Cloud-based collaborative prostate cancer multi-modal data analysis system

Abstract

The present application relates to the technical field of computer-aided diagnosis, in particular to a prostate cancer multi-modal data analysis system based on cloud collaboration, which comprises a biochemical image joint acquisition module, a constraint hyperplane space construction module, a lesion texture feature extraction module, a projection correction module and a risk assessment module.In the present application, biochemical antigen concentration and gland volume parameters are converted into density reference values, a geometric subspace perpendicular to the azimuth axis is constructed in the feature space to generate a constraint normal vector, the lesion gray level dispersion is extracted and the feature deviation distance is calculated in this direction, the inner product operation is used to perform subtraction correction on the feature vectors exceeding the tolerance threshold, the texture noise interference caused by benign hyperplasia is eliminated, the image features are precisely aligned in the physiological dimension, and the specificity and accuracy of the malignant risk assessment are improved by comparing the cosine value of the calibrated feature vector with the typical cancer prototype in the cloud.

Description

Technical Field

[0001] This invention relates to the field of computer-aided diagnostic technology, and in particular to a cloud-based collaborative multimodal data analysis system for prostate cancer. Background Technology

[0002] The field of computer-aided diagnostic technology encompasses the digital technology of acquiring, storing, extracting features, and comprehensively analyzing medical images, pathological slides, physiological signals, and clinical biochemical indicators using computer hardware and software. It constructs auxiliary interpretation logic by combining medical imaging, biostatistics, pattern recognition algorithms, and pathophysiology principles to assist clinicians in locating, detecting, segmenting, classifying, and quantifying lesions in the human body, thus providing objective data support for the formulation of disease diagnosis and treatment plans. The traditional prostate cancer multimodal data analysis system refers to an integrated architecture that comprehensively evaluates patients' prostate-specific antigen serum indicators, multi-parameter magnetic resonance imaging sequences, and biopsy pathological images. It relies on the hospital's internal LAN and local image archiving and communication system workstations to directly retrieve MRI T2-weighted imaging and diffusion-weighted imaging data, and combines them with biochemical values ​​measured in the laboratory. Physicians manually delineate regions of interest on the imaging plane, or use locally pre-built convolutional neural networks and support vector machine classifiers to automatically label suspicious lesions in the peripheral zone and transition zone of the prostate and calculate the probability of benign or malignant lesions. Finally, it generates an analysis report containing lesion location and risk level based on the Gleason scoring criteria.

[0003] Traditional prostate cancer analysis architectures rely on local workstations to independently process biochemical indicators and magnetic resonance imaging data. They rely solely on physicians' manual delineation or pre-set classifiers for mechanical probability calculations, ignoring the spatial geometric mapping relationship between physiological density benchmarks and lesion texture features. This results in a lack of vectorized constraints on image grayscale distribution features based on antigen concentration. When faced with interference from atypical hyperplastic tissue, it is unable to correct feature deviations through a unified hyperplane benchmark, causing dimensional alignment errors when comparing lesion feature vectors with pathological prototypes, making it difficult to accurately define the boundary of malignancy risk. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a cloud-based collaborative multimodal data analysis system for prostate cancer.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a cloud-based collaborative prostate cancer multimodal data analysis system includes:

[0006] The biochemical imaging combined acquisition module acquires T2-weighted magnetic resonance imaging data and tPSA concentration values, counts prostate capsule pixels and calculates the total prostate volume by combining slice thickness and spacing, and divides the tPSA concentration value by the total prostate volume to generate a PSA density baseline value.

[0007] The constrained hyperplane space construction module defines the origin and azimuth axis in the feature space based on the PSA density reference value, constructs a benign geometric subspace perpendicular to the azimuth axis, and generates a geometric constraint normal vector.

[0008] The lesion texture feature extraction module defines the lesion region of sequentially scanned T2-weighted magnetic resonance imaging data based on the geometric constraint normal vector, extracts grayscale dispersion and texture regularity values, and generates the original image feature vector.

[0009] The projection correction module calculates the inner product of the original image feature vector and the geometric constraint normal vector, divides it by the vector magnitude to obtain the feature deviation distance value, compares the feature deviation distance value with the preset tolerance threshold, and if it exceeds the preset tolerance threshold, performs vector subtraction correction; otherwise, it retains the value and generates a physiological alignment feature vector.

[0010] The risk assessment module calls the prototype vector of typical prostate cancer features in the cloud, calculates the cosine of the angle between the physiologically aligned feature vector and the prototype vector of typical prostate cancer features, and generates a quantitative index of malignancy risk.

[0011] As a further aspect of the present invention, the PSA density benchmark value includes biochemical antigen concentration parameters and gland volume space parameters; the geometric constraint normal vector includes spatial orientation guidance components and hyperplane normal coefficients; the original image feature vector includes pixel grayscale distribution statistics and lesion texture topology parameters; the physiological alignment feature vector includes feature projection correction components and spatial geometric calibration terms; and the malignancy risk quantification index includes feature similarity probability values ​​and lesion carcinogenesis tendency scores.

