio of a single lesion minus a preset significant growth threshold. If the difference is greater than zero, the over-threshold growth is taken as the difference; if it is less than or equal to zero, the over-threshold growth is taken as zero.

[0037] The amplitude normalization weight is a weight that normalizes the above-threshold growth of all lesions to the range of 0-1. The larger the value, the higher the proportion of the growth of that lesion among all growing lesions. Specifically, the first step is to calculate the sum of the above-threshold growth of all lesions (including non-growing candidate lesions) (the above-threshold growth of non-growing candidate lesions is zero and does not affect the sum); the second step is that if the sum is zero (no growing candidate lesions), the amplitude normalization weight of all lesions is set to zero; if the sum is not zero, the amplitude normalization weight of a single lesion is equal to the above-threshold growth of that lesion divided by the sum.

[0038] The joint weight is a comprehensive indicator that combines spatial clustering (spatial distance weight) and growth rate (amplitude normalization weight). It is used to simultaneously consider the impact of the spatial distribution and growth intensity of lesions on the judgment of oligoprogression. Specifically, the joint weight of a single lesion is equal to the spatial distance weight of that lesion multiplied by its amplitude normalization weight.

[0039] Effective progression count is an indicator that quantifies the total contribution of lesions with spatial clustering and significant growth. Only growing candidate lesions will have a positive contribution to this count, while non-growing candidate lesions will have a zero contribution. Specifically, the first step is to multiply the joint weight of each lesion by its growth candidate label (1 for yes, 0 for no) to obtain the progression contribution value of a single lesion. The second step is to calculate the sum of the progression contribution values ​​of all lesions, and the result is the effective progression count.

[0040] The effective total count is an indicator that quantifies the overall contribution of all lesions (including controlled and progressive lesions) in terms of space and magnitude. It is used to compare with the effective progression count to determine the relative proportion of progressive lesions. Specifically, the effective total count is equal to the sum of the combined weights of all lesions (including growth candidate and non-growth candidate lesions).

[0041] The small proportion boundary parameter is the critical ratio value that defines oligoprogression (small proportion of progression) and diffuse progression (large proportion of progression). It is used to determine whether the proportion of effective progression count to the total effective count falls into the oligoprogression category. Specifically, referring to the clinical definition of oligoprogression (≤3-5 lesion progression), the small proportion boundary parameter is set to 0.2 (that is, the proportion of effective progression count to the total effective count is ≤20% and is judged as small proportion, corresponding to clinical few lesion progression).

[0042] The first prior shape parameter is the first shape parameter of the Beta probability distribution, which reflects the prior belief about the proportion of progressive lesions (such as a bias towards a small proportion of progression) and is used to quantify prior information in Bayesian statistical inference. Specifically, based on the proportion of oligoprogressive cases in clinical practice (about 30%), the first prior shape parameter is set to 3 (corresponding to the prior tendency of a small proportion of progression, the mean of the Beta distribution = the first prior shape parameter / (the first prior shape parameter + the second prior shape parameter)).

[0043] The second prior shape parameter is the second shape parameter of the Beta probability distribution. Together with the first prior shape parameter, it determines the shape of the prior distribution and is used to balance the strength of the prior belief about the proportion of progression. Specifically, combining the first prior shape parameter (3) with the proportion of clinical oligoprogression (30%), the second prior shape parameter is set to 7 (at this time, the mean of the Beta distribution = 3 / (3+7) = 0.3, matching the clinical prior).

[0044] The first sum is the first shape parameter of the Beta posterior distribution, which is obtained by fusing the first prior shape parameter with the effective progress count, and is used for subsequent calculation of the regularized incomplete beta function; specifically, the first sum is equal to the first prior shape parameter plus the effective progress count.

[0045] The second sum is the second shape parameter of the Beta posterior distribution, which is obtained by fusing the second prior shape parameter with the effective total count minus the effective progress count. Together with the first sum, it constitutes the posterior distribution. Specifically, the second sum is equal to the second prior shape parameter plus (the effective total count minus the effective progress count).

[0046] The regularized incomplete beta function value quantifies the probability that the proportion of progress is less than or equal to the small-proportion boundary parameter, reflecting the strength of evidence for oligoprogression. A larger value indicates a higher probability of oligoprogression. Specifically, the regularized incomplete beta function value is calculated using numerical methods (such as the Newton-Coates method). The steps are as follows: First, determine the integration interval from 0 to the small-proportion boundary parameter. Second, calculate the integrand (i.e., the probability density function of the Beta function, which is equal to x raised to the power of (first sum - 1) multiplied by (1 - x) raised to the power of (second sum - 1), and then divided by the Beta function value (the Beta function value is calculated using the gamma function, i.e., Gamma(first sum) × Gamma(second sum) / Gamma(first sum + second sum)). Third, integrate the integrand over the integration interval; the result is the regularized incomplete beta function value.

[0047] Spatial-weighted oligoprogression mixture margin is a continuous indicator designed to quantify the strength of evidence for oligoprogressive morphology in tumors. It transforms three key clinical characteristics of oligoprogression—spatial clustering of lesions, growth rate, and low-proportion progression—into quantitative signals that can be directly used as edge attributes in graph neural networks through statistical modeling. Essentially, spatial-weighted oligoprogression mixture margin is a Bayesian transformation of the strength of evidence for the clinical judgment of whether progressive lesions constitute a low proportion. Specifically, the regularized incomplete beta function value in the formula represents the probability, under a Beta posterior distribution (integrating prior clinical experience and current lesion data), that the proportion of progressive lesions ≤ the low-proportion boundary parameter (e.g., 20%). A larger value indicates stronger statistical evidence for low-proportion progression. Dividing the regularized incomplete beta function value by (1 minus this value) transforms this probability into the ratio of the probability of low-proportion progression to the probability of high-proportion / diffuse progression (i.e., the dominance ratio), which better highlights the two progression morphologies. The difference is that the natural logarithm of the ratio is taken to transform the non-negative odds ratio into a continuous value that can be positive or negative (when the margin is positive, the larger the value, the stronger the evidence for oligoprogression, corresponding to the typical oligoprogression scenario in clinical practice where a few lesions are spatially clustered and have a high growth rate, supporting the semantic coupling of continuing the original systemic therapy + local therapy; when the margin is negative or close to zero, it represents weak evidence for oligoprogression, which is more inclined to diffuse progression, corresponding to the potential need to change systemic therapy). At the same time, the logarithmic transformation makes the margin more sensitive to changes in the probability ratio, ensuring that the strength of evidence can still be effectively distinguished in the boundary scenario between oligoprogression and non-oligoprogression.

[0048] It should be noted that oligoprogression is a specific pattern of lesion progression that occurs in cancer patients during systemic treatment (such as chemotherapy, targeted therapy, immunotherapy, etc.), including:

[0049] 1. The number of lesions that have progressed is small. Clinically, this is usually defined as only ≤3-5 lesions showing clear progression (exceeding the increase in measurement noise), while the vast majority of the remaining lesions remain under control (stable volume or shrinkage).

[0050] 2. Spatial clustering of progressive lesions, meaning that a small number of progressive lesions are concentrated in the same or adjacent anatomical areas, rather than being diffusely distributed throughout the body.

[0051] From a clinical perspective, oligoprogression is a spatial focal resistance caused by tumor clonal heterogeneity. Only a portion of tumor cell clones develop resistance to the current systemic therapy and cause local lesion progression, while the remaining sensitive clones are still controlled by treatment. Therefore, the clinical management strategy is not to immediately switch systemic therapies, but to eliminate drug-resistant progressive lesions through local treatments (such as radiofrequency ablation and stereotactic radiotherapy) while continuing the original systemic therapy, in order to delay the change of treatment regimen and prolong the benefit time of the original regimen.

[0052] It should be noted that non-oligoprogression is a lesion progression pattern relative to oligoprogression. It refers to situations where, during systemic therapy, lesion progression in cancer patients does not conform to the characteristics of minority and spatial clustering, and is divided into two categories:

[0053] 1. Diffuse progression, meaning that most or even all measurable lesions in the patient's body have progressed, indicating that the tumor as a whole is resistant to the current systemic therapy (rather than local clonal resistance).

[0054] 2. Multifocal progression, which means that the number of progressive lesions exceeds 5 and is spatially dispersed (such as simultaneous progression in multiple non-adjacent sites such as the lungs, liver, and bones).

[0055] In clinical management, non-oligoprogression usually means that the current systemic therapy can no longer effectively control the tumor. Therefore, doctors often choose to change the systemic therapy regimen (such as using targeted drugs with different mechanisms of action, adjusting the chemotherapy regimen, etc.) rather than treating the progressive lesions with local treatment alone.

[0056] It should be noted that lesion segmentation is a preprocessing step in tumor medical imaging that separates the region representing the tumor lesion from normal tissue (such as muscle, blood vessels, fat, normal organs, etc.) in the image through algorithms or manual assistance. Commonly used segmentation methods in clinical practice include traditional image segmentation algorithms (such as thresholding based on the difference in grayscale / signal between the lesion and normal tissue, and edge detection based on the edge features of the lesion, which are suitable for images with clear lesion boundaries and high tissue contrast, such as high-density metastases in CT) and deep learning segmentation algorithms (such as U-Net and its improved models, which learn lesion features by training a large number of labeled images, and are suitable for complex lesions with blurred boundaries and high overlap with normal tissue signals, such as brain metastases in MRI).

[0057] In one embodiment of the present invention, a graph is constructed comprising one image node and three treatment text nodes. The three treatment text nodes correspond to continuing the original system treatment, adding local treatment, and changing the system treatment, respectively. Directed edges are established from the image node to each treatment text node, and the spatially weighted oligoprogression mixing margin is written as an edge attribute to each directed edge, including:

[0058] Define a single-stage graph, which includes a set of nodes, a set of directed edges, and edge attribute functions;

[0059] The node set contains an image node, a treatment text node corresponding to continuing the original system treatment, a treatment text node corresponding to adding local treatment, and a treatment text node corresponding to changing the system treatment;

[0060] The set of directed edges contains three directed edges: the first directed edge is from the image node to the corresponding treatment text node that continues the original system treatment; the second directed edge is from the image node to the corresponding treatment text node that adds local treatment; and the third directed edge is from the image node to the corresponding treatment text node that changes the system treatment.

[0061] The edge attribute function is used to assign an attribute value to each directed edge, which is the spatially weighted oligo-progression mixture margin;

[0062] Define an edge attribute vector that contains the attribute values ​​of the first directed edge, the second directed edge, and the third directed edge in sequence.

[0063] A single-stage graph is a graph structure corresponding to one imaging assessment cycle for a cancer patient. A single stage specifically refers to the treatment decision cycle corresponding to two adjacent stages of cancer medical imaging (such as CT images 2 months and 4 months after treatment). This graph only focuses on the association between the current stage of imaging and treatment options and does not contain cross-stage temporal information. Its core function is to structure the relationship between imaging features and treatment semantics within a single cycle.

[0064] The node set is the core entity unit of the graph structure, retaining only nodes that are strongly related to treatment decisions in oligoprogression scenarios: the image node stores the fusion features of the two adjacent phases of the current stage (such as the encoded vector of lesion volume and spatial distribution), and the three treatment text nodes correspond to three types of treatment options in oligoprogression and non-oligoprogression scenarios in clinical practice (continue the original plan, add local treatment, change the plan).

[0065] Image nodes are the carriers of image information in the graph structure. Their initial representation needs to be generated by an image coding network (such as inputting the lesion segmentation results and relative volume changes of two images into a CNN or Transformer model, and outputting a 256-dimensional or 512-dimensional vector). This node does not directly output decisions, but only transmits the message of image feature association to the treatment text node through directed edges.

[0066] The treatment text node corresponding to continuing the original system treatment indicates that the original system treatment should continue. The initial representation is a learnable vector (such as a 256-dimensional randomly initialized vector). The representation is subsequently updated through weighted messages passed by the image node. The final output is the probability of choosing to continue the original system treatment. This corresponds to the original solution still being effective in oligoprogression scenarios, and the treatment logic does not require switching lines.

[0067] The treatment text node corresponding to adding local treatment represents adding local treatment (such as radiofrequency ablation or stereotactic radiotherapy) on the basis of continuing the original systemic treatment. It is initially represented as a learnable vector, and after updating, it outputs the probability of choosing to add local treatment, which corresponds to the clinical strategy of clearing local drug-resistant lesions and maintaining the benefit of the original treatment in oligoprogression scenarios.

[0068] The treatment text node corresponding to the change of systemic treatment indicates that the current systemic treatment is stopped and replaced with a new systemic treatment plan (such as changing to a different targeted drug or adjusting the chemotherapy regimen). The initial representation is a learnable vector. After the update, the output is the probability of choosing to change the systemic treatment. This corresponds to the handling logic of the original plan becoming resistant to drugs in non-oligoprogression scenarios, which requires changing the treatment path.

[0069] The directed edge set is a directional channel connecting image features and treatment semantics. The design of the direction from image nodes to treatment text nodes clarifies the unidirectionality of information transmission (only image features transmit related information to treatment decisions, and treatment semantics do not affect image features in the reverse). The three edges correspond to three potential associations: image → continue treatment, image → add local treatment, and image → change treatment, ensuring that each treatment option can independently receive image information.

[0070] The first directed edge is the association channel between image features and the semantics of continuing the original systemic treatment, used to convey information on the matching degree between the strength of oligoprogression evidence in the image and continued treatment.

[0071] The second directed edge is the association channel between image features and the semantics of local treatment, used to convey information on the matching degree between the strength of oligoprogression evidence in the image and local treatment.

[0072] The third directed edge is the association channel between image features and the semantics of systemic replacement therapy, used to convey information on the matching degree between the strength of oligoprogression evidence in the image and systemic replacement therapy.

