A method for predicting lymph node metastasis of bladder cancer

Abstract

This invention discloses a method for predicting lymph node metastasis in bladder cancer, belonging to the field of bladder tumor image processing technology. This method acquires arterial and venous phase CT volume data and initial mask volume data of the target region. It then uses multi-scale three-dimensional morphological erosion operations to generate a set of sequentially nested shell regions, and calculates the activity bias value based on the normalized distance from each shell region to its boundary. A three-dimensional activity confidence tensor is generated by mapping the local temporal variance of voxels to the activity bias value, and this tensor is used as a spatial gating weight to perform voxel-by-voxel multiplication on the initial spatiotemporal gradient tensor, thereby extracting the enhanced spatiotemporal gradient tensor. Finally, the volume percentage of the retained voxels is statistically analyzed and input into a classifier, outputting a prediction auxiliary report. This scheme effectively solves the feature dilution problem caused by necrosis noise within complex tumors through spatial topological decoupling and spatiotemporal feature gating, achieving high-precision non-invasive prediction of preoperative lymph node metastasis risk.

Description

Technical Field

[0001] This invention relates to the field of bladder tumor image processing technology, and more specifically, this application relates to a method for predicting lymph node metastasis in bladder cancer. Background Technology

[0002] In the field of medical image processing, preoperative prediction of lymph node metastasis in bladder cancer based on enhanced CT images is an important step in assisting clinical diagnosis. Traditional image-aided processing techniques typically treat the primary tumor as a homogeneous entity, extracting macroscopic image features within a global tumor mask based on a single-phase static scan image and then fusing and reducing the dimensions.

[0003] However, this conventional extraction method ignores the complex spatial topological heterogeneity within the tumor, resulting in low feature extraction accuracy of traditional models when faced with physiological features containing large areas of irregular necrosis, which easily leads to missed diagnoses.

[0004] Specifically, in applications involving massive tumors with extensive irregular necrosis and delayed contrast agent diffusion, a large, ischemic necrotic area with no blood supply typically exists at the tumor center. Existing technologies, when processing such data, forcibly apply globally averaged feature extraction logic. This results in low-frequency, stable background noise from the necrotic area at the tumor center severely diluting and masking the extremely weak effective signals from the invasive regions at the tumor margins during computation. Simultaneously, due to the lack of dynamic feature mining in the temporal dimension, existing technologies completely lose the dynamic retention characteristics of the contrast agent by the peripheral microvascular network. This dilution of information in the spatial dimension and the lack of features in the temporal dimension prevent conventional methods from mapping the true microscopic invasive dynamics of the tumor at its source.

[0005] In summary, the systemic technical deficiency in the imaging scenario of large tumors, where traditional extraction methods cannot shield low-frequency noise generated by necrotic areas, leading to the dilution or even obliteration of the dynamic spatiotemporal invasion characteristics of high-risk edges, is the key to improving prediction accuracy. Summary of the Invention

[0006] To address the aforementioned technical problems, a method for predicting lymph node metastasis in bladder cancer is provided. This technical solution resolves the issues raised in the background section.

[0007] In a first aspect, embodiments of this application provide a method for predicting lymph node metastasis in bladder cancer, comprising the following steps: acquiring arterial phase CT volume data of a target region within a preset three-dimensional spatial coordinate system, including the first CT intensity value of each voxel, venous phase CT volume data of a target region, including the second CT intensity value of each voxel, and initial mask volume data characterizing the three-dimensional contour of the bladder tumor; performing multi-scale three-dimensional morphological erosion operations on the initial mask volume data, progressing from the centroid of the bladder tumor to the boundary, to obtain a set of sequentially nested shell regions, and calculating the corresponding activity bias value for each shell region based on the three-dimensional Euclidean normalized distance from the boundary of the initial mask volume data; for each target voxel located within the initial mask volume data, calculating the local time of its first CT intensity value and second CT intensity value. The variance is calculated, and a weighted sum is obtained based on the activity bias value of its corresponding shell region. The weighted sum is then input into a preset activation function and mapped to the activity confidence value of the voxel. The activity confidence values ​​of all target voxels are matrix-concatenated according to a preset three-dimensional spatial coordinate system to construct a three-dimensional activity confidence tensor for the target voxels. The voxel-by-voxel difference between the venous phase CT volume data and the arterial phase CT volume data is calculated to generate an initial spatiotemporal gradient tensor. The three-dimensional activity confidence tensor and the initial spatiotemporal gradient tensor are multiplied voxel-by-voxel to obtain an enhanced spatiotemporal gradient tensor. Voxels whose enhanced spatiotemporal gradient tensor is greater than a preset signal threshold are identified as cut-off voxels. Their volume proportion in the initial mask volume data is calculated and input into a pre-trained classifier to output an auxiliary report containing lymph node metastasis prediction values.