[0012] As a further aspect of the present invention, the biochemical image acquisition module includes:

[0013] The image biochemical data analysis submodule acquires T2-weighted magnetic resonance imaging data and tPSA concentration values, identifies the prostate capsule edge trajectory within the imaging data, traverses the internal region of the edge trajectory, accumulates the number of discrete pixels, and generates prostate capsule pixel statistics.

[0014] The gland volume space calculation submodule calls the pixel statistics of the prostate capsule, introduces the scanning layer thickness parameter, layer spacing parameter and single pixel plane physical area, and combines the morphological calibration factor and boundary convergence coefficient to identify the three-dimensional volume information of the gland after spatial geometric correction, and calculates the total volume value of the prostate.

[0015] The biochemical density benchmark mapping submodule calls the total prostate volume value and tPSA concentration value to construct the ratio relationship between antigen mass and gland volume, quantifies the distribution density of biochemical indicators per unit volume, and generates a PSA density benchmark value.

[0016] As a further aspect of the present invention, the total prostate volume is expressed by the formula:

[0017] The pixel statistics of the prostate capsule and the physical area of ​​a single pixel are obtained. The physical area conversion of the two in the planar dimension is performed to extract the true physical area features at the slice level. Then, the physical area features are combined with the scanning slice thickness parameters and the interslice spacing parameters to perform interslice expansion in the depth direction and three-dimensional geometric space integration to construct a preliminary organ volume assessment. Finally, for this preliminary organ volume assessment, a preset morphological calibration factor is introduced to perform three-dimensional morphological distortion compensation processing, and the boundary convergence coefficient is combined simultaneously to perform numerical approximation and error convergence correction calculations on the edge contour. Finally, the total prostate volume value that accurately represents the three-dimensional spatial distribution of the target organ is obtained.

[0018] As a further aspect of the present invention, the constrained hyperplane space construction module includes:

[0019] The feature space benchmark definition submodule filters the benign sample set based on the PSA density benchmark value, calculates the distribution centroid of the benign sample set in the multidimensional feature space, sets the origin of the coordinate system, plans and identifies the principal component directions of the feature distribution, and generates the feature space orientation axis parameters.

[0020] The benign subspace construction submodule calls the feature space orientation axis parameters, performs spatial orthogonal decomposition operation, constructs a geometric plane perpendicular to the principal component direction, defines the projection coverage of benign samples on the geometric plane, and generates the boundary of the benign geometric subspace.

[0021] The constraint normal generation submodule calls the PSA density reference value, feature space orientation axis parameters, and benign geometric subspace boundary limits to analyze the degree of deviation between antigen density and benign distribution, calculate the hyperplane normal deflection weight, and combine the orientation axis parameters to perform vector synthesis to generate the geometric constraint normal vector.

[0022] As a further aspect of the present invention, the lesion texture feature extraction module includes:

[0023] The scanning area guidance and definition submodule loads T2-weighted magnetic resonance imaging data according to the geometric constraint normal vector, establishes a scanning index along the spatial gradient direction indicated by the geometric constraint normal vector, identifies the tissue signal abrupt change edge in the cross section orthogonal to the geometric constraint normal vector, identifies the lesion voxel set through gradient determination, and generates the spatial boundary of the lesion area.

[0024] The grayscale texture feature extraction submodule calls the spatial boundary of the lesion region, extracts the brightness intensity value of the pixel, calculates the second moment statistics of the grayscale histogram, quantifies the tissue density dispersion, constructs the grayscale co-occurrence matrix to analyze the topological structure of the pixel spatial arrangement, calculates and obtains the energy spectrum, contrast and inverse moment attributes, and generates grayscale dispersion and texture rule parameters.

[0025] The feature vector multidimensional aggregation submodule performs numerical standardization operations on the gray-level statistical values ​​and texture topology attributes of the differentiated physical dimensions for the gray-level dispersion and texture rule parameters, eliminates the differences in the numerical distribution scale between feature components, and performs data serial stitching operation according to the preset feature dimension order to generate the original image feature vector.

[0026] As a further aspect of the present invention, the projection correction module includes:

[0027] The projection deviation quantization submodule performs a dot product operation on the original image feature vector and the geometric constraint normal vector to extract the projection intensity of the feature on the normal constraint axis. At the same time, it analyzes the Euclidean norm of the geometric constraint normal vector, divides the dot product result by the Euclidean norm, performs numerical normalization processing, and generates the feature deviation distance value.

[0028] The tolerance discrimination decision submodule calls the feature deviation distance value and the geometric constraint normal vector, reads the preset tolerance threshold parameter, performs a numerical comparison, and determines whether the feature deviation distance value exceeds the allowable range of benign physiological fluctuations. If the determination result is out of bounds, the product of the deviation distance and the geometric constraint normