Imaging automatic grading and presenting method for curative effects of tumor drugs

Abstract

The invention discloses an imaging automatic grading and presenting method for the curative effect of a tumor drug, and relates to the technical field of image processing.The method comprises the steps that a multi-stage dynamic enhanced image is obtained, and a blood supply activity spectrum representing the tumor microenvironment function is constructed; extracting multi-time-point image omics characteristics, constructing a time sequence and calculating a sample entropy so as to generate a time sequence heterogeneity index for quantifying dynamic evolution irregularity; jointly inputting the atlas and the indexes into a conditional generative network, and predicting and synthesizing a tumor heterogeneity evolution prediction map at a future evaluation time point; and finally, fusing the anatomical image, the blood supply map and the predictive map to generate an integrated view, calculating a comprehensive curative effect level in combination with blood supply and heterogeneity indexes, and automatically identifying a high-risk evolution region at the same time. According to the method, the problem that traditional evaluation is subjective and lagged is solved, conversion from retrospective analysis to prospective and spatialization prediction is achieved, and objective, accurate and visual decision support is provided for clinic.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to an automatic imaging grading and presentation method for the efficacy of tumor drugs. Background Technology

[0002] In the field of modern personalized cancer treatment, accurate and dynamic evaluation of drug efficacy is the cornerstone for optimizing treatment plans and improving patients' quality of life. Therefore, technological advancements in this field have evolved from traditional macroscopic morphological assessments that rely solely on changes in tumor size to comprehensive functional and structural assessments that integrate multimodal medical imaging.

[0003] However, the existing technology system has encountered two fundamental bottlenecks in its progress toward this advanced goal, which limit the further enhancement of its clinical application value.

[0004] In current technological practice, information from different imaging modalities and analytical dimensions—such as anatomical images characterizing tumor spatial boundaries, functional parametric maps reflecting the vascular microenvironment (e.g., blood supply activity), and radiomics features quantifying tissue disorder—is typically presented as independent image sequences, parametric maps, or quantitative data reports. This places a significant cognitive burden on clinicians, requiring them to rely on their professional experience to perform complex mental integration of this highly heterogeneous and multidimensional information in order to form a comprehensive judgment of the tumor's condition.

[0005] This integration process, reliant on subjective experience, is not only inefficient but also prone to introducing discrepancies in interpretation. The fragmentation of information leads to a fragmentation of insight. All assessment methods within the existing technological framework, regardless of their complexity, are essentially "retrospective" or "descriptive." Their function is limited to depicting changes in the tumor from the start of treatment to the current examination point. Existing technologies are incapable of predicting the evolutionary trends of tumors in future treatment cycles, particularly identifying and locating the spatial evolution of potential drug-resistant subregions. This results in clinical decision-making being passively relegated to a "response mode" rather than an active "anticipatory mode." Clinicians can only intervene after confirming tumor progression or drug resistance, unable to proactively deploy strategies based on predictions of future risks—undoubtedly a potential waste of the valuable treatment window. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an automated imaging grading and presentation method for the efficacy of tumor drugs, thereby resolving the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: Step 1: Obtain anatomical structure images and dynamic contrast-enhanced images of the patient at the baseline time point before drug treatment and at the drug cycle time point, and fuse them to obtain a four-dimensional image dataset.

[0008] Step 2: Based on the curve composed of the temporal signal intensity S(t) of the dynamic contrast-enhanced images in the four-dimensional image dataset, construct a voxel-level pharmacokinetic model to generate a tumor blood supply activity map. After iteratively optimizing the pharmacokinetic model based on plasma transport rate and extracellular space volume fraction, calculate the average blood supply activity index of the entire tumor volume of interest, preset the physiological efficacy threshold, and preliminarily determine the immediate efficacy of the drug on the tumor vascular system.

[0009] Step 3: Based on the four-dimensional image dataset, extract radiomics features through image processing technology, and construct a temporal heterogeneity index, denoted as THI, to characterize the complexity of the evolution of tumor internal structure over treatment time; preset a drug resistance risk threshold to assess the potential risk of tumors developing heterogeneous drug resistance.

[0010] Step 4: Using the tumor blood supply activity map as a spatial condition and the time series of the temporal heterogeneity index (THI) as an evolutionary driver, input the data into a pre-trained conditional image generation network to synthesize a tumor heterogeneity evolution prediction map at a future evaluation time point Tn+1. Simultaneously, by calculating the structural similarity between the tumor heterogeneity evolution prediction map and the actual evolution map, a generation fidelity index, denoted as GFI, is obtained. If the GFI is qualified, image retention processing is performed; otherwise, image rejection processing is performed.

[0011] Step 5: Using the anatomical structure image as the base layer, fuse and overlay the tumor blood supply activity map and the preserved tumor heterogeneity evolution prediction map to generate a fused view; and combine the average blood supply activity index and the temporal heterogeneity index to automatically output a comprehensive drug efficacy level.

[0012] Preferably, at the baseline time point and the drug cycle time point, a first magnetic resonance imaging scan is performed on the tumor region to acquire anatomical structure images; after bias field correction of the anatomical structure images using the N4ITK algorithm, a first reference image set is obtained.

[0013] At the baseline time point and the drug cycle time point, a second dynamic contrast-enhanced magnetic resonance imaging scan was performed on the same tumor region. By intravenously injecting contrast agent and performing continuous rapid imaging, second time-series image data reflecting the perfusion and penetration process of contrast agent in the tumor were acquired.

[0014] Using the first reference image set at the baseline time point as a reference benchmark, a non-rigid image registration algorithm is employed to perform 3D spatial correction on the second time-series image data at the drug cycle time point. Then, within the aligned first reference image set at the baseline time point, a deep learning segmentation model is used to identify and obtain the 3D volume of interest defining the spatial extent of the tumor. Specifically, this includes:

[0015] The first reference image set of the aligned baseline time points is used as the three-dimensional voxel input data and fed into the pre-trained three-dimensional convolutional neural network segmentation model to perform forward propagation calculation. The model performs hierarchical feature extraction and semantic classification on each voxel in the three-dimensional voxel input data, thereby generating a three-dimensional probability map.

[0016] Threshold segmentation is performed on the three-dimensional probability map, and the set of voxels whose probability values ​​are not less than a preset segmentation threshold τ is defined as the three-dimensional volume of interest.

[0017] Based on the three-dimensional volume of interest, the signal intensity value of each voxel within the tumor region is extracted from all aligned second temporal image data, normalized, and a four-dimensional image dataset with a unified coordinate system is constructed.

[0018] Preferably, in the four-dimensional image dataset, the signal intensity values ​​of each voxel at all dynamic contrast-enhanced scanning time points are extracted to form a curve composed of time-series signal intensity S(t), and the time-series signal intensity S(t) is converted into contrast agent concentration C using the relationship between magnetic resonance signal intensity and contrast agent concentration. data (t);

[0019] The conversion steps are as follows: Contrast agent molecules can alter the magnetic field environment of the surrounding water molecules, thereby shortening the longitudinal relaxation time of the tissue. This physical change is related to the temporal signal intensity S(t) and the contrast agent concentration C. data (t) all have definite mathematical relationships, and the steps are as follows:

[0020] For each time point t on the temporal signal intensity S(t), the dynamic longitudinal relaxation time T1(t) corresponding to the current time point is solved by inversely based on the signal equation of magnetic resonance imaging.

[0021] Convert the dynamic longitudinal relaxation time T1(t) corresponding to the current time point into its reciprocal to obtain the dynamic relaxation rate R1(t);

[0022] Subtracting the baseline relaxation rate R10 of the tissue before contrast agent injection from the dynamic relaxation rate R1(t) yields the net increase in relaxation rate ΔR1(t) that is entirely contributed by the contrast agent.