[0008] Secondly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting lymph node metastasis in bladder cancer.

[0009] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0010] 1. To address the common problem of internal necrosis in massive tumors, this approach generates "sequentially nested shell regions" and assigns differentiated "activity bias values," achieving data separation between tumor invasion boundaries and inert centers. Compared to traditional techniques that directly extract global features, this method prevents low-frequency noise in necrotic areas from diluting key metastatic signals, significantly improving the signal-to-noise ratio of effective information in the "enhanced spatiotemporal gradient tensor," and eliminating the feature masking effect caused by tumor heterogeneity at the signal source.

[0011] 2. By utilizing the "local temporal variance" calculated from "arterial phase CT volume data" and "venous phase CT volume data," the dynamic blood perfusion heterogeneity of contrast agent at the tumor's active edge was accurately captured. Combined with gating operations performed using the "three-dimensional activity confidence tensor," this ensured that the "enhanced spatiotemporal gradient tensor" could accurately map the dynamic trapping effect of the microvascular network. This overcomes the bottleneck of existing technologies that cannot reconstruct invasion dynamics based solely on single-phase images, and compensates for the predictive limitations of static anatomical representations through the spatiotemporal dynamic dimension.

[0012] 3. This scheme abandons the black-box end-to-end prediction approach and uses "volume percentage" as a key bridging indicator to directly link the microscopic distribution of "retained voxels" with the macroscopic risk of metastasis. This first-principles-based calculation process gives the output "lymph node metastasis prediction value" a clear biological and physical basis, meeting the stringent requirements of clinical diagnosis for algorithm interpretability and establishing a highly interpretable physical-logical prediction link. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating the steps of a method for predicting lymph node metastasis in bladder cancer, provided in an embodiment of this application.

[0014] Figure 2 This is a schematic diagram of the logic flow of a method for predicting lymph node metastasis in bladder cancer, provided in an embodiment of this application. Detailed Implementation

[0015] This application provides a method for predicting lymph node metastasis in bladder cancer, which solves the technical problems in the prior art where the low-frequency noise generated by tumor center necrosis severely masks weak invasion signals at the edge due to the use of global mean extraction logic, and the bottleneck in prediction accuracy caused by the inability to capture dynamic contrast-enhanced features.

[0016] This solution addresses the feature dilution phenomenon caused by tumor heterogeneity in medical imaging by constructing a underlying logic from spatial topological decoupling to spatiotemporal feature gating. First, based on the biological distribution characteristics of tumors, this solution reconstructs static initial mask data into a shell structure progressing from the center to the boundary through multi-scale morphological operations. The logical starting point of this step lies in recognizing the inherent differences in cell activity at different depths of the tumor. By establishing sequentially nested shell regions, a physical spatial benchmark can be provided for subsequent differentiated signal weight allocation. Subsequently, an activity bias value is introduced to quantify each shell, a design that breaks the traditional assumption of treating the tumor as a homogeneous entity. In the temporal dimension, this solution calculates the local temporal variance by comparing the intensity fluctuations during the arterial and venous phases to capture the hemodynamic characteristics of the contrast agent at the voxel level. The logic is further deepened by nonlinearly fusing this temporal activity representation with the aforementioned spatial topological bias, mapping it to an activity confidence value. The constructed three-dimensional activity confidence tensor is no longer a simple grayscale distribution but includes multidimensional weights of spatial location and metabolic intensity.

[0017] To extract pure invasion features at the physical level, this scheme utilizes a three-dimensional activity confidence tensor and an initial spatiotemporal gradient tensor generated from the arteriovenous phase difference to perform gated inhibition. Only voxels possessing both anatomically active localization and significant hemodynamic gradients are assigned high weights. In this way, noise from the central necrosis zone, which would otherwise interfere with prediction results, is effectively shielded, thereby accurately identifying cutoff voxels strongly correlated with metastasis risk within the enhanced spatiotemporal gradient tensor. Finally, through statistical analysis and classification of volume proportions, a logical closed loop from underlying physical signals to high-level risk prediction is achieved.