[0073] The edge attribute function is the specific execution rule for assigning the spatially weighted oligoprogression mixture margin to three directed edges. It is necessary to ensure that the attribute values ​​of the three edges are consistent with the spatially weighted oligoprogression mixture margin of the current stage (because the oligoprogression evidence strength corresponding to the image of the same stage is unique, the three edges share the same spatially weighted oligoprogression mixture margin value). Specifically, the input of the edge attribute function is the directed edge identifier (such as edge 1, edge 2, edge 3) and the spatially weighted oligoprogression mixture margin of the current stage, and the output is the attribute value of the corresponding directed edge. Regardless of which directed edge identifier is input, the spatially weighted oligoprogression mixture margin of the current stage is directly used as the attribute value of that edge.

[0074] The edge attribute vector is a vector formed by integrating the attribute values ​​of three directed edges (all of which are spatially weighted oligoprogression mixture margins) in a fixed order (continue treatment edge → add local treatment edge → change treatment edge). Its core function is to facilitate the subsequent batch calling of the attribute values ​​of the three edges by the graph neural network (such as simultaneously inputting the gating mapping module to calculate the directional weights of the three edges), avoiding the efficiency loss of calling the attribute of a single edge. Specifically, the edge attribute vector is equal to a three-dimensional vector formed by arranging the attribute values ​​of the first directed edge, the second directed edge, and the third directed edge in the order of their attribute values. Since the attribute values ​​of the three edges are all spatially weighted oligoprogression mixture margins at the current stage, the vector can be represented as [spatially weighted oligoprogression mixture margin value, spatially weighted oligoprogression mixture margin value, spatially weighted oligoprogression mixture margin value].

[0075] In one embodiment of the present invention, gating mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain orientation weights, and edge messages for each directed edge are generated by combining the initial representation of the image nodes, including:

[0076] Set the gate mapping slope parameter and the gate mapping bias parameter, and use a logic function to perform the gate mapping; wherein, the logic function is calculated by dividing one by the sum of one and the natural exponential function value, where the exponent of the natural exponential function is a negative input value;

[0077] To calculate the continuation direction weight, first calculate the product of the spatially weighted oligoprogression mixing margin and the gated mapping slope parameter, then add the product to the gated mapping bias parameter to obtain the input value, and substitute the input value into the logic function to obtain the continuation direction weight.

[0078] Subtract the continuing direction weight from one to obtain the changing direction weight;

[0079] Configure a set of positive weight matrices and a set of negative weight matrices for the treatment text nodes that correspond to continuing the original system treatment, the treatment text nodes that correspond to adding local treatment, and the treatment text nodes that correspond to changing the system treatment.

[0080] For directed edges from image nodes to corresponding treatment text nodes that continue the original system treatment, directed edges from image nodes to corresponding treatment text nodes that add local treatment, and directed edges from image nodes to corresponding treatment text nodes that change the system treatment, the edge messages of the corresponding directed edges are calculated. First, the product of the positive weight matrix corresponding to the corresponding node and the initial representation of the image node is calculated and multiplied by the continuing direction weight to obtain the first result. Then, the product of the reverse weight matrix corresponding to the corresponding node and the initial representation of the image node is calculated and multiplied by the changing direction weight to obtain the second result. The first result and the second result are added together to obtain the edge message of the corresponding directed edge.

[0081] The gating slope parameter is a parameter that adjusts the sensitivity of the input value to the output of the logic function. The larger the value, the more significant the change in the weight of the continuation direction caused by the change in the input value (the more sensitive the gating response). The smaller the value, the smoother the gating response. Specifically, the initial value of the gating slope parameter is set in the range of 0.1-0.5, and the specific value is determined by cross-validation (take the value that minimizes the shape alignment loss in the training set, such as 0.3 when the loss is minimized in the validation set, then set it to 0.3).

[0082] The gating mapping bias parameter is a parameter that adjusts the baseline position of the logic function output. It is used to counteract the impact of the overall input value offset on the continuation direction weight, ensuring that the continuation direction weight is close to 0.5 (neutral state) when the input value is at a moderate level. Specifically, the initial value of the gating mapping bias parameter is set to the range of -0.2 to 0.2. It is specifically initialized through training (taking a value that makes the initial continuation direction weight close to 0.1 in the no-progress scenario and close to 0.8 in the minimal-progress scenario, such as initially setting it to 0.1).

[0083] A logic function is an activation function that compresses input values ​​of any range into the 0-1 interval. It is used to ensure that the weights (output) in the continuing direction conform to the weight value characteristics (0≤weight≤1). Its function is to transform the linear combination of inputs into nonlinear weight values.

[0084] The direction weight is a weight that quantifies the tendency of image information to be transmitted to the continuation / addition of local treatment semantics. It takes a value of 0-1. The larger the value, the more likely it is to transmit a stronger message to the continuation / addition of local treatment node.

[0085] The change direction weight is a weight that quantifies the tendency of image information to be transmitted to the change system treatment semantics. It is complementary to the continuation direction weight (the sum of the two is 1), and takes a value of 0-1. The larger the value, the more likely it is to transmit a stronger message to the change system treatment node.

[0086] The positive weight matrix corresponding to the treatment text node for continuing the original system treatment is a linear transformation matrix that transmits the message supporting continued treatment to the original system treatment node from the initial representation of the image node. Its core function is to map the image features to feature dimensions that match the semantics of continued treatment. Specifically, the dimension of the positive weight matrix is ​​set to the dimension of the initial representation of the image node × the dimension of the embedding of the treatment text node (e.g., if the initial representation of the image is a 256-dimensional vector and the embedding of the treatment node is 256-dimensional, then the matrix is ​​256 rows × 256 columns). The initial value is initialized using Xavier (randomly sampled from a normal distribution with a mean of 0 and a variance of 2 / (image dimension + treatment node dimension)).

[0087] The positive weight matrix corresponding to the treatment text node for adding local treatment and the treatment text node for changing system treatment is a linear transformation matrix that transmits the message supporting adding local treatment to the initial representation of the image node. Its function is the same as the positive matrix of the continuing treatment node, only the parameter values ​​are different (to adapt to the semantics of adding local treatment). Specifically, the dimension setting is the same as the positive matrix of the continuing treatment node (e.g., 256 rows × 256 columns), the initial value is also initialized using Xavier, and it is independent of other positive matrix parameters (does not share parameters).

[0088] The positive weight matrix corresponding to the treatment text node for adding local treatment and the treatment text node corresponding to changing system treatment is a linear transformation matrix that transmits the message supporting the change treatment to the image node from the initial representation of the image node. Although the change treatment depends more on the inverse weight matrix, this positive weight matrix is ​​used to cover boundary scenarios (such as messages that slightly support the change treatment under weak or minimal progression). Specifically, the dimension setting is the same as the first two types of positive matrices (such as 256 rows × 256 columns), the initial value is initialized using Xavier, and the parameters are independent of other positive matrices.

[0089] The inverse weight matrix corresponding to the treatment text node that continues the original system treatment is a linear transformation matrix that transmits the message against continuing treatment to the original system treatment node from the initial representation of the image node. It has the opposite function of the forward matrix and is used to weaken the message strength of the continuing treatment node in non-independent progression scenarios. Specifically, the dimension is the same as the forward matrix of the same node (e.g., 256 rows × 256 columns), the initial value is initialized using Xavier, and the parameters are independent of the forward matrix.

[0090] The inverse weight matrix corresponding to the treatment text node with local treatment is a linear transformation matrix that transmits the message against local treatment to the local treatment node from the initial representation of the image node. It is used to weaken the message strength of the local treatment node in non-sparse progression scenarios. Specifically, the dimension is the same as the forward matrix of the same node (e.g., 256 rows × 256 columns), the initial value is initialized using Xavier, and the parameters are independent of the forward matrix.

[0091] The inverse weight matrix corresponding to the treatment text node of the system change treatment is a linear transformation matrix that transmits the anti-treatment change message to the system change treatment node from the initial representation of the image node. It has the opposite function of the forward matrix and is used to weaken the message strength of the treatment change node in oligoprogression scenarios. Specifically, the dimension is the same as the forward matrix of the same node (e.g., 256 rows × 256 columns), the initial value is initialized using Xavier, and the parameters are independent of the forward matrix.

[0092] The initial representation of an image node is a high-dimensional vector (e.g., 256-dimensional or 512-dimensional) transformed from the features of tumor medical imaging (such as lesion segmentation results, volume changes, and spatial distribution) through an encoding network (such as CNN or Transformer). It is the original image feature carrier for generating side messages, and its dimension needs to match the dimension of the positive / negative weight matrix; there is no computational requirement (the encoding network parameters are optimized during training).

[0093] The edge message corresponding to the directed edge for continuing the original system treatment is a quantitative carrier of the semantic association between image features and continued treatment. It integrates information that supports continued treatment (first outcome) and opposes continued treatment (second outcome), and the final value reflects the combined influence of the two.

[0094] The edge message corresponding to the directed edge of the local treatment is a quantitative carrier of the semantic association between image features and the local treatment. The logic is consistent with the edge message of the continued treatment, and it only depends on the positive / negative matrix of the local treatment node.

[0095] The edge message corresponding to the directed edge of the system treatment replacement is a quantitative carrier of the semantic association between image features and treatment replacement. The logic is the same as the first two types of edge messages, and it only depends on the positive / negative matrix of the treatment replacement node.

[0096] In one embodiment of the present invention, attention temperature mapping is performed based on spatially weighted oligoprogression mixing margin to obtain and normalize the attention coefficients of each directed edge, including:

[0097] The attention temperature slope parameter is set, and the product of the spatially weighted oligoprogression mixing margin and the attention temperature slope parameter is calculated. This product is used as the exponent of the natural exponential function to calculate the attention temperature parameter value.

[0098] Configure a first attention projection vector for the treatment text node that continues the original system treatment, configure a second attention projection vector for the treatment text node that adds local treatment, and configure a third attention projection vector for the treatment text node that changes the system treatment.

[0099] The correlation scores are calculated for directed edges from image nodes to treatment text nodes that continue the original system treatment, directed edges from image nodes to treatment text nodes that add local treatment, and directed edges from image nodes to treatment text nodes that change the system treatment. The correlation scores are obtained by multiplying the corresponding attention projection vector with the initial representation of the image node.

[0100] The unnormalized attention coefficients are calculated for directed edges from image nodes to treatment text nodes that continue the original system treatment, directed edges from image nodes to treatment text nodes that add local treatment, and directed edges from image nodes to treatment text nodes that change the system treatment. The product of the attention temperature parameter value and the correlation score of the corresponding directed edge is used as the exponent of the natural exponential function to calculate the unnormalized attention coefficients.

[0101] Calculate the sum of the unnormalized attention coefficients of the directed edges from the image node to the corresponding treatment text node that continues the original system treatment, the directed edges from the image node to the corresponding treatment text node that adds local treatment, and the directed edges from the image node to the corresponding treatment text node that changes the system treatment.

[0102] Calculate the attention coefficients of directed edges from image nodes to the corresponding treatment text nodes that continue the original system treatment, directed edges from image nodes to the corresponding treatment text nodes that add local treatment, and directed edges from image nodes to the corresponding treatment text nodes that change the system treatment. The attention coefficient is the unnormalized attention coefficient of the corresponding directed edge divided by the total attention.

[0103] The attention temperature slope parameter is a parameter that adjusts the degree of influence of the input value on the attention temperature parameter value. The larger the value, the more significant the change in the temperature parameter value caused by the change in the input value; the smaller the value, the more gradual the change in the temperature parameter value. Specifically, the initial value of the attention temperature slope parameter is set in the range of 0.05-0.2, and the specific value is determined through training and validation (the value that best matches the attention allocation with the clinical decision in the training set is selected; for example, if the highest matching degree in the validation set corresponds to 0.1, then it is set to 0.1).

[0104] The attention temperature parameter is a parameter that controls the sharpness of the attention coefficient distribution. The larger the value, the more attention is concentrated on a few edges (the weight of highly relevant edges is amplified), and the smaller the value, the more balanced the attention distribution.

[0105] The first attention projection vector is a vector that maps the initial representation of the image node to the semantics of the continued original system treatment. Its function is to transform the image features into a dimension with computable relevance, ensuring that the relevance score can be calculated with the initial representation of the image. Specifically, the dimension of the first attention projection vector is consistent with the dimension of the initial representation of the image node (e.g., if the initial representation of the image is 256-dimensional, then the projection vector is 256-dimensional). The initial value is initialized using Xavier (randomly sampled from a normal distribution with a mean of 0 and a variance of 2 / (the dimension of the initial representation of the image)).

[0106] The second attention projection vector is a vector that maps the initial representation of the image node to the semantics of local treatment. Its function is the same as that of the first attention projection vector, only the parameter values ​​are different (to adapt to the semantics of local treatment). Specifically, its dimension is the same as that of the initial representation of the image node (e.g., 256 dimensions), the initial value is initialized using Xavier, and the parameters are independent of those of the first attention projection vector (they do not share parameters).

[0107] The third attention projection vector is a vector that maps the initial representation of the image node to the semantics of the treatment of the changed system. Its function is the same as the first two types of projection vectors, only the parameter values ​​are different (to adapt to the semantics of the treatment of the changed system). Specifically, its dimension is the same as the dimension of the initial representation of the image node (e.g., 256 dimensions), the initial value is initialized using Xavier, and it is independent of the parameters of the first and second attention projection vectors.

[0108] The correlation score of the directed edge corresponding to the continuation of the original system treatment is an indicator that quantifies the degree of semantic matching between the initial image representation and the continuation of the original system treatment. The larger the value, the higher the correlation between the image features and the semantics of the treatment. Specifically, the correlation score is equal to the vector inner product of the first attention projection vector and the initial image node representation (i.e., the sum of the corresponding dimension elements).

[0109] The correlation score of the directed edge corresponding to the local treatment is an indicator that quantifies the degree of semantic matching between the initial image representation and the local treatment. The larger the value, the higher the correlation between the image features and the semantics of the treatment. Specifically, the correlation score is equal to the dot product of the second attention projection vector and the initial image node representation.

[0110] The correlation score of the directed edge corresponding to the system treatment is an indicator that quantifies the degree of semantic matching between the initial image representation and the system treatment. The larger the value, the higher the correlation between the image features and the semantics of the treatment. Specifically, the correlation score is equal to the dot product of the third attention projection vector and the initial image node representation.

[0111] The unnormalized attention coefficients of the directed edges corresponding to the continuation of the original system treatment are attention weights that have not been calibrated by weight allocation, reflecting the original strength of the semantic correlation between the image and the continuation of treatment.

[0112] The unnormalized attention coefficients of the directed edges corresponding to local treatment are attention weights that have not been calibrated by weight allocation, reflecting the original strength of the semantic correlation between the image and the local treatment.

[0113] The unnormalized attention coefficients of the directed edges corresponding to the system replacement treatment are attention weights that have not been calibrated by weight allocation, reflecting the original strength of the semantic correlation between the image and the replacement treatment.

[0114] The attention sum is the cumulative value of the three unnormalized attention coefficients. Its purpose is to serve as the normalization denominator, ensuring that the sum of the normalized attention coefficients is 1.

[0115] The attention coefficient of the directed edge corresponding to the continuation of the original system treatment is the final attention weight after weight calibration, with a value of 0-1. The larger the value, the higher the weight of the message transmitted from the image to the continuation treatment node.

[0116] The attention coefficient of the directed edge corresponding to the local treatment is the final attention weight after weight calibration, with a value of 0-1. The larger the value, the higher the weight of the message passed from the image to the local treatment node.

[0117] The attention coefficient of the directed edge corresponding to the system replacement treatment is the final attention weight after weight calibration, with a value of 0-1. The larger the value, the higher the weight of the message passed from the image to the replacement treatment node.

[0118] In one embodiment of the present invention, the representations of the three treatment text nodes are updated by weighting the corresponding edge messages with attention coefficients, and the probabilities corresponding to the three treatment text nodes are output, including:

[0119] Calculate the weighted edge messages for directed edges from image nodes to the corresponding treatment text nodes that continue the original system treatment, directed edges from image nodes to the corresponding treatment text nodes that add local treatment, and directed edges from image nodes to the corresponding treatment text nodes that change the system treatment. The weighted edge message is the product of the attention coefficient of the corresponding directed edge and the corresponding edge message.

[0120] The weighted edge messages are processed using rectified linear unit functions. The rectified linear unit function is calculated by taking the larger value between the input value and zero.

[0121] The directed edges from the image node to the corresponding treatment text node that continues the original system treatment, the directed edges from the image node to the corresponding treatment text node that adds local treatment, and the weighted edge messages from the image node to the corresponding treatment text node that changes the system treatment are respectively input into the rectified linear unit function to obtain the updated representation of the corresponding node.

[0122] Configure a first output mapping vector for the treatment text node that continues the original system treatment, configure a second output mapping vector for the treatment text node that adds local treatment, and configure a third output mapping vector for the treatment text node that changes the system treatment.

[0123] Calculate the output scores for the treatment text nodes that continue the original system treatment, the treatment text nodes that add local treatment, and the treatment text nodes that change the system treatment. The output score is the product of the corresponding output mapping vector and the updated representation of the node.

[0124] Calculate the natural exponential function values ​​corresponding to the output scores of the treatment text nodes that continue the original system treatment, the treatment text nodes that add local treatment, and the treatment text nodes that change the system treatment, and calculate the natural exponential function values ​​and values.

[0125] Calculate the probability of the treatment text node corresponding to continue the original system treatment, the treatment text node corresponding to add local treatment, and the treatment text node corresponding to change the system treatment. The probability is the natural exponential function value corresponding to the node output score divided by the sum of the natural exponential function values.

[0126] The weighted edge message corresponding to the directed edge of the local treatment is the fusion result of the attention weight and the image-treatment semantic association message. The larger the value, the stronger the effective information transmitted from the image to the original system treatment node.

[0127] The weighted edge message corresponding to the directed edge of the system treatment is the fusion result of attention weight and image-local treatment semantic association message. The larger the value, the stronger the effective information transmitted from the image to the local treatment node.

[0128] The weighted edge message corresponding to the directed edge continuing the original system treatment is the fusion result of the attention weight and the semantic association message between the image and the replacement system treatment. The larger the value, the stronger the effective information transmitted from the image to the replacement system treatment node.

[0129] The Rectified Linear Unit (ReLU) function is an activation function that introduces nonlinear feature transformation. Its core function is to filter out negative information in weighted edge messages (negative values ​​are set to zero), retaining only positive valid information and avoiding negative interference from affecting the update of the treatment node representation.

[0130] The updated representation of the treatment text node corresponding to the continuation of the original system treatment is a new feature after fusing effective image information into the continuation of the original system treatment node. It replaces the initial representation for subsequent output probability calculation and reflects the strength of image support for the treatment option.

[0131] The updated representation of the treatment text node corresponding to the addition of local treatment is a new feature after fusing effective image information with the addition of local treatment node. It replaces the initial representation for subsequent output probability calculation and reflects the strength of image support for the treatment option.

[0132] The updated representation of the treatment text node corresponding to the system treatment is a new feature after the system treatment node is replaced and the effective information of the image is fused. It replaces the initial representation and is used for subsequent output probability calculation, reflecting the strength of the image's support for the treatment option.

[0133] The first output mapping vector is a linear mapping carrier that transforms the updated representation of the original system's treatment nodes into output scores, ensuring that the updated features can adapt to the dimensional requirements of probability calculation. Specifically, the dimension of the first output mapping vector is consistent with the dimension of the updated representation of the corresponding treatment text node (e.g., if the updated representation is 256-dimensional, then the vector is 256-dimensional). The initial values ​​are initialized using Xavier (randomly sampled from a normal distribution with a mean of 0 and a variance of 2 / the dimension of the updated representation).

[0134] The second output mapping vector is a linear mapping carrier that transforms the updated representation of the local treatment node into the output score. Its function is the same as the first output mapping vector, only the parameter values ​​are different (to adapt to the semantics of the local treatment node). Specifically, its dimension is the same as the dimension of the updated representation of the local treatment node (e.g., 256 dimensions), the initial value is initialized using Xavier, and the parameters are independent of the first output mapping vector (they do not share parameters).

[0135] The third output mapping vector is a linear mapping carrier that transforms the updated representation after changing the system treatment node into an output score. Its function is the same as the first two types of output mapping vectors, only the parameter values ​​are different (to adapt to the semantics of changing the system treatment). Specifically, its dimension is the same as the dimension of the updated representation after changing the system treatment node (e.g., 256 dimensions), the initial value is initialized using Xavier, and the parameters are independent of the first and second output mapping vectors.

[0136] The output score of the treatment text node corresponding to the continuation of the original system treatment is an indicator that quantifies the intermediate strength of the transformation of the updated representation of the continuation of the original system treatment node into probability. The larger the value, the higher the recommendation probability in the subsequent calculation. Specifically, the output score is equal to the inner product of the first output mapping vector and the vector of the updated representation of the continuation of the original system treatment node (i.e., the sum of the corresponding dimension elements).

[0137] The output score of the treatment text node corresponding to the addition of local treatment is an indicator that quantifies the intermediate strength of the transformation of the updated representation of the local treatment node into probability. The larger the value, the higher the recommendation probability in the subsequent calculation. Specifically, the output score is equal to the inner product of the second output mapping vector and the vector of the updated representation of the local treatment node.

[0138] The output score of the treatment text node corresponding to the system treatment change is an indicator that quantifies the intermediate strength of the transformation of the updated representation of the system treatment node into probability. The larger the value, the higher the recommendation probability in the subsequent calculation. Specifically, the output score is equal to the inner product of the third output mapping vector and the vector of the updated representation of the system treatment node.

[0139] The natural exponential function value is a non-linear amplification of the output score.

[0140] The sum of the natural exponential function values ​​is the cumulative sum of the natural exponential function values ​​corresponding to the output scores of the three treatment text nodes. Its function is to serve as the denominator for probability normalization, ensuring that the sum of the probabilities of the three treatment options is 1.

[0141] The probability is the final recommended probability of continuing the original systemic treatment, adding local treatment, or changing the systemic treatment option. It takes a value of 0-1. The larger the value, the higher the confidence of the treatment plan recommended by the model.

[0142] In one embodiment of the present invention, supervised training is performed using labels extracted from historical treatment texts corresponding to three treatment text nodes. Simultaneously, shape alignment constraints are used to ensure that the gating mapping and the spatially weighted oligoprogression mixture margin change in the same direction, resulting in a target graph neural network, comprising:

[0143] Three tags are extracted from the historical treatment text. Each tag corresponds to a treatment text node that continues the original system treatment, a treatment text node that adds local treatment, and a treatment text node that changes the system treatment. Each tag has a value of zero or one, and the sum of the values ​​of the three tags is one.

[0144] To calculate the supervised classification loss, first calculate the product of the logarithm of the probability of each label and the corresponding treatment text node, then calculate the sum of the three products, and finally take the negative of the sum to obtain the supervised classification loss.

[0145] Set the shape alignment mapping slope parameter, and use a logic function to construct the shape alignment mapping. The logic function is calculated as: one divided by one and the sum of the natural exponential function value. The exponent of the natural exponential function is: the product of the negative spatial weighted oligoprogression mixing margin and the shape alignment mapping slope parameter. Input the spatial weighted oligoprogression mixing margin into the logic function to obtain the target gating curve result.

[0146] To calculate the shape alignment loss, first calculate the difference between the continuation direction weight and the target gating curve result, then calculate the square of the difference to obtain the shape alignment loss;

[0147] Set a weight parameter, which is a value greater than zero, and calculate the sum of the products of the supervised classification loss, the weight parameter, and the shape alignment loss to obtain the total loss;

[0148] Define the parameter set, which includes image coding network parameters, forward weight matrix group, backward weight matrix group, attention projection vector, output mapping vector, gating mapping slope parameter, gating mapping bias parameter, attention temperature slope parameter, and shape alignment mapping slope parameter.

[0149] Set a learning rate parameter, which is a value greater than zero, and calculate the gradient of the total loss with respect to each parameter in the parameter set;

[0150] The updated parameter set is obtained by subtracting the product of the learning rate parameter and the gradient of the corresponding parameter from each parameter in the parameter set.

[0151] Based on the updated parameter set, the target graph neural network is obtained.

[0152] Historical treatment texts are textual data that stores records of a patient's past treatment decisions (such as treatment plan descriptions, medical orders, and treatment instructions in electronic medical records). They must contain clear statements regarding continuing the original systemic treatment, adding local treatment, or changing systemic treatment. They are the original data source for extracting true treatment labels.

[0153] The three labels are binary identifiers that quantify real historical treatment decisions. 1 represents that the treatment option is the actual clinically adopted plan, and 0 represents that it was not adopted. The constraint that the sum is one is consistent with the actual clinical scenario of selecting only one core treatment strategy each time (such as when choosing to add local treatment, the label is [0,1,0]).

[0154] Supervised classification loss is an indicator that measures the difference between the model's output treatment probability and the historical true label. The smaller the value, the higher the consistency between the model's prediction and the actual clinical decision.

[0155] The shape alignment mapping slope parameter is a parameter that adjusts the steepness of the target gating curve. The larger the value, the more sensitive the curve is to changes in input (small changes in input can cause large fluctuations in output), and the smaller the value, the flatter the curve (changes in input have a weak impact on output). Specifically, the initial value of the shape alignment mapping slope parameter is set in the range of 0.1-0.5, and the specific value is determined through cross-validation (different values ​​are tested in the training set, and the value that minimizes the shape alignment loss is selected. For example, if the minimum loss in the validation set corresponds to 0.3, then the parameter is set to 0.3).

[0156] One divided by the sum of one and the natural exponential function value, the exponent of the natural exponential function is: the product of the negative spatially weighted oligo-progression mixing margin and the shape alignment mapping slope parameter; Explanation: The shape alignment mapping is a function that generates the ideal reference curve that the gating mapping should follow. Its role is to provide a benchmark for the continuation direction weights that change in the same direction as the spatially weighted oligo-progression mixing margin, ensuring that the gating mapping does not have a logical reversal.

[0157] The target gating curve result is the output value of the shape alignment mapping, which ranges from 0 to 1 and increases monotonically with the increase of the spatially weighted oligo-progression mixing margin. It is a reference standard for judging whether the weight of the continuation direction is logical.

[0158] Shape alignment loss is an indicator that measures the degree of deviation between the continuation direction weight and the target gating curve result. The smaller the value, the stronger the homogeneity between the continuation direction weight and the spatially weighted oligoprogression mixing margin.

[0159] The weight parameter is a coefficient that balances the proportion of supervised classification loss (fitting historical data) and shape alignment loss (ensuring logical consistency) in the total loss. The larger the value, the stronger the influence of shape alignment constraints on model training; the smaller the value, the stronger the influence of supervised classification. Specifically, the initial value of the weight parameter is set in the range of 0.01-0.1, and it is dynamically adjusted during training (if the continuation direction weight and the spatially weighted oligoprogression mixture margin are reversed, increase the weight parameter to 0.05-0.1; if the model prediction deviates too much from the historical labels, decrease the weight parameter to 0.01-0.03).