[0023] The net increase in relaxation rate ΔR1(t) caused by contrast agent is related to the contrast agent concentration C in the tissue. data(t) shows a linear proportional relationship, and the contrast agent concentration C is obtained after conversion. data (t).

[0024] Preferably, the contrast agent concentration C of each voxel within the three-dimensional volume of interest is obtained. data (t), and the contrast agent concentration C for each voxel. data (t) Perform voxel-level pharmacokinetic model fitting to determine the plasma transport rate K, which characterizes vascular permeability. trans and the extravascular extracellular space volume fraction V, which characterizes tissue structure e ;

[0025] The voxel-level pharmacokinetic model fitting process is implemented through an iterative optimization algorithm, specifically including:

[0026] Based on prior physiological knowledge, the plasma transport rate K... trans and extravascular extracellular space volume fraction V e Set initial parameter values ;

[0027] Start an iterative optimization loop, and perform the following operations in each iteration:

[0028] Residual and Jacobian matrix calculation: Based on the parameter value P of the current iteration and the preset pharmacokinetic model, calculate the first residual sum of squares SSR1;

[0029] The pharmacokinetic model was used to predict the contrast agent concentration, and the predicted contrast agent concentration C was calculated. model (t) and the contrast agent concentration C data The differences between (t) at each time point constitute the residual vector r:

[0030] Simultaneously, the predicted concentration of K was calculated using a voxel-level pharmacokinetic model. trans and V e partial derivatives , forming the Jacobian matrix J;

[0031] And solve the system of linear equations to obtain the parameter update step size ΔP;

[0032] The experimental parameter value Ptrial = P + ΔP is generated based on the calculated parameter update step size ΔP, and a second experimental residual sum of squares SSR2 is calculated based on the experimental parameter value Ptrial. The second experimental residual sum of squares SSR2 is compared with the first residual sum of squares SSR1 to evaluate the effectiveness of the parameter update step size ΔP, including:

[0033] To ensure the stable convergence of the iterative optimization loop and prevent optimization divergence caused by excessive parameter update step size ΔP due to the high nonlinearity of the pharmacokinetic model, a non-negative adaptive damping factor λ is introduced to dynamically adjust between the fast-converging Gauss-Newton method and the robust gradient descent method.

[0034] If SSR2 is less than SSR1, the update step size ΔP is determined to be valid, and the current parameter value P is updated to the experimental parameter value Ptrial, while the damping factor λ is decreased; conversely, if SSR2 is greater than or equal to SSR1, the parameter update step size ΔP is determined to be invalid, the current parameter value P is kept unchanged, and the damping factor λ is increased.

[0035] A first threshold and a second threshold are set. If the norm of the parameter update step size ΔP is less than the first threshold, or the relative change of the residual sum of squares is less than the second threshold, or the number of iterations reaches the maximum value, the iteration is terminated and the final parameter value P is used as the fitting result of the current voxel; if the conditions are not met, the next iteration continues.

[0036] Preferably, the final parameter value P is extracted after each termination iteration, and the plasma transport rate K of each voxel within parameter value P is extracted. trans and extravascular extracellular space volume fraction V e Arrange each voxel on its original three-dimensional spatial coordinates (x, y, z) to generate K. trans 3D parametric map and V e Three-dimensional parametric map;

[0037] Extract K trans K of all voxels in the 3D parametric map trans The arithmetic mean of the values ​​is used to obtain the average blood supply activity index;

[0038] Extract V e V of all voxels in the 3D parametric map e The arithmetic mean of the values ​​is used to obtain the average extracellular space volume fraction.

[0039] A preset physiological efficacy threshold is set. If the average blood supply activity index is less than the physiological efficacy threshold, it means that the drug inhibits the vascular permeability and blood flow perfusion of the tumor within a given cycle, achieving the expected anti-angiogenic or vascular normalization effect, and generating a first-action efficacy label.

[0040] If the average blood supply activity index is greater than or equal to the physiological effectiveness threshold, it means that the drug failed to effectively inhibit the blood supply activity of the tumor within a given cycle, and the effect was unqualified, generating a secondary action invalid label.

[0041] and K trans 3D parametric map and V eThe three-dimensional parametric maps are fused to form a tumor blood supply activity map.

[0042] Preferably, based on the four-dimensional image dataset, within the three-dimensional volume of interest, for the three-dimensional image data at each time point, several radiomics features of tumor internal texture and morphology are extracted.

[0043] For each of the several types of image omics features, a time series consisting of its feature values ​​at all time points is constructed to obtain several types of time feature sequences;

[0044] For the constructed time feature series, its sample entropy value is calculated to quantify the irregularity of the evolution of the corresponding radiomics features over time; and all the sample entropy values ​​calculated from all radiomics features are combined in a weighted linear combination to generate the temporal heterogeneity index (THI).

[0045] A drug resistance risk threshold is preset, and the temporal heterogeneity index (THI) is compared with the drug resistance risk threshold;

[0046] If the temporal heterogeneity index (THI) is greater than the drug resistance risk threshold, the tumor's internal structure is determined to be unstable, with a potential risk of developing heterogeneous drug resistance, and a third high-risk drug resistance label is generated; otherwise, a fourth low-risk drug resistance label is generated.

[0047] Preferably, radiomics features include first-order statistical features and second-order statistical features;

[0048] First-order statistical features include: the sum of squares of voxel intensity values, the entropy value (a measure of the disorder of voxel intensity distribution), the skewness of gray-level distribution, and the kurtosis of gray-level distribution;

[0049] Second-order statistical features include: texture roughness, morphological sphericity, and maximum three-dimensional diameter.

[0050] Preferably, the tumor blood supply activity map generated in step two is obtained. The tumor blood supply activity map is used as a spatial condition, and the time series of the temporal heterogeneity index (THI) mentioned in step three is used as the temporal evolution driving data. Both are input into a pre-trained conditional image generation network. Forward reasoning is performed through the conditional image generation network to predict and synthesize the tumor heterogeneity evolution prediction map at the preset future evaluation time point Tn+1.

[0051] The conditional generative adversarial network includes: a generator that receives spatial conditions and temporal evolution drivers and generates a tumor heterogeneity evolution prediction map, and a discriminator that distinguishes the differences between the generated tumor heterogeneity evolution prediction map and the real evolution map;

[0052] After obtaining the true evolution map at the future assessment time point Tn+1 and extracting the true tumor heterogeneity feature map based on the true evolution map, the similarity is measured with the tumor heterogeneity evolution prediction map at the future assessment time point Tn+1 to obtain the fidelity index GFI.

[0053] A preset fidelity threshold is set, preferably 0.85. When the fidelity index (GFI) is lower than the fidelity threshold, it is considered unqualified, and the tumor heterogeneity evolution prediction map at the current future evaluation time point Tn+1 is removed. When the fidelity index (GFI) is greater than or equal to the fidelity threshold, it is considered qualified, and the image is retained.

[0054] Preferably, the anatomical structure image is used as the base rendering layer, and within its three-dimensional coordinate space, the tumor blood supply activity map and the tumor heterogeneity evolution prediction map at the future assessment time point Tn+1 are fused to generate a fused view that can simultaneously display the tumor anatomical location, current blood supply status and future heterogeneity evolution trend.

[0055] Based on the blood supply activity index and the temporal heterogeneity index, the drug efficacy index is obtained by a weighted summation method, and a first drug efficacy threshold and a second drug efficacy threshold are preset.

[0056] When the drug efficacy index is less than the second drug efficacy threshold, the first drug efficacy level is generated.

[0057] When the drug efficacy index is between the first drug efficacy threshold and the second drug efficacy threshold, a second drug efficacy level is generated.

[0058] When the drug efficacy index is greater than the first drug efficacy threshold, a third drug efficacy level is generated.

[0059] Preferably, a critical roughness threshold segmentation operation is performed on the tumor heterogeneity evolution prediction map at the future evaluation time point Tn+1, the predicted texture roughness value of each voxel is extracted, a critical roughness threshold is preset, and voxels with texture roughness values ​​higher than or equal to the critical roughness threshold are aggregated to automatically generate targeted high-risk evolution regions.

[0060] This invention provides an automated imaging grading and presentation method for the efficacy of tumor drugs. It has the following beneficial effects:

[0061] (1) By accurately fitting the pharmacokinetic model at the voxel level, the original image signal is transformed into a blood supply activity index with clear physiological significance, and an objective judgment is made based on the preset physiological effectiveness threshold. This achieves a fundamental shift from traditional, morphology-based subjective assessment to standardized, microenvironment-based functional quantification, which promotes the improvement of the accuracy and early sensitivity of efficacy assessment.

[0062] (2) By constructing time series of radiomics features and introducing sample entropy to quantify the irregularity of their dynamic evolution, the dynamic instability of tumor internal structures can be captured and condensed into a temporal heterogeneity index. This elevates the assessment from a static description of the current state to a prospective prediction of future drug resistance risks, providing a quantitative early warning signal for intervention before clinical drug resistance manifests.

[0063] (3) By utilizing a conditional image generation network, the blood supply activity map representing the current physiological basis is combined with the heterogeneity index driving temporal evolution, enabling accurate prediction of the three-dimensional spatial distribution of future high-risk areas. This breaks through the limitations of traditional methods that can only provide global, non-spatial risk indicators, and for the first time, it visually identifies potential drug resistance "hotspot" subregions, providing a basis for decision-making to achieve precise spatial intervention.

[0064] (4) By generating an integrated view that incorporates tumor anatomical location, current blood supply status, and future evolution trends, and by integrating multi-dimensional quantitative indicators into a single, graded drug efficacy index, complex analytical results are transformed into intuitive, easy-to-understand, and standardized clinical conclusions. This not only reduces the information integration burden on clinicians, but also improves the efficiency and feasibility of clinical decision-making by automatically delineating high-risk target areas and directly transforming abstract predictions into specific and actionable intervention goals. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the steps of the present invention;

[0066] Figure 2 This is a flowchart illustrating steps one, two, and three of the present invention;

[0067] Figure 3 This is a schematic diagram of step four of the present invention;

[0068] Figure 4 This is a schematic diagram of step five of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Example 1

[0071] Please see Figure 1 This invention provides an automatic imaging grading and presentation method for the efficacy of tumor drugs, comprising the following steps:

[0072] Step 1: Obtain anatomical images and dynamic contrast-enhanced images of the patient at baseline time points before drug treatment and at drug cycle time points, and fuse them to obtain a four-dimensional image dataset.

[0073] Step 2: Based on the curve composed of the temporal signal intensity S(t) of the dynamic contrast-enhanced images in the four-dimensional image dataset, construct a voxel-level pharmacokinetic model to generate a tumor blood supply activity map. After iteratively optimizing the pharmacokinetic model based on plasma transport rate and extracellular space volume fraction, calculate the average blood supply activity index of the entire tumor volume of interest, preset the physiological efficacy threshold, and preliminarily determine the immediate efficacy of the drug on the tumor vascular system.

[0074] Step 3: Based on the four-dimensional image dataset, extract radiomics features through image processing technology, and construct a temporal heterogeneity index, denoted as THI, to characterize the complexity of the evolution of tumor internal structure over treatment time; preset a drug resistance risk threshold to assess the potential risk of tumors developing heterogeneous drug resistance.

[0075] Step 4: Using the tumor blood supply activity map as a spatial condition and the time series of the temporal heterogeneity index (THI) as an evolutionary driver, input the data into a pre-trained conditional image generation network to synthesize a tumor heterogeneity evolution prediction map at a future evaluation time point Tn+1. Simultaneously, by calculating the structural similarity between the tumor heterogeneity evolution prediction map and the actual evolution map, a generation fidelity index, denoted as GFI, is obtained. If the GFI is qualified, image retention processing is performed; otherwise, image rejection processing is performed.

[0076] Step 5: Using the anatomical structure image as the base layer, fuse and overlay the tumor blood supply activity map and the preserved tumor heterogeneity evolution prediction map to generate a fused view; and combine the average blood supply activity index and the temporal heterogeneity index to automatically output a comprehensive drug efficacy level.

[0077] In this embodiment, the present invention, through the fusion view generation technology in step five, seamlessly overlays the underlying anatomical structure, the tumor blood supply activity map representing the current physiological state, and the heterogeneity prediction map indicating future evolution trends into a unified three-dimensional space. This effectively solves the problem of fragmented and isolated multidimensional information in existing technologies, transforming complex analysis results into an intuitive and unified decision-making interface, reducing the cognitive integration load on clinicians, and making the comprehensive judgment of the current state and future risks of tumors more efficient and accurate. The conditional image generation network in step four breaks through the fundamental limitation of existing assessment methods that can only perform retrospective descriptions. It enables the prospective prediction of the spatial evolution trend of tumor internal heterogeneity at future time points, allowing clinical decision-making to shift from a passive "response mode" to an active "predictive mode." Finally, the comprehensive efficacy level automatically generated by combining the quantitative indicators in steps two and three provides objective and reliable data support for formulating prospective and individualized treatment strategy adjustments, helping to seize the valuable treatment window.

[0078] Example 2

[0079] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 2 Specifically, at the baseline time point and the drug cycle time point, a first magnetic resonance imaging scan is performed on the tumor region to acquire anatomical structure images; the N4ITK algorithm is used to perform bias field correction on the anatomical structure images to eliminate signal intensity inhomogeneity and obtain a first reference image set;

[0080] At the baseline time point and the drug cycle time point, a second dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) scan was performed on the same tumor region. By intravenously injecting contrast agent and performing continuous rapid imaging, second time-series image data reflecting the perfusion and penetration process of contrast agent in the tumor were acquired.

[0081] Using the first reference image set at the baseline time point as a reference benchmark, a non-rigid image registration algorithm is used to perform three-dimensional spatial correction on the second time-series image data at the drug cycle time point. Then, in the aligned first reference image set at the baseline time point, a deep learning segmentation model is used to identify and obtain the three-dimensional volume of interest that defines the spatial extent of the tumor.

[0082] The step of using a deep learning segmentation model to identify and obtain the three-dimensional volume of interest that defines the spatial extent of the tumor specifically includes:

[0083] The first reference image set of the aligned baseline time points is used as three-dimensional voxel input data and fed into a pre-trained three-dimensional convolutional neural network segmentation model to perform forward propagation calculation. The model performs hierarchical feature extraction and semantic classification on each voxel in the three-dimensional voxel input data, thereby generating a three-dimensional probability map.

[0084] Threshold segmentation is performed on the three-dimensional probability map, and the set of voxels whose probability values ​​are not less than a preset segmentation threshold τ is defined as the three-dimensional volume of interest.

[0085] Each voxel in the three-dimensional probability map has a probability value between 0 and 1, representing its belonging to the tumor region.

[0086] The three-dimensional probability map is subjected to voxel-level binarization processing. Specifically, for any voxel with coordinates (x,y,z), the voxel probability value P(x,y,z) is compared with a preset segmentation threshold τ. If P(x,y,z)≥τ, the output value of the current voxel is set to 1; otherwise, it is 0 and used as a background voxel.

[0087] All voxels with an output value of 1 are statistically analyzed to form a binary mask region representing the three-dimensional spatial boundary of the tumor, thus obtaining the three-dimensional volume of interest.