[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0019] like Figure 1 The diagram shown is a step-by-step illustration of a method for predicting lymph node metastasis in bladder cancer provided in an embodiment of this application. The method for predicting lymph node metastasis in bladder cancer includes the following steps: acquiring arterial phase CT volume data of a target region within a preset three-dimensional spatial coordinate system and the first CT intensity value of each voxel, venous phase CT volume data and the second CT intensity value of each voxel, and initial mask volume data characterizing the three-dimensional contour of the bladder tumor.

[0020] The preset three-dimensional spatial coordinate system is constructed based on metadata from medical digital imaging and communication standards. It is established by parsing the physical pixel spacing and scan slice thickness parameters of the original medical image files to ensure that each voxel has an absolute and uniform physical spatial scale. Before input, arterial phase CT volume data and venous phase CT volume data must undergo rigid image registration processing with a unified coordinate system to ensure geometric positional correspondence.

[0021] A multi-scale three-dimensional morphological erosion operation is performed on the initial mask data, progressing from the centroid representing the bladder tumor to the boundary, to obtain a set of nested shell regions. Based on the three-dimensional Euclidean normalized distance from each shell region to the boundary of the initial mask data, the corresponding active bias value is calculated for each shell region.

[0022] Multi-scale 3D morphological erosion is a mature mathematical morphological processing technique in the field of computer vision. It utilizes specific structuring elements to perform boundary shrinkage operations on a set of voxels in 3D space, thereby exploring the internal topological geometry of an object and stripping away non-core regions. The active bias value imparts a probabilistic meaning to spatial depth.

[0023] For each target voxel located within the initial mask data, calculate the local temporal variance of its first CT intensity value and second CT intensity value, and perform a weighted summation operation with the activity bias value of its shell region accordingly. The weighted summation result is then input into a preset activation function and mapped to the activity confidence value of the target voxel.

[0024] Local temporal variance accurately characterizes the dramatic hemodynamic fluctuations of a single target voxel before and after contrast agent injection. Weighted summation integrates the temporal-dimensional fluctuations with the spatial-dimensional depth quantities to generate a comprehensive evaluation index.

[0025] The activity confidence values ​​of all target voxels are matrix-stitched according to a preset three-dimensional spatial coordinate system to construct the three-dimensional activity confidence tensor of the target voxels;

[0026] The data dimensions of the three-dimensional activity confidence tensor are absolutely consistent with the initial mask volume data. The value of each coordinate point inside the tensor intuitively represents the probability of tumor pathological activity at the corresponding spatial location, forming a physical spatial mask.

[0027] Calculate the voxel-by-voxel difference between venous phase CT volume data and arterial phase CT volume data to generate the initial spatiotemporal gradient tensor;

[0028] The initial spatiotemporal gradient tensor cancels out the background grayscale of the human body's baseline tissue through pure physical difference operations, directly reflecting the total amount of contrast agent retained in the microvascular network within the tissue.

[0029] Perform voxel-wise multiplication on the three-dimensional activity confidence tensor and the initial spatiotemporal gradient tensor to obtain the enhanced spatiotemporal gradient tensor;

[0030] The voxel-by-voxel multiplication operation, also known as Hadamard product processing, uses a three-dimensional activity confidence tensor to perform nonlinear physical shielding on the initial spatiotemporal gradient tensor, forcibly setting the invalid gradient values ​​corresponding to necrotic dead zones and background tissues to zero, thus preserving the pure edge active gradient signal.

[0031] The corresponding voxels whose enhanced spatiotemporal gradient tensor is greater than the preset signal threshold are identified as cut-off voxels. Their volume proportion in the initial mask volume data is calculated and input into the pre-trained classifier, and an auxiliary report containing lymph node metastasis prediction values ​​is output.

[0032] The preset signal threshold was obtained by statistically analyzing historical data of real contrast agent permeation residue from a large sample. Its effective range was set to 10 HU to 30 HU, with a typical engineering value of 15 HU, used to determine whether microvessels exhibit abnormal contrast agent retention. The pre-trained classifier was constructed using a standard supervised learning architecture, and its dataset was derived from a retrospective clinical cohort with postoperative pathological lymph node metastasis results that served as the gold standard.

[0033] Figure 2 This is a schematic diagram of the logical flow of a method for predicting lymph node metastasis in bladder cancer, provided in an embodiment of this application. This embodiment addresses the application scenario of massive tumors accompanied by extensive irregular necrosis and delayed contrast agent diffusion, overcoming the inherent defect of traditional medical image processing that treats the tumor as a uniform entity for global feature extraction. It utilizes multi-scale three-dimensional morphological erosion operations and local temporal variance to construct a three-dimensional activity confidence tensor, accurately isolating low-frequency background noise generated by the large, bloodless, ischemic necrotic area. Using the three-dimensional activity confidence tensor as spatial gating weights effectively constrains the extraction range of spatiotemporal gradients, ensuring that the final enhanced spatiotemporal gradient tensor can purely and realistically map the contrast agent dynamic retention characteristics of the newly formed microvascular network at the tumor invasion edge, significantly improving the accuracy and clinical interpretability of preoperative non-invasive lymph node metastasis prediction from the root of the physical signal.

[0034] Furthermore, the internal spatial structure of irregularly shaped, massive tumors is difficult to smoothly segment using conventional Euclidean geometric boundaries. Therefore, conventional overall analysis can lead to assessment bias because it ignores the layered decrease in tumor activity. To non-destructively deconstruct the three-dimensional internal physical topology of any irregularly shaped, massive tumor and construct a spatial analysis benchmark, further processing is performed in the multi-scale three-dimensional morphological erosion operation stage:

[0035] Define a monotonically increasing sequence of morphological erosion radii;

[0036] The monotonically increasing morphological erosion radius sequence is set as an arithmetic sequence of positive integers based on voxel units, with typical engineering values ​​set to voxel lengths of 1, 3, 5, 7, and 9. The maximum radius threshold of the sequence is limited by the maximum inscribed sphere radius of the initial mask volume data to ensure that the final morphological erosion operation does not cause the mask data to completely disappear.

[0037] Using each corrosion radius in the corrosion radius sequence, recursive three-dimensional morphological corrosion operations are performed on the initial mask data to generate multiple nested mask sequences with different shrinkage degrees.

[0038] The recursive 3D morphological erosion operation dynamically calls the current erosion radius to perform voxel stripping of the initial mask data at progressively deeper levels.

[0039] Following the order of progression from the centroid to the boundary, voxel-level Boolean difference operations are performed on two adjacent nested body masks, and the resulting closed surrounding voxel sets are defined as the corresponding shell regions.

[0040] Voxel-level Boolean difference operations can precisely extract the independent closed three-dimensional space between two masks from a mathematical perspective.

[0041] This embodiment solves the technical problem of the difficulty in segmenting the internal spatial structure of irregularly shaped giant tumors by using recursive three-dimensional morphological erosion operations and voxel-level Boolean difference operations, and realizes highly robust mathematical analysis and stripping of the spatial depth topology of any irregularly shaped tumor from the surface to the interior.

[0042] Furthermore, the physical size limitations of CT scanner detectors lead to blurred tissue boundaries. When performing only global calculations, the outermost voxel of a bladder tumor inevitably contains physical signals from both tumor tissue and normal bladder wall tissue. This physical aliasing generates a large spurious local temporal variance when subtracting two CT scans, easily penetrating variance gating and causing severe false positive predictions. Therefore, edge isolation is necessary.

[0043] By performing a three-dimensional morphological boundary extraction operation on the initial mask data to extract the width of the single voxel, the outermost single-layer voxel set data is isolated.

[0044] The three-dimensional morphological boundary extraction operation of the single-voxel width is achieved by subtracting the data features of the initial mask data from the data features obtained after eroding one voxel size.

[0045] Determine whether the coordinates of the target voxel belong to the outermost single-layer voxel set data;

[0046] If the coordinates of the target voxel are determined to belong to the outermost single-layer voxel set data, the mapping result of its weighted summation value is ignored, and the activity confidence value of the target voxel is forcibly reset to 0.

[0047] This embodiment solves the problem of partial volume aliasing effect of voxels caused by the limited physical resolution of CT scanning equipment by extracting the three-dimensional morphological boundary of the voxel width and using a hard truncation mechanism that forces a reset to a value of 0. This achieves precise removal of physical aliasing artifacts at tissue boundaries and physical purification of underlying dynamic features.

[0048] Furthermore, to eliminate non-rigid registration misalignment artifacts caused by uncontrollable physiological peristalsis or respiratory displacement of the patient during voxel difference calculation, due to the deformation and displacement of non-rigid organs between two CT scans, even with pre-performed elastic registration, microscopic physical misalignments will still remain at the edges. The preceding workflow will mistake this pure positional shift for contrast agent elution gradients, generating high-intensity spurious residual signals. After obtaining the enhancement spatiotemporal gradient tensor, and before extracting the truncated voxel data, spatial orthogonality verification is performed:

[0049] By extracting voxels from the three-dimensional shell region of the initial mask data and applying a three-dimensional spatial differential operator to calculate its gradient direction, the volume surface normal vector of the initial mask data in the preset three-dimensional spatial coordinate system is generated.