[0160] Total loss is an optimization metric that combines the model’s data fitting ability with logical consistency. The goal of model training is to minimize total loss, ensuring that the model not only fits historical treatment decisions but also conforms to the logical relationship between spatially weighted oligoprogression mixture margin and gating mapping.

[0161] The parameter set is the collective term for all learnable parameters in a model that need to be optimized through training. It covers the entire process of image feature extraction, message passing, attention calculation, and output mapping, and is a decisive factor in model performance.

[0162] Image coding network parameters are network parameters that transform the original tumor medical contrast image features (such as lesion grayscale, edge, and volume) into the initial representation of image nodes (such as the convolution kernel weights and biases of CNNs, the self-attention weights of Transformers, etc.), which directly affect the accuracy of image feature extraction.

[0163] The learning rate parameter controls the magnitude of each parameter update. If the value is too large, it can easily lead to parameter oscillation and non-convergence; if the value is too small, it can easily lead to slow training speed. Specifically, the initial value of the learning rate parameter is set in the range of 0.0001-0.001, and a learning rate decay strategy is adopted (every 10 epochs of training, the learning rate is multiplied by 0.9 until it drops to 1 / 10 of the initial value and then remains unchanged).

[0164] The gradient is the partial derivative of the total loss with respect to a single parameter. It reflects the direction and degree of the influence of parameter changes on the total loss (positive gradient indicates that increasing the parameter leads to an increase in loss, and negative gradient indicates that increasing the parameter leads to a decrease in loss) and is the core basis for parameter updates. Specifically, the gradient is calculated using the backpropagation algorithm. Starting from the total loss, the gradient is derived in reverse according to the forward calculation process of the model, and the partial derivative of each parameter with respect to the total loss is calculated in turn (e.g., the gradient of convolutional layer parameters is calculated through convolution backpropagation, and the gradient of fully connected layer parameters is calculated through the chain rule).

[0165] The updated parameter set is the combination of parameters optimized after one training iteration. The update direction of each parameter is opposite to the gradient (to ensure that the total loss is reduced), and the update magnitude is controlled by the learning rate parameter.

[0166] The target graph neural network is the final model obtained after the parameter set has been trained and iterated for multiple rounds (until the total loss converges, and the convergence criterion is that the change in total loss is less than 0.001 for 5 consecutive epochs). It has stable auxiliary decision-making ability and can output the probability of three types of treatment options by inputting new images from two adjacent periods.

[0167] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A personalized treatment method for cancer patients integrating artificial intelligence, characterized in that, include: After acquiring two adjacent contrast-enhanced tumor images, registering and segmenting the lesions, a spatially weighted oligoprogression margin is generated based on preset parameters, including: The relative volume change ratio is obtained based on the change in the current volume of each lesion compared to the previous volume; When the relative volume change ratio is greater than the preset significant growth threshold, the corresponding lesion is marked as a candidate lesion for growth. Based on the spatial location of each lesion, the spatial aggregation center of the growing candidate lesion set is obtained; Set spatial kernel scale parameters, calculate the squared value of the distance between the spatial location of each lesion and the spatial aggregation center, and obtain the spatial distance weight of each lesion through the squared value of the distance to the spatial aggregation center; When the difference between the relative volume change ratio of each lesion and the preset significant growth threshold is positive, the difference is taken as the over-threshold growth amount; calculate the sum of the over-threshold growth amounts of all lesions, divide the over-threshold growth amount of each lesion by the sum of the over-threshold growth amounts, and obtain the amplitude normalization weight of each lesion. Multiply the spatial distance weight of each lesion by the amplitude normalization weight to obtain the joint weight of each lesion; calculate the sum of the products of the joint weights of all lesions and the corresponding growth candidate markers to obtain the effective progression count; calculate the sum of the joint weights of all lesions to obtain the total effective count. Set a small-scale boundary parameter, a priori shape parameter one, and a priori shape parameter two. Calculate the first sum of the priori shape parameter one and the effective progress count, and the second sum of the priori shape parameter two and the difference between the effective total count and the effective progress count. Using a small-scale boundary parameter as the cutoff value, calculate the regularized incomplete beta function values ​​corresponding to the first and second sums; The spatially weighted oligoprogression mixture margin is obtained by calculating the logarithm of the ratio of the regularized incomplete beta function value to a value minus the regularized incomplete beta function value. Construct a graph containing one image node and three treatment text nodes. The three treatment text nodes correspond to continuing the original system treatment, adding local treatment, and changing the system treatment, respectively. Establish directed edges from the image node to each treatment text node, and write the spatial weighted oligoprogression mixing margin as an edge attribute into each directed edge. Gating mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain orientation weights, and edge messages of each directed edge are generated by combining the initial representation of the image node; attention temperature mapping is performed based on spatially weighted oligoprogression hybrid margin to obtain attention coefficients of each directed edge and normalize them; The representations of the three treatment text nodes are updated by weighting the edge messages with attention coefficients, and the probabilities corresponding to the three treatment text nodes are output respectively. Supervised training is performed using labels extracted from historical treatment texts corresponding to three treatment text nodes. Simultaneously, shape alignment constraints are used to ensure that the gating mapping and spatially weighted oligoprogression mixing margin change in the same direction, resulting in a target graph neural network. The target graph neural network is used to output the probability of continuing the original system treatment, adding local treatment, or changing the system treatment.

2. The method for individualized treatment of cancer patients integrating artificial intelligence according to claim 1, characterized in that, Construct a graph containing one image node and three treatment text nodes. The three treatment text nodes correspond to continuing the original system treatment, adding local treatment, and changing the system treatment, respectively. Establish directed edges from the image node to each treatment text node, and write the spatially weighted oligoprogression mixture margin as an edge attribute into each directed edge, including: Define a single-stage graph, which includes a set of nodes, a set of directed edges, and edge attribute functions; The node set contains an image node, a treatment text node corresponding to continuing the original system treatment, a treatment text node corresponding to adding local treatment, and a treatment text node corresponding to changing the system treatment; The set of directed edges contains three directed edges: the first directed edge is from the image node to the corresponding treatment text node that continues the original system treatment; the second directed edge is from the image node to the corresponding treatment text node that adds local treatment; and the third directed edge is from the image node to the corresponding treatment text node that changes the system treatment. The edge attribute function is used to assign an attribute value to each directed edge, which is the spatially weighted oligo-progression mixture margin; Define an edge attribute vector that contains the attribute values ​​of the first directed edge, the second directed edge, and the third directed edge in sequence.