[0088] The steps for the joint calibration experiment of segmentation threshold τ are as follows: Prepare a calibration dataset containing multiple groups of cases (e.g., N≥50 cases), which should cover typical clinical scenarios with different tumor sizes, shapes, locations and image qualities;

[0089] For each case in this dataset, there is a three-dimensional probability map generated by the method of the present invention, as well as a three-dimensional tumor mask that is manually drawn and jointly confirmed by one or more senior radiologists as the "gold standard".

[0090] The search range for the preset segmentation threshold τ is set to [0.2, 0.9], and the search step size is 0.01;

[0091] For each candidate segmentation threshold τ value within the search range:

[0092] Volume-level binarization is performed on the 3D probability maps of all cases in the calibration dataset using the current candidate τ value to generate segmentation results to be evaluated;

[0093] The segmentation results to be evaluated are compared with the corresponding "gold standard" mask at the voxel level, and the similarity index is calculated by the distance similarity method. The preferred value of the preset segmentation threshold τ is determined to be the value that maximizes the similarity index on the calibration dataset, and this preferred value usually falls within the range of [0.4, 0.7]; in specific cases without specific clinical preferences, the preferred value is 0.5.

[0094] Based on the three-dimensional volume of interest, the signal intensity value of each voxel within the tumor region is extracted from all aligned second temporal image data, normalized, and a four-dimensional image dataset with a unified coordinate system is constructed.

[0095] In the four-dimensional image dataset, for each voxel located within the three-dimensional volume of interest, the signal intensity values ​​of each voxel at all dynamic contrast-enhanced scanning time points (t=0,1,2,...,n), where n represents the total number of time points, are extracted to form a curve composed of temporal signal intensity S(t). Then, using the relationship between magnetic resonance signal intensity and contrast agent concentration, the temporal signal intensity S(t) is converted into contrast agent concentration C. data (t);

[0096] The conversion steps are as follows: Contrast agent molecules can alter the magnetic field environment of the surrounding water molecules, thereby shortening the longitudinal relaxation time of the tissue. This physical change is related to the temporal signal intensity S(t) and the contrast agent concentration C. data (t) all have definite mathematical relationships, and the steps are as follows:

[0097] For each time point t of the time-series signal intensity S(t), the dynamic longitudinal relaxation time T1(t) corresponding to the current time point is solved by inversely based on the signal equation of magnetic resonance imaging. The signal equation is:

[0098]

[0099] S(t) is the signal strength measured at time t, and (t) represents the longitudinal relaxation time of the organization at time t. Represents a constant related to proton density and device gain; TR represents the sequence repetition time. Indicates the flip angle of the radio frequency pulse; it is a known parameter set during scanning.

[0100] Convert the dynamic longitudinal relaxation time T1(t) corresponding to the current time point into its reciprocal to obtain the dynamic relaxation rate R1(t): ;

[0101] Subtracting the baseline relaxation rate R10 of the tissue before contrast agent injection from the dynamic relaxation rate R1(t) yields the net increase in relaxation rate ΔR1(t) entirely contributed by the contrast agent: Where T10 is the baseline relaxation time of the tissue before contrast agent injection; and the net increase in relaxation rate ΔR1(t) represents the purely physical effect produced by the contrast machine.

[0102] The net increase in relaxation rate ΔR1(t) caused by contrast agent is related to the contrast agent concentration C in the tissue. data (t) shows a linear proportional relationship, and the contrast agent concentration C is obtained after conversion. data (t): in, The longitudinal relaxation rate of the contrast agent is an inherent physical constant of the contrast agent, usually obtained from technical manuals or literature.

[0103] Longitudinal relaxation rate It was configured within a numerical range that matches the physical properties of commonly used clinical gadolinium-based contrast agents (GBCAs);

[0104] The lower limit of the numerical range is set to be no less than 3.0 L·mol⁻¹. -1 ·s -1 The technical objective is to exclude relaxation rate values ​​of atypical or high molecular weight contrast agents, focusing instead on low molecular weight extracellular fluid contrast agents used in routine dynamic contrast enhancement analysis; the upper limit of the numerical range is set to no higher than 6.0 L·mmol. -1 ·s -1 Its technical purpose is to cover the relaxation rate range of most commercially available gadolinium-based contrast agents under the mainstream clinical magnetic field strength of 1.5 Tesla to 3.0 Tesla.

[0105] Obtain the contrast agent concentration C for each voxel within the three-dimensional volume of interest. data (t), and the contrast agent concentration C for each voxel. data (t) Perform voxel-level pharmacokinetic model fitting to determine the plasma transport rate K, which characterizes vascular permeability. trans and the extravascular extracellular space volume fraction V, which characterizes tissue structure e ;

[0106] The voxel-level pharmacokinetic model fitting step is implemented through an iterative optimization algorithm, specifically including:

[0107] Based on prior physiological knowledge, the plasma transport rate K is... trans and extravascular extracellular space volume fraction V e Set initial parameter values Wherein, the plasma transport rate K trans The initial value is set at 0.01 min. -1 up to 0.5 min-1 Within the range, V e The initial value is set in the dimensionless range of 0.05 to 0.6;

[0108] Start an iterative optimization loop, and perform the following operations in each iteration:

[0109] Residual and Jacobian matrix calculation: Based on the parameter value P of the current iteration and the preset pharmacokinetic model, calculate the first residual sum of squares SSR1;

[0110] The pharmacokinetic model was used to predict the contrast agent concentration, and the predicted contrast agent concentration C was calculated. model (t) and the contrast agent concentration C data The differences between (t) at each time point constitute the residual vector r: ;

[0111] Among them, C data (t1) and C data (t n ) refers to the time points t1 and t2. n The contrast agent concentration was actually measured and calculated by processing four-dimensional image data; C model (t1) and C model (t n ) refers to the time points t1 and t2. n Contrast agent concentration predicted by pharmacokinetic model;

[0112] Meanwhile, the computational model predicts the concentration of K trans and V e partial derivatives , forming the Jacobian matrix J; ;

[0113] In this matrix, the first and second columns of the Jacobian matrix J represent the model-predicted concentration pairs with K, respectively. trans and V e The sensitivity of the parameter update step size ΔP is determined by solving the system of linear equations. ,in, Let J denote the transpose of the Jacobian matrix J. This represents a non-negative adaptive damping factor. It is the identity matrix;

[0114] The experimental parameter value Ptrial = P + ΔP is generated based on the calculated parameter update step size ΔP, and a second experimental residual sum of squares SSR2 is calculated based on the experimental parameter value Ptrial. The second experimental residual sum of squares SSR2 is compared with the first residual sum of squares SSR1 to evaluate the effectiveness of the parameter update step size ΔP, including:

[0115] If SSR2 is less than SSR1, the update step size ΔP is determined to be valid, and the current parameter value P is updated to the experimental parameter value Ptrial, while the damping factor λ is decreased; conversely, if SSR2 is greater than or equal to SSR1, the parameter update step size ΔP is determined to be invalid, the current parameter value P is kept unchanged, and the damping factor λ is increased.

[0116] A first threshold and a second threshold are set. When the norm of the parameter update step size ΔP is less than the first threshold, or the relative change of the residual sum of squares is less than the second threshold, or the number of iterations reaches the maximum value; if the conditions are not met, the next iteration continues; if the conditions are met, the iteration is terminated and the final parameter value P is used as the fitting result of the current voxel.

[0117] Extract the final parameter value P after each termination iteration, and extract the plasma transport rate K for each voxel within parameter value P. trans and extravascular extracellular space volume fraction V e Arrange each voxel on its original three-dimensional spatial coordinates (x, y, z) to generate K. trans 3D parametric map and V e Three-dimensional parametric map;

[0118] Extract K trans K of all voxels in the 3D parametric map trans The arithmetic mean of the values ​​is used to obtain the average blood supply activity index;

[0119] Extract V e V of all voxels in the 3D parametric map e The arithmetic mean of the values ​​is used to obtain the average extracellular space volume fraction.

[0120] A preset physiological efficacy threshold is set. If the average blood supply activity index is less than the physiological efficacy threshold, it means that the drug inhibits the vascular permeability and blood flow perfusion of the tumor within a given cycle, achieving the expected anti-angiogenic or vascular normalization effect, and generating a first-action efficacy label.

[0121] If the average blood supply activity index is greater than or equal to the physiological effectiveness threshold, it means that the drug failed to effectively inhibit the blood supply activity of the tumor within a given cycle, and the effect was unqualified, generating a secondary action invalid label.

[0122] The core of plasma transport rate is measuring tumor vascular permeability and blood perfusion. In most solid tumors, neovascularization exhibits high permeability due to structural abnormalities, disordered arrangement, and porous walls, and is termed "leaky" vessels. This chaotic vascular system is the basis for rapid tumor growth. Therefore, an untreated, highly active tumor with high permeability (Kk) will... transThe value will be high. Mechanism of action of anti-angiogenic drugs: The core objective of these drugs evaluated in this technical approach is to attack the tumor's vascular system. Their mechanisms of action are mainly twofold:

[0123] Anti-angiogenesis: Inhibits the formation of new blood vessels, causing tumors to "starve".

[0124] Vascular normalization: Repairing leaky blood vessels, making their structure and function more like those of normal tissue vessels, thereby improving the delivery of subsequent chemotherapy drugs; regardless of the mechanism mentioned above, effective drug action will inevitably lead to reduced permeability and blood perfusion in the tumor vascular system. This is characterized by the following quantitative indicators:

[0125] If K after treatment trans Below this threshold, it means that the drug has successfully suppressed tumor angiogenesis activity from a high, pathological level to a low, inhibited level. This directly proves that the drug is effective.

[0126] Conversely, if K trans If the value is still higher than or equal to the threshold, it indicates that the drug has failed to effectively alter the vascular microenvironment of the tumor, and its effect is ineffective or insignificant.

[0127] Physiological efficacy threshold: set at 0.15 min⁻¹ based on extensive research; below this value, the drug is considered to have significantly inhibited vasoactivity biologically. Examples of tumor samples are shown in Table 1 below:

[0128] Table 1: Examples of tumor samples

[0129] Tumor sample number Total number of voxels of interest N Mean blood supply activity index Mean extracellular space volume Physiological effectiveness threshold (preferred value) Evaluation results Sample 001 31452 0.08 0.45 0.15 First function effective label Sample 002 45870 0.25 0.18 0.15 First-function invalid label Sample 003 28910 0.16 0.28 0.15 Secondary function invalid label Sample 004 (Pre-treatment baseline) 30500 0.35 0.22 0.15 (not applicable) Sample 005 38200 0.11 0.21 0.15 First function effective label Sample 006 25550 0.14 0.32 0.15 First function effective label

[0130] Among them, the average blood supply activity index of sample 001 was significantly lower than the threshold, indicating that vascular permeability was effectively inhibited. At the same time, the average extracellular space volume fraction was high, which may suggest that the treatment caused partial cell necrosis and loosened the tissue structure.

[0131] The mean blood supply activity index of sample 002 was much higher than the threshold, indicating that the tumor blood supply remained active. At the same time, the mean extravascular extracellular space volume fraction was low, suggesting that the tumor cells were still highly dense and the drug failed to achieve the expected effect.

[0132] Sample 003 had an average blood supply activity index slightly above the threshold, and was therefore deemed invalid according to the judgment rules. This demonstrates the objectivity and rigor of the method. The average extravascular extracellular space volume was at a moderate level.

[0133] Sample 004 is a typical pre-treatment sample with a very high mean blood supply activity index and a low V. eThe value, characterized by high vascular permeability and high cell density, represents an aggressive feature that provides a benchmark for subsequent efficacy evaluation.

[0134] Sample 005 had a mean blood supply activity index that was clearly below the threshold, thus it was deemed effective. Unlike sample 001, this sample had a lower mean extravascular extracellular space volume fraction, suggesting that the drug's main mechanism of action might be "vascular normalization," i.e., repairing vascular leakage, reducing tissue edema, and thus making the tissue structure more compact.

[0135] Sample 006 had an average blood supply activity index just below the threshold, classifying it as "borderline effective." This method can accurately capture this weak but significant response. Simultaneously, the average extravascular extracellular space volume fraction showed a moderate increase, suggesting a certain degree of cell-killing effect, but not as pronounced as in sample 001.

[0136] and K trans 3D parametric map and V e The three-dimensional parametric maps are fused to form a tumor blood supply activity map.

[0137] In this embodiment, existing technologies rely on clinicians to mentally integrate separated images, which this invention replaces with a complete automated processing workflow. From bias field correction using the N4ITK algorithm and accurate registration of multimodal images using non-rigid algorithms, to three-dimensional tumor segmentation using deep learning models, each step ensures standardized data processing. More importantly, by linking the average post-treatment blood supply activity index with a defined physiological effectiveness threshold (e.g., 0.15 min...), the process is streamlined. -1 By making a rigid comparison, this invention transforms the ambiguous question of "whether the therapeutic effect is significant" into a binary, objective calculation result (such as the comparison between 0.16 and 0.15 in sample 003), which helps to eliminate subjective differences in the evaluation process and ensures a high degree of consistency and reliability of the evaluation results.

[0138] Traditional methods only calculate the mean parameters of the entire tumor, while this invention independently performs iterative optimization-based pharmacokinetic model fitting, including the Levenberg-Marquardt algorithm, on each voxel within the three-dimensional volume of interest. This in-depth analysis at the microscopic level not only generates high-resolution K... trans and V e Three-dimensional parametric mapping, more importantly, can provide profound insights into the mechanism of drug action through the combined changes of these two parameters. As shown in the example, sample 001 (low K) trans High V e ) and sample 005 (low K) trans Low V eAlthough both were deemed "effective," this method can further reveal two distinct biological effects: "anti-angiogenic cell necrosis" and "vascular normalization." This ability to differentiate between them is completely lacking in existing technologies, providing greater depth of information for understanding the mechanisms of personalized drug action.

[0139] This invention describes in detail how, from the original magnetic resonance time-series signal intensity S(t), through signal equations, relaxation rate transformation, and then utilizing the intrinsic physical constant A1 of the contrast agent, the contrast agent concentration C is finally calculated accurately. data The entire process of (t) is described. This rigorous link ensures that the input data for subsequent pharmacokinetic models has a solid physical and physiological basis, thus enabling the final output K to be... trans and V e The values ​​accurately reflect the permeability of tumor blood vessels and the characteristics of tissue structure. This first-principles-based calculation method, compared to empirical analysis based solely on image grayscale changes, yields more robust and biologically interpretable results, providing stronger data support for clinical decision-making.

[0140] Example 3

[0141] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 2 Specifically, based on the four-dimensional image dataset, within the three-dimensional volume of interest, for each time point of the three-dimensional image data, several predefined radiomics features are extracted to characterize the internal texture and morphology of the tumor.

[0142] For each of the several types of image omics features, a time series consisting of its feature values ​​at all time points is constructed to obtain several types of time feature sequences;

[0143] For the constructed time feature series, its sample entropy value is calculated to quantify the irregularity of the evolution of the corresponding radiomics features over time; and all the sample entropy values ​​calculated from all radiomics features are combined in a weighted linear combination to generate the temporal heterogeneity index (THI).

[0144] A drug resistance risk threshold is preset, and the temporal heterogeneity index (THI) is compared with the drug resistance risk threshold;

[0145] If the temporal heterogeneity index (THI) is greater than the drug resistance risk threshold, the tumor's internal structure is determined to be unstable, with a potential risk of developing heterogeneous drug resistance, and a third high-risk drug resistance label is generated; otherwise, a fourth low-risk drug resistance label is generated.

[0146] Radiomics features include first-order statistical features and second-order statistical features;

[0147] First-order statistical features include: the sum of squares of voxel intensity values, the entropy value (a measure of the disorder of voxel intensity distribution), the skewness of gray-level distribution, and the kurtosis of gray-level distribution;

[0148] Second-order statistical features include: texture roughness, morphological sphericity, and maximum three-dimensional diameter.

[0149] In this embodiment, efficacy assessment is elevated from describing static features to quantifying and predicting the complexity of dynamic evolution. By constructing time series for various radiomics features and innovatively introducing sample entropy to measure the irregularity of their evolutionary process, this invention can capture the dynamic instability of tumor internal structures that traditional methods cannot identify. The resulting temporal heterogeneity index, denoted as THI, condenses this complex dynamic process into an objective and quantifiable risk indicator. After comparison with a preset threshold, tumors with high drug resistance potential can be prospectively identified. This directly overcomes the fundamental deficiency of existing technologies, which can only perform retrospective assessments, providing crucial and quantifiable decision-making basis for predicting and adjusting treatment strategies before clinical drug resistance manifests, thereby achieving proactive management of tumor evolution risk.

[0150] Example 4

[0151] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 3 Specifically, the tumor blood supply activity map generated in step two is obtained. Using the tumor blood supply activity map as a spatial condition and the time series of the temporal heterogeneity index (THI) mentioned in step three as temporal evolution driving data, they are jointly input into a pre-trained conditional image generation network. Forward inference is performed through the conditional image generation network to predict and synthesize a tumor heterogeneity evolution prediction map at a preset future assessment time point Tn+1. The prediction map is used to characterize the spatial distribution of the tumor internal region evolving into a highly heterogeneous state at the future assessment time point.

[0152] The conditional generative adversarial network includes: a generator for receiving the spatial constraints and temporal evolution drivers and generating the prediction graph, and a discriminator for determining the difference between the generated prediction graph and the real evolution graph;

[0153] After obtaining the true evolution map at the future assessment time point Tn+1 and extracting the true tumor heterogeneity feature map based on the true evolution map, the similarity is measured with the tumor heterogeneity evolution prediction map at the future assessment time point Tn+1 to obtain the fidelity index GFI.

[0154] A preset fidelity threshold is set, preferably 0.85. When the fidelity index (GFI) is lower than the fidelity threshold, it is considered unqualified, and the tumor heterogeneity evolution prediction map at the current future evaluation time point Tn+1 is removed. When the fidelity index (GFI) is greater than or equal to the fidelity threshold, it is considered qualified, and the image is retained.

[0155] In this embodiment, the paradigm shift from "describing the past" to "predicting the future" brings a forward-looking perspective to tumor treatment efficacy evaluation. By introducing a conditional image generation network, this invention overcomes the fundamental limitation of existing technologies, which can only perform retrospective analyses.

[0156] More specifically, this method is not a simple trend extrapolation, but rather achieves precise spatial localization of future high-heterogeneity regions by using a "tumor blood supply activity atlas" representing the physiological basis as a spatial constraint and a "temporal heterogeneity index" reflecting evolutionary dynamics as a driving force. This enables clinicians to, for the first time, visually and spatially anticipate high-risk subregions that may develop drug resistance, providing a basis for decision-making regarding preventative interventions or adjustments to treatment plans, and promoting a shift in clinical management from passive response to proactive anticipation.

[0157] Example 5

[0158] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 4 Specifically, using the anatomical structure image as the base rendering layer, the tumor blood supply activity map and the tumor heterogeneity evolution prediction map at the future assessment time point Tn+1 are fused in its three-dimensional coordinate space to generate a fused view that can simultaneously display the tumor anatomical location, current blood supply status and future heterogeneity evolution trend.

[0159] Based on the blood supply activity index and the temporal heterogeneity index, a weighted summation is performed to obtain the drug efficacy index, and a first drug efficacy threshold and a second drug efficacy threshold are preset. Figure 4 Threshold 1 is the first drug efficacy threshold; threshold 2 is the second drug efficacy threshold.

[0160] When the drug efficacy index is less than the second drug efficacy threshold, the first drug efficacy level is generated.

[0161] When the drug efficacy index is between the first drug efficacy threshold and the second drug efficacy threshold, a second drug efficacy level is generated.

[0162] When the drug efficacy index is greater than the first drug efficacy threshold, a third drug efficacy level is generated.

[0163] The third-level drug efficacy is better than the second-level drug efficacy, and the second-level drug efficacy is better than the first-level drug efficacy.

[0164] A critical roughness threshold segmentation operation is performed on the tumor heterogeneity evolution prediction map at the future evaluation time point Tn+1. The predicted texture roughness value of each voxel is extracted, a critical roughness threshold is preset, and voxels with texture roughness values ​​higher than or equal to the critical roughness threshold are aggregated to automatically generate targeted high-risk evolution regions.

[0165] The critical roughness threshold is derived from: preparing a calibration dataset containing multiple groups of cases (e.g., N≥80 cases), which should cover typical clinical scenarios with different tumor types, treatment regimens, and baseline heterogeneity levels; for each case in this dataset, it includes: a three-dimensional tumor heterogeneity evolution prediction map generated by the method of this invention at time point Tn for a future assessment time point Tn+1; and a clinical outcome label (e.g., labeled as "disease progression" or "disease stabilization / remission") obtained through actual clinical follow-up of the patient as the "gold standard"; assuming that the texture roughness values ​​in the prediction map have been normalized to the [0,1] interval, the search range of the critical roughness threshold is set to [0.3,0.9], with a search step size of 0.01; for each candidate critical roughness threshold within the search range:

[0166] Thresholding is performed on the evolutionary prediction maps of all cases in the calibration dataset using the current candidate values. If any case's prediction map contains a voxel with a roughness higher than a certain value, the prediction result for that case is classified as "high risk".

[0167] The predicted outcomes ("high risk" or "non-high risk") of all cases are compared with the corresponding "gold standard" clinical outcome labels to calculate a comprehensive predictive performance index for comprehensively evaluating predictive accuracy; the preferred value of the critical roughness threshold is determined to be the candidate value that maximizes the comprehensive predictive performance index across the entire calibration dataset.

[0168] The calibration results based on large-scale clinical data show that the preferred value typically falls within the range of [0.65, 0.80]. A joint calibration experiment was conducted. This experiment first prepared a calibration dataset containing multiple groups of cases (e.g., N≥50 cases), designed to cover typical clinical scenarios with different tumor types, treatment regimens, and baseline heterogeneity levels. For each case in the dataset, a three-dimensional tumor heterogeneity evolution prediction map for a future assessment time point Tn+1, generated by the method of this invention, and a clinical outcome label obtained through actual clinical follow-up as the "standard" (e.g., labeled "disease progression" or "disease stability / remission"). With the texture roughness values ​​in the evolution prediction map normalized to the range of [0, 1], the search range for the critical roughness threshold was set to [0.3, 0.9], with a search step size of 0.01.