[0050] The three-dimensional spatial differential operator adopts the standard three-dimensional Sobel operator in this field. It synthesizes a spatial three-dimensional gradient vector by performing convolution difference calculations in the three orthogonal coordinate axes X, Y, and Z respectively to accurately represent the surface normal vector of the volume.

[0051] For each voxel in the enhanced spatiotemporal gradient tensor, the difference gradient vector in the preset three-dimensional coordinate system is calculated by applying a three-dimensional spatial differential operator.

[0052] Calculate the cosine similarity of the angle between the surface normal vector and the difference gradient vector for the same voxel;

[0053] Cosine similarity provides a nonlinear metric that discards absolute numerical values ​​and only evaluates the consistency of orientation in pure geometric space.

[0054] If the absolute value of the cosine similarity of the included angle is greater than the preset polarization verification threshold, the corresponding voxel is determined to be a physical misalignment artifact offset along the normal direction, and the value of the corresponding voxel in the enhanced spatiotemporal gradient tensor is overwritten with the value 0.

[0055] The preset polarization verification threshold is set based on the criterion in spatial geometry that two vectors with nearly parallel angles are considered to be in a displacement state. The legal value range is 0.85 to 0.95, and the typical engineering value is set at 0.90. Real tumor microvascular contrast agent leakage exhibits isotropic diffusion characteristics, while registration misalignment artifacts inevitably exhibit unidirectional physical characteristics that are strictly polarized along the normal direction of the organ boundary.

[0056] If the absolute value of the cosine similarity of the included angle is not greater than the preset polarization verification threshold, the value of the corresponding voxel in the enhanced spatiotemporal gradient tensor data will remain unchanged.

[0057] This embodiment solves the technical problem of false high gradient signals generated at the edges due to non-rigid registration misalignment caused by the patient's physiological peristalsis by calculating the cosine similarity of the included angle and combining it with a preset polarization verification threshold for comparison and elimination. It achieves the unique technical effect of eliminating spatiotemporal dynamic noise by utilizing pure geometric orthogonality prior.

[0058] Furthermore, in order to accurately quantify the spatial distribution gradient of biological activity of a tumor from its internal necrotic core to its external invasive boundary, a smooth and biologically consistent normalized decay index is needed.

[0059] Obtain the preset edge attenuation control coefficient;

[0060] The preset method for obtaining the edge attenuation control coefficient is as follows: A retrospective statistical analysis was performed on bladder tumor images with known lymph node metastasis pathology results in the sample set. Scatter plots of microvessel density distribution at different tumor depths were extracted. A negative exponential correlation curve between microvessel density and distance from the boundary was generated using the least squares method. The attenuation rate parameter of this curve was extracted as the coefficient value. The effective range of this coefficient is [insert range here]. to Typical engineering value .

[0061] Obtain the average three-dimensional Euclidean distance from all voxels within any shell region to the boundary of the initial mask data, and obtain the maximum three-dimensional Euclidean distance from the tumor centroid of the initial mask data to the boundary.

[0062] The division operation yields the three-dimensional Euclidean normalized distance value, as shown in the following formula:

[0063] ,in, Indicates the first The three-dimensional Euclidean normalized distance values ​​of each shell region Indicates the first The average 3D Euclidean distance from all voxels in each shell region to the initial mask volume data boundary. This represents the maximum three-dimensional Euclidean distance from the tumor centroid to the boundary in the initial mask volume data.

[0064] The active bias value corresponding to the shell region is calculated using the following negative exponential decay formula:

[0065] ,in, Indicates the first The activity bias value of each shell region The preset edge attenuation control coefficient, Represents the natural constant.

[0066] This embodiment calculates the activity bias value by combining three-dimensional Euclidean normalized distance with a negative exponential decay formula, which solves the technical defect of traditional feature extraction methods that cannot accurately and continuously characterize the spatial heterogeneity of internal hierarchies, and realizes the precise mathematical quantification of the spatial distribution gradient of tumor biological activity.

[0067] Furthermore, the local temporal variance is enormous and has no absolute upper limit, while the activity bias value range is limited. If they are directly added without processing, the large variance, being constant, will cause gradient explosion, completely destroying the subsequent probability gating mechanism. To achieve a nonlinear smooth mapping between spatial topological bias and local temporal dynamic variance, and other multidimensional heterogeneous feature data, a nonlinear activation mapping is performed:

[0068] Get the weighted sum result Weighted summation result The generation logic includes independently configured scalar weight constants for each component. These weight constants are pre-calibrated and set by evaluating the influence of spatial morphology and temporal dynamics on tumor invasiveness using the analytic hierarchy process.