3. The method for individualized treatment of cancer patients integrating artificial intelligence according to claim 2, characterized in that, The orientation weights are obtained by performing gated mapping based on spatially weighted oligoprogression hybrid margin, and edge messages for each directed edge are generated by combining the initial representation of the image nodes, including: Set the gate mapping slope parameter and the gate mapping bias parameter, and use a logic function to perform the gate mapping; wherein, the logic function is calculated by dividing one by the sum of one and the natural exponential function value, where the exponent of the natural exponential function is a negative input value; To calculate the continuation direction weight, first calculate the product of the spatially weighted oligoprogression mixing margin and the gated mapping slope parameter, then add the product to the gated mapping bias parameter to obtain the input value, and substitute the input value into the logic function to obtain the continuation direction weight. Subtract the continuing direction weight from one to obtain the changing direction weight; Configure a set of positive weight matrices and a set of negative weight matrices for the treatment text nodes corresponding to continuing the original system treatment, the treatment text nodes corresponding to adding local treatment, and the treatment text nodes corresponding to changing the system treatment. For directed edges from image nodes to corresponding treatment text nodes that continue the original system treatment, from image nodes to corresponding treatment text nodes that add local treatment, and from image nodes to corresponding treatment text nodes that change the system treatment, first calculate the product of the positive weight matrix corresponding to the corresponding node and the initial representation of the image node, and multiply by the continuing direction weight to obtain the first result. Then calculate the product of the reverse weight matrix corresponding to the corresponding node and the initial representation of the image node, and multiply by the changing direction weight to obtain the second result. Add the first result and the second result to obtain the edge message of the corresponding directed edge.

4. The method for individualized treatment of cancer patients integrating artificial intelligence according to claim 3, characterized in that, Attention temperature mapping is performed based on spatially weighted oligo-progression hybrid margin to obtain and normalize the attention coefficients of each directed edge, including: The attention temperature slope parameter is set, and the product of the spatially weighted oligoprogression mixing margin and the attention temperature slope parameter is calculated. This product is used as the exponent of the natural exponential function to calculate the attention temperature parameter value. Configure a first attention projection vector for the treatment text node that continues the original system treatment, configure a second attention projection vector for the treatment text node that adds local treatment, and configure a third attention projection vector for the treatment text node that changes the system treatment. The correlation scores are calculated for directed edges from image nodes to treatment text nodes that continue the original system treatment, directed edges from image nodes to treatment text nodes that add local treatment, and directed edges from image nodes to treatment text nodes that change the system treatment. The correlation scores are obtained by multiplying the corresponding attention projection vector with the initial representation of the image node. Calculate the unnormalized attention coefficients of directed edges from image nodes to the corresponding treatment text nodes that continue the original system treatment, directed edges from image nodes to the corresponding treatment text nodes that add local treatment, and directed edges from image nodes to the corresponding treatment text nodes that change the system treatment. Calculate the sum of the unnormalized attention coefficients of the directed edges from the image node to the corresponding treatment text node that continues the original system treatment, the directed edges from the image node to the corresponding treatment text node that adds local treatment, and the directed edges from the image node to the corresponding treatment text node that changes the system treatment. Calculate the attention coefficients of directed edges from image nodes to the corresponding treatment text nodes that continue the original system treatment, directed edges from image nodes to the corresponding treatment text nodes that add local treatment, and directed edges from image nodes to the corresponding treatment text nodes that change the system treatment. The attention coefficient is the sum of the unnormalized attention coefficients of the corresponding directed edges divided by the unnormalized attention coefficients.

5. The method for individualized treatment of cancer patients integrating artificial intelligence according to claim 4, characterized in that, The representations of the three treatment text nodes are updated by weighting the corresponding edge messages with attention coefficients, and the probabilities corresponding to the three treatment text nodes are output, including: Calculate the weighted edge messages for directed edges from image nodes to the corresponding treatment text nodes that continue the original system treatment, directed edges from image nodes to the corresponding treatment text nodes that add local treatment, and directed edges from image nodes to the corresponding treatment text nodes that change the system treatment. The weighted edge message is the product of the attention coefficient of the corresponding directed edge and the corresponding edge message. The weighted edge messages are processed using rectified linear unit functions. The rectified linear unit function is calculated by taking the larger value between the input value and zero. The directed edges from the image node to the corresponding treatment text node that continues the original system treatment, the directed edges from the image node to the corresponding treatment text node that adds local treatment, and the weighted edge messages from the image node to the corresponding treatment text node that changes the system treatment are respectively input into the rectified linear unit function to obtain the updated representation of the corresponding node. Configure a first output mapping vector for the treatment text node that continues the original system treatment, configure a second output mapping vector for the treatment text node that adds local treatment, and configure a third output mapping vector for the treatment text node that changes the system treatment. Calculate the output scores for the treatment text nodes that continue the original system treatment, the treatment text nodes that add local treatment, and the treatment text nodes that change the system treatment. The output score is the product of the corresponding output mapping vector and the updated representation of the node. Calculate the natural exponential function values ​​corresponding to the output scores of the treatment text nodes that continue the original system treatment, the treatment text nodes that add local treatment, and the treatment text nodes that change the system treatment, and calculate the natural exponential function values ​​and values. Calculate the probability of the treatment text node corresponding to continue the original system treatment, the treatment text node corresponding to add local treatment, and the treatment text node corresponding to change the system treatment. The probability is the natural exponential function value corresponding to the node output score divided by the sum of the natural exponential function values.

6. The method for individualized treatment of cancer patients integrating artificial intelligence according to claim 5, characterized in that, Supervised training is performed using labels extracted from historical treatment texts corresponding to three treatment text nodes. Simultaneously, shape alignment constraints are used to ensure that the gating mapping and spatially weighted oligoprogression mixture margin change in the same direction, resulting in a target graph neural network, including: Three tags are extracted from the historical treatment text. Each tag corresponds to a treatment text node that continues the original system treatment, a treatment text node that adds local treatment, and a treatment text node that changes the system treatment. Each tag has a value of zero or one, and the sum of the values ​​of the three tags is one. To calculate the supervised classification loss, first calculate the product of the logarithm of the probability of each label and the corresponding treatment text node, then calculate the sum of the three products, and finally take the negative of the sum to obtain the supervised classification loss. Set the shape alignment mapping slope parameter, and use a logic function to construct the shape alignment mapping. The logic function is calculated as: one divided by one and the sum of the natural exponential function value. The exponent of the natural exponential function is: the product of the negative spatial weighted oligoprogression mixing margin and the shape alignment mapping slope parameter. Input the spatial weighted oligoprogression mixing margin into the logic function to obtain the target gating curve result. To calculate the shape alignment loss, first calculate the difference between the continuation direction weight and the target gating curve result, then calculate the square of the difference to obtain the shape alignment loss; Set a weight parameter, which is a value greater than zero, and calculate the sum of the products of the supervised classification loss, the weight parameter, and the shape alignment loss to obtain the total loss; Define the parameter set, which includes image coding network parameters, forward weight matrix group, backward weight matrix group, attention projection vector, output mapping vector, gating mapping slope parameter, gating mapping bias parameter, attention temperature slope parameter, and shape alignment mapping slope parameter. Set a learning rate parameter, which is a value greater than zero, and calculate the gradient of the total loss with respect to each parameter in the parameter set; The updated parameter set is obtained by subtracting the product of the learning rate parameter and the gradient of the corresponding parameter from each parameter in the parameter set. Based on the updated parameter set, the target graph neural network is obtained.

Images