[0169] During calibration, the system iterates through every candidate critical roughness threshold within the search range. For each candidate value, the system uses it to perform threshold segmentation on the evolution prediction graphs of all cases in the calibration dataset. If any case's prediction graph contains a voxel with a roughness higher than the current candidate critical roughness threshold, the prediction result for that case is classified as "high risk." Subsequently, the prediction results of all cases are compared with the corresponding "standard" clinical outcome labels, and the accuracy of the current candidate threshold is evaluated by calculating a comprehensive predictive performance index. The preferred comprehensive predictive performance index is the Youden's Index (J), which is calculated as J = Sensitivity + Specificity - 1. The Youden's Index, first proposed by W.J. Youden in 1950, is a fundamental and widely used statistical indicator in fields such as medical diagnostics, biostatistics, and machine learning—it balances the ability to detect "disease progression" cases (sensitivity) and the ability to correctly exclude "disease stability / remission" cases (specificity).

[0170] Ultimately, the preferred value of the critical roughness threshold was determined to be the candidate critical roughness threshold that maximizes the Youden index across the entire calibration dataset. Calibration results based on large-scale clinical data show that this preferred value typically falls within the range of [0.65, 0.80]. In a specific clinical application scenario aimed at maximizing the differentiation between disease progression and stability over the next 12 months, the preferred value of the critical roughness threshold is 0.72.

[0171] In this embodiment, by generating a fused view, a spatiotemporal integrated presentation of multidimensional heterogeneous information is achieved, effectively addressing the problems of information fragmentation and excessive cognitive load. Existing technologies present anatomical, physiological, and risk assessment information separately, while this invention seamlessly overlays the underlying anatomical structure image, the blood supply atlas representing the current physiological state, and the heterogeneity prediction map indicating future evolutionary trends onto the same three-dimensional coordinate space. This provides clinicians with an intuitive and unified decision-making interface, enabling them to clearly grasp the "location," "current state," and "future" of the tumor, thus improving diagnostic efficiency and accuracy. A comprehensive drug efficacy rating is constructed, reducing the complexity of the analysis process to clear conclusions that can guide clinical practice. By weighting and combining the blood supply activity index, which reflects immediate drug action, with the temporal heterogeneity index, which predicts long-term drug resistance risk, this invention outputs not isolated parameters, but a comprehensive assessment rating that integrates "current efficacy" and "future risk." This multi-level classification (such as first, second, and third drug efficacy levels) provides a more refined and standardized reference for clinical decision-making, making efficacy assessments comparable for different patients and treatment cycles. Transforming prospective predictions into specific, actionable clinical goals represents a leap from "risk warning" to "precise targeting." By thresholding the tumor heterogeneity evolution prediction map, this method can automatically and objectively delineate "targeted high-risk evolution regions" most likely to evolve into high-risk states in the future. This provides precise spatial targets for truly prospective and personalized treatment interventions, such as guiding localized intensified treatment or closer monitoring of specific subregions, thereby completely transforming clinical strategies from a passive "response mode" to a proactive "prediction and intervention mode."

[0172] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value, it is acceptable.

[0173] The above formulas are all derived from software simulation using a large amount of data and are selected to be close to the actual values. The coefficients in the formulas are set by those skilled in the art according to the actual situation. The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any equivalent substitutions or changes made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the protection scope of the present invention.

Claims

1. An automated imaging grading and presentation method for the efficacy of tumor drugs, characterized in that, Includes the following steps: Step 1: Obtain anatomical images and dynamic contrast-enhanced images of the patient at baseline time points before drug treatment and at drug cycle time points, and fuse them to obtain a four-dimensional image dataset. Step 2: Based on the curve composed of the temporal signal intensity S(t) of the dynamic contrast-enhanced images in the four-dimensional image dataset, a voxel-level pharmacokinetic model is constructed to generate a tumor blood supply activity map. After iterative optimization of the pharmacokinetic model based on plasma transport rate and extracellular space volume fraction, the average blood supply activity index of the entire tumor volume of interest is calculated. A physiological efficacy threshold is preset to preliminarily determine the immediate efficacy of the drug on the tumor vascular system. Step 3: Based on the four-dimensional image dataset, extract radiomics features through image processing technology, and construct a temporal heterogeneity index, denoted as THI, to characterize the complexity of the evolution of the internal structure of the tumor over time. Preset a drug resistance risk threshold to assess the potential risk of tumors developing heterogeneous drug resistance; Step 4: Using the tumor blood supply activity map as a spatial condition and the time series of the temporal heterogeneity index (THI) as an evolutionary driver, input the data into a pre-trained conditional image generation network to synthesize a tumor heterogeneity evolution prediction map at a future evaluation time point Tn+1. Simultaneously, by calculating the structural similarity between the tumor heterogeneity evolution prediction map and the actual evolution map, a generation fidelity index, denoted as GFI, is obtained. If the GFI is qualified, image retention processing is performed; otherwise, image rejection processing is performed. Step 5: Using the anatomical structure image as the base layer, fuse and overlay the tumor blood supply activity map and the preserved tumor heterogeneity evolution prediction map to generate a fused view; and combine the average blood supply activity index and the temporal heterogeneity index to automatically output a comprehensive drug efficacy level.

2. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 1, characterized in that, At the baseline time point and the drug cycle time point, a first magnetic resonance imaging scan is performed on the tumor region to acquire anatomical structure images; after bias field correction of the anatomical structure images by the N4ITK algorithm, a first reference image set is obtained. At the baseline time point and the drug cycle time point, a second dynamic contrast-enhanced magnetic resonance imaging scan was performed on the same tumor region. By intravenously injecting contrast agent and performing continuous rapid imaging, second time-series image data reflecting the perfusion and penetration process of contrast agent in the tumor were acquired. Using the first reference image set at the baseline time point as a reference benchmark, a non-rigid image registration algorithm is employed to perform 3D spatial correction on the second time-series image data at the drug cycle time point. Then, within the aligned first reference image set at the baseline time point, a deep learning segmentation model is used to identify and obtain the 3D volume of interest defining the spatial extent of the tumor. Specifically, this includes: The first reference image set of the aligned baseline time points is used as the three-dimensional voxel input data and fed into the pre-trained three-dimensional convolutional neural network segmentation model to perform forward propagation calculation. The model performs hierarchical feature extraction and semantic classification on each voxel in the three-dimensional voxel input data, thereby generating a three-dimensional probability map. Threshold segmentation is performed on the three-dimensional probability map, and the set of voxels whose probability values ​​are not less than a preset segmentation threshold τ is defined as the three-dimensional volume of interest. Based on the three-dimensional volume of interest, the signal intensity value of each voxel within the tumor region is extracted from all aligned second temporal image data, normalized, and a four-dimensional image dataset with a unified coordinate system is constructed.

3. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 1, characterized in that, In the four-dimensional image dataset, the signal intensity values ​​of each voxel at all dynamic contrast-enhanced scanning time points are extracted to form a curve composed of time-series signal intensity S(t). Then, using the relationship between magnetic resonance signal intensity and contrast agent concentration, the time-series signal intensity S(t) is converted into contrast agent concentration C. data (t); The conversion steps are as follows: Contrast agent molecules can alter the magnetic field environment of the surrounding water molecules, thereby shortening the longitudinal relaxation time of the tissue. This physical change is related to the temporal signal intensity S(t) and the contrast agent concentration C. data (t) all have definite mathematical relationships, and the steps are as follows: For each time point t on the temporal signal intensity S(t), the dynamic longitudinal relaxation time T1(t) corresponding to the current time point is solved by inversely based on the signal equation of magnetic resonance imaging. Convert the dynamic longitudinal relaxation time T1(t) corresponding to the current time point into its reciprocal to obtain the dynamic relaxation rate R1(t); Subtracting the baseline relaxation rate R10 of the tissue before contrast agent injection from the dynamic relaxation rate R1(t) yields the net increase in relaxation rate ΔR1(t) that is entirely contributed by the contrast agent. The net increase in relaxation rate ΔR1(t) caused by contrast agent is related to the contrast agent concentration C in the tissue. data (t) shows a linear proportional relationship, and the contrast agent concentration C is obtained after conversion. data (t).

4. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 1, characterized in that, Obtain the contrast agent concentration C for each voxel within the three-dimensional volume of interest. data (t), and the contrast agent concentration C for each voxel. data (t) Perform voxel-level pharmacokinetic model fitting to determine the plasma transport rate K, which characterizes vascular permeability. trans and the extravascular extracellular space volume fraction V, which characterizes tissue structure e ; The voxel-level pharmacokinetic model fitting process is implemented through an iterative optimization algorithm, specifically including: Based on prior physiological knowledge, the plasma transport rate K... trans and extravascular extracellular space volume fraction V e Set initial parameter values ; Start an iterative optimization loop, and perform the following operations in each iteration: Residual and Jacobian matrix calculation: Based on the parameter value P of the current iteration and the preset pharmacokinetic model, calculate the first residual sum of squares SSR1; The pharmacokinetic model was used to predict the contrast agent concentration, and the predicted contrast agent concentration C was calculated. model (t) and the contrast agent concentration C data The differences between (t) at each time point constitute the residual vector r: Simultaneously, the predicted concentration of K was calculated using a voxel-level pharmacokinetic model. trans and V e partial derivatives , forming the Jacobian matrix J; And solve the system of linear equations to obtain the parameter update step size ΔP; The experimental parameter value Ptrial = P + ΔP is generated based on the calculated parameter update step size ΔP, and a second experimental residual sum of squares SSR2 is calculated based on the experimental parameter value Ptrial. The second experimental residual sum of squares SSR2 is compared with the first residual sum of squares SSR1 to evaluate the effectiveness of the parameter update step size ΔP, including: To ensure the stable convergence of the iterative optimization loop and prevent optimization divergence caused by excessive parameter update step size ΔP due to the high nonlinearity of the pharmacokinetic model, a non-negative adaptive damping factor λ is introduced to dynamically adjust between the fast-converging Gauss-Newton method and the robust gradient descent method. If SSR2 is less than SSR1, the update step size ΔP is determined to be valid, and the current parameter value P is updated to the experimental parameter value Ptrial, while the damping factor λ is decreased; conversely, if SSR2 is greater than or equal to SSR1, the parameter update step size ΔP is determined to be invalid, the current parameter value P is kept unchanged, and the damping factor λ is increased. A first threshold and a second threshold are set. If the norm of the parameter update step size ΔP is less than the first threshold, or the relative change of the residual sum of squares is less than the second threshold, or the number of iterations reaches the maximum value, the iteration is terminated and the final parameter value P is used as the fitting result of the current voxel; if the conditions are not met, the next iteration continues.

5. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 4, characterized in that, Extract the final parameter value P after each termination iteration, and extract the plasma transport rate K for each voxel within parameter value P. trans and extravascular extracellular space volume fraction V e Arrange each voxel on its original three-dimensional spatial coordinates (x, y, z) to generate K. trans 3D parametric map and V e Three-dimensional parametric map; Extract K trans K of all voxels in the 3D parametric map trans The arithmetic mean of the values ​​is used to obtain the average blood supply activity index; Extract V e V of all voxels in the 3D parametric map e The arithmetic mean of the values ​​is used to obtain the average extravascular extracellular space volume fraction. A preset physiological efficacy threshold is set. If the average blood supply activity index is less than the physiological efficacy threshold, it means that the drug inhibits the vascular permeability and blood flow perfusion of the tumor within a given cycle, achieving the expected anti-angiogenic or vascular normalization effect, and generating a first-action efficacy label. If the average blood supply activity index is greater than or equal to the physiological effectiveness threshold, it means that the drug failed to effectively inhibit the blood supply activity of the tumor within a given cycle, and the effect was unqualified, generating a secondary action invalid label. and K trans 3D parametric map and V e The three-dimensional parametric maps are fused to form a tumor blood supply activity map.

6. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 1, characterized in that, Based on the four-dimensional image dataset, within the three-dimensional volume of interest, for each time point of the three-dimensional image data, several radiomics features of tumor internal texture and morphology are extracted. For each of the several types of image omics features, a time series consisting of its feature values ​​at all time points is constructed to obtain several types of time feature sequences; For the constructed time feature series, its sample entropy value is calculated to quantify the irregularity of the evolution of the corresponding radiomics features over time; and all the sample entropy values ​​calculated from all radiomics features are combined in a weighted linear combination to generate the temporal heterogeneity index (THI). A drug resistance risk threshold is preset, and the temporal heterogeneity index (THI) is compared with the drug resistance risk threshold. If the temporal heterogeneity index (THI) is greater than the drug resistance risk threshold, the tumor's internal structure is determined to be unstable, with a potential risk of developing heterogeneous drug resistance, and a third high-risk drug resistance label is generated. Conversely, a fourth low-risk drug resistance label is generated.

7. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 6, characterized in that, Radiomics features include first-order statistical features and second-order statistical features; First-order statistical features include: the sum of squares of voxel intensity values, the entropy value (a measure of the disorder of voxel intensity distribution), the skewness of gray-level distribution, and the kurtosis of gray-level distribution; Second-order statistical features include: texture roughness, morphological sphericity, and maximum three-dimensional diameter.

8. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 1, characterized in that, The tumor blood supply activity map generated in step two is obtained. The tumor blood supply activity map is used as a spatial condition, and the time series of the temporal heterogeneity index THI mentioned in step three is used as the temporal evolution driving data. They are input together into a pre-trained conditional image generation network. Forward reasoning is performed through the conditional image generation network to predict and synthesize the tumor heterogeneity evolution prediction map at the preset future evaluation time point Tn+1. The conditional generative adversarial network includes: a generator that receives spatial conditions and temporal evolution drivers and generates a tumor heterogeneity evolution prediction map, and a discriminator that distinguishes the differences between the generated tumor heterogeneity evolution prediction map and the real evolution map; After obtaining the true evolution map at the future assessment time point Tn+1 and extracting the true tumor heterogeneity feature map based on the true evolution map, the similarity is measured with the tumor heterogeneity evolution prediction map at the future assessment time point Tn+1 to obtain the fidelity index GFI. A preset fidelity threshold is set. If the fidelity index (GFI) is lower than the fidelity threshold, the image is considered unqualified and the tumor heterogeneity evolution prediction map at the current future evaluation time point Tn+1 is removed. If the fidelity index (GFI) is greater than or equal to the fidelity threshold, the image is considered qualified and is retained.

9. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 1, characterized in that, Using the anatomical structure image as the base rendering layer, the tumor blood supply activity map and the tumor heterogeneity evolution prediction map at the future assessment time point Tn+1 are fused in its three-dimensional coordinate space to generate a fused view that can simultaneously display the tumor anatomical location, current blood supply status and future heterogeneity evolution trend. Based on the blood supply activity index and the temporal heterogeneity index, the drug efficacy index is obtained by a weighted summation method, and a first drug efficacy threshold and a second drug efficacy threshold are preset. When the drug efficacy index is less than the second drug efficacy threshold, the first drug efficacy level is generated. When the drug efficacy index is between the first drug efficacy threshold and the second drug efficacy threshold, a second drug efficacy level is generated. When the drug efficacy index is greater than the first drug efficacy threshold, a third drug efficacy level is generated.

10. The method for automatic imaging grading and presentation of tumor drug efficacy according to claim 9, characterized in that, A critical roughness threshold segmentation operation is performed on the tumor heterogeneity evolution prediction map at the future evaluation time point Tn+1. The predicted texture roughness value of each voxel is extracted, a critical roughness threshold is preset, and voxels with texture roughness values ​​higher than or equal to the critical roughness threshold are aggregated to automatically generate targeted high-risk evolution regions.

Images