[0069] Using the Sigmoid activation function as the preset activation function, calculate the activity confidence value:

[0070] ,in, This represents the confidence level of activity. Represents the natural constant. This represents the weighted summation result.

[0071] This embodiment solves the technical problem of gradient explosion caused by the inconsistency of dimensions among multi-source heterogeneous physical quantities and the interference of extreme outliers by obtaining the weighted summation result and calling the Sigmoid activation function for nonlinear mapping. It realizes the unified conversion of complex spatiotemporal physical signals into normalized active confidence values ​​with clear probabilistic gating significance.

[0072] Furthermore, since the core process highly depends on the difference in CT intensity values ​​per voxel, high-frequency discrete noise is exponentially amplified by variance calculations, completely obscuring the real, subtle hemodynamic gradient signals. To effectively suppress the thermal noise of the CT equipment hardware and the high-frequency pulse discrete noise generated by the image reconstruction algorithm itself, pre-processing spatial filtering and purification are performed:

[0073] Temporal positioning occurs before calculating the local time variance.

[0074] Using a preset Gaussian smoothing filter kernel, three-dimensional spatial convolution filtering operations are performed on the first CT intensity value and the second CT intensity value of each voxel in the arterial phase CT volume data and the venous phase CT volume data, respectively, to generate smooth arterial phase CT volume data and smooth venous phase CT volume data that represent the removal of high-frequency pulse discrete noise.

[0075] The preset Gaussian smoothing filter kernel is set to a 3D spatial convolution kernel with a normally distributed weight matrix, and the kernel size is set to a typical engineering value. Voxel space, typical value of standard deviation parameter is Three-dimensional spatial convolution filtering operations smoothly diffuse abnormally isolated high-frequency noise into the surrounding neighborhood.

[0076] Numerical values ​​are re-extracted from the smooth arterial phase CT volume data and the smooth venous phase CT volume data based on voxel coordinates to obtain new first CT intensity values ​​and new second CT intensity values, and thus new local temporal variance.

[0077] This embodiment solves the technical problem of severe local temporal variance calculation caused by inherent high-frequency discrete noise in the original medical image data by introducing a preset Gaussian smoothing filter kernel to perform three-dimensional spatial convolution filtering operation, and realizes effective smoothing and high-fidelity restoration of the voxel-level CT intensity change reference signal.

[0078] Furthermore, to establish an end-to-end metastasis probability assessment mapping link with high clinical interpretability, clinicians cannot directly understand the non-linear S-shaped correspondence between linear volume percentage and actual biological metastasis risk. Directly outputting percentage data would lead to a significant reduction in the efficacy of auxiliary diagnosis. Therefore, a transformation derivation is performed in the final risk output stage:

[0079] Obtain a pre-stored logistic regression network as a pre-trained classifier. The logistic regression network contains trained and converged intercept parameters and slope weight parameters.

[0080] The logistic regression network uses a standard generalized linear regression model. The training dataset is obtained by collecting medical data containing voxel volume percentage features and gold-standard binary classification labels; the prediction error is calculated using a binary cross-entropy loss function; and the intercept and slope weight parameters are iteratively updated using a stochastic gradient descent algorithm. The model convergence termination condition is strictly set to continuous... The cross-entropy loss value on the validation set fluctuates less than [amount] within [number] training iterations. .

[0081] After multiplying the volume percentage and slope weight parameters, the intercept parameter is added to calculate the logarithmic probability value.

[0082] The logarithmic probability value removes the upper and lower bound constraints of the independent variable in mathematical space, giving it the ability to fully map to the real number field.

[0083] The logarithmic odds value is input into an exponential function to convert it into a percentage value of metastasis probability, which is then used as a lymph node metastasis prediction value and written into the auxiliary report.

[0084] This embodiment solves the technical problem that high-dimensional physical indicators are difficult to directly assist clinicians in intuitive diagnosis and decision-making by loading a logistic regression network containing trained convergence parameters for transformation operations. It realizes the accurate transformation of abstract volume ratio physical indicators into intuitive and highly interpretable transition probability percentage values.

[0085] Furthermore, due to differences in tube voltage settings among different CT scanning devices or variations in patients' own circulatory metabolic rates, there are scenarios where the background signal-to-noise ratio (SNR) of contrast-enhanced images exhibits significant individual fluctuations. In such scenarios, if a fixed empirical value is used as the preset signal threshold, a large number of "false positive" voxels will be retained for patients with high background noise, while crucial "true positive" transfer signals will be missed for patients with low sensitivity to contrast agents.

[0086] Temporal node localization occurs after the enhanced spatiotemporal gradient tensor is obtained, and during the process of determining the truncated voxels. The steps for obtaining the preset signal threshold include: defining the preset signal threshold as... From the nested shell regions, extract the data of the first nested shell located at the innermost layer; the data of the first nested shell corresponds to the region where the geometric centroid of the tumor is located.

[0087] Obtain the set of target voxels corresponding to the coordinates of the first nested shell data in the enhanced spatiotemporal gradient tensor; calculate the mean gradient intensity of the target voxel set to obtain the background noise baseline data; the background noise baseline data characterizes the residual physical values ​​of avascular necrotic tissue in the enhanced spatiotemporal gradient tensor under the current imaging conditions. Calculate the background noise baseline data using a preset sensitivity coefficient to generate a preset signal threshold, the specific formula of which is as follows:

[0088] ,in, Indicates the preset signal threshold. This represents the preset sensitivity coefficient. This represents the augmented spatiotemporal gradient tensor. This represents the first nested shell data. This represents the function for calculating the mean.

[0089] For the preset sensitivity coefficient The acquisition method was as follows: by statistically analyzing the images of the calibrated control group, the signal standard deviation multiple between normal muscle tissue and known necrotic tissue was calculated, and the typical engineering value was set as follows. to .

[0090] Through the above technical solution, this embodiment reuses the shell region data generated in the previous step and uses the physiological dead zone inside the tumor as an "in-situ reference" to achieve dynamic adaptive calibration of the preset signal threshold. This processing mechanism solves the technical problem in the core basic solution that the fixed threshold cannot adapt to individual image differences, and ensures the physical validity of the voxel cutoff determination from the underlying logic of data processing.

[0091] This embodiment solves the technical problem that the core basic scheme is prone to misidentifying equipment noise as micro-transfer signals in complex imaging environments by enhancing the nonlinear neighborhood verification of the spatiotemporal gradient tensor and the three-dimensional activity confidence tensor. It achieves a significant reduction in the false positive rate of the transfer prediction value in the final auxiliary report while ensuring the sensitivity of detecting small lesions.

[0092] This application also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements a method for predicting lymph node metastasis in bladder cancer.

[0093] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0098] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting lymph node metastasis in bladder cancer, characterized in that, Includes the following steps: Acquire arterial phase CT volume data of the target region within a preset three-dimensional spatial coordinate system and the first CT intensity value of each voxel, venous phase CT volume data and the second CT intensity value of each voxel, as well as initial mask volume data characterizing the three-dimensional contour of the bladder tumor. A multi-scale three-dimensional morphological erosion operation is performed on the initial mask data, progressing from the centroid representing the bladder tumor to the boundary, to obtain a set of nested shell regions. Based on the three-dimensional Euclidean normalized distance from each shell region to the boundary of the initial mask data, the corresponding active bias value is calculated for each shell region. For each target voxel located within the initial mask data, calculate the local temporal variance of its first CT intensity value and second CT intensity value, and then perform a weighted summation with the activity bias value of its shell region to obtain the weighted summation result, which is then input into a preset activation function and mapped to the activity confidence value of the voxel. The activity confidence values ​​of all target voxels are matrix-stitched according to a preset three-dimensional spatial coordinate system to construct a three-dimensional activity confidence tensor of the target voxels. The voxel-by-voxel difference between the venous phase CT volume data and the arterial phase CT volume data is calculated to generate an initial spatiotemporal gradient tensor. The three-dimensional activity confidence tensor and the initial spatiotemporal gradient tensor are multiplied voxel-by-voxel to obtain the enhanced spatiotemporal gradient tensor. The corresponding voxels whose enhanced spatiotemporal gradient tensor is greater than the preset signal threshold are identified as cut-off voxels. Their volume proportion in the initial mask volume data is calculated and input into the pre-trained classifier, and an auxiliary report containing lymph node metastasis prediction values ​​is output.

2. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, After mapping to the activity confidence value of the target voxel, the process also includes edge hard truncation operations based on three-dimensional geometric isolation: By performing a three-dimensional morphological boundary extraction operation on the initial mask data to extract the width of the single voxel, the outermost single-layer voxel set data is isolated. Determine whether the coordinates of the target voxel belong to the outermost single-layer voxel set data; If the coordinates of the target voxel are determined to belong to the outermost single-layer voxel set data, the mapping result of its weighted summation value is ignored, and the activity confidence value of the target voxel is forcibly reset to 0.

3. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, After obtaining the enhanced spatiotemporal gradient tensor, artifact removal operations based on the principle of geometric orthogonality are also included: By extracting voxels from the three-dimensional shell region of the initial mask data and applying a three-dimensional spatial differential operator to calculate its gradient direction, the volume surface normal vector of the initial mask data in the preset three-dimensional spatial coordinate system is generated. For each voxel in the enhanced spatiotemporal gradient tensor, the difference gradient vector in the preset three-dimensional coordinate system is calculated by applying a three-dimensional spatial differential operator. Calculate the cosine similarity of the angle between the surface normal vector and the difference gradient vector for the same voxel; If the absolute value of the cosine similarity of the included angle is greater than the preset polarization verification threshold, the corresponding voxel is determined to be a physical misalignment artifact offset along the normal direction, and the value of the corresponding voxel in the enhanced spatiotemporal gradient tensor is overwritten with the value 0. If the absolute value of the cosine similarity of the included angle is not greater than the preset polarization verification threshold, the value of the corresponding voxel in the enhanced spatiotemporal gradient tensor data will remain unchanged.

4. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, The specific calculation steps for the active bias value include: Obtain the preset edge attenuation control coefficient; Obtain the average three-dimensional Euclidean distance from all voxels within any shell region to the boundary of the initial mask data, and obtain the maximum three-dimensional Euclidean distance from the tumor centroid to the boundary of the initial mask data; The division operation yields the three-dimensional Euclidean normalized distance value, as shown in the following formula: ,in, Indicates the first The three-dimensional Euclidean normalized distance values ​​of each shell region Indicates the first The average 3D Euclidean distance from all voxels in each shell region to the initial mask volume data boundary. This represents the maximum three-dimensional Euclidean distance from the tumor centroid to the boundary in the initial mask volume data. The active bias value corresponding to the shell region is calculated using the following negative exponential decay formula: ,in, Indicates the first The activity bias value of each shell region The preset edge attenuation control coefficient, Represents the natural constant.

5. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, The specific steps for calculating the activity confidence value include: Obtain the weighted summation result: Using the Sigmoid activation function as the preset activation function, calculate the activity confidence value: ,in, This represents the confidence level of activity. Represents the natural constant. This represents the weighted summation result.

6. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, Before calculating the local time variance, the following steps are also included: Using a preset Gaussian smoothing filter kernel, three-dimensional spatial convolution filtering operations are performed on the first CT intensity value and the second CT intensity value of each voxel in the arterial phase CT volume data and the venous phase CT volume data, respectively, to generate smooth arterial phase CT volume data and smooth venous phase CT volume data that represent the removal of high-frequency pulse discrete noise. Voxel intensity values ​​are re-extracted from smooth venous phase CT volume data and smooth arterial phase CT volume data to obtain new first CT intensity values ​​and new second CT intensity values, and thus new local temporal variance.

7. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, The output includes a supplementary report containing predicted values ​​for lymph node metastasis. The specific calculation steps include: Obtain a pre-stored logistic regression network as a pre-trained classifier. The logistic regression network contains trained and converged intercept parameters and slope weight parameters. After multiplying the volume percentage and slope weight parameters, the intercept parameter is added to calculate the logarithmic probability value. The logarithmic odds value is input into an exponential function to convert it into a percentage value of metastasis probability, which is then used as a lymph node metastasis prediction value and written into the auxiliary report.

8. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, The specific steps for generating a set of nested shell regions include: Obtain a set of pre-defined monotonically increasing morphological erosion radius sequences; Using each corrosion radius in the corrosion radius sequence, a recursive three-dimensional morphological corrosion operation is performed on the initial mask data to generate multiple nested mask sequences with different shrinkage degrees. Following the order of progression from the centroid to the boundary, voxel-level Boolean difference operations are performed on two adjacent nested body masks, and the resulting closed surrounding voxel sets are defined as the corresponding shell regions.

9. The method for predicting lymph node metastasis in bladder cancer according to claim 1, characterized in that, The steps for obtaining the preset signal threshold include: Extract the data of the first nested shell located at the innermost layer from the sequentially nested shell regions; Obtain the set of target voxels in the enhanced spatiotemporal gradient tensor that correspond to the coordinates of the first nested shell data; Calculate the mean gradient intensity of the target voxel set to obtain the background noise baseline data; The background noise baseline data is weighted using a preset sensitivity coefficient to generate a preset signal threshold.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-9.

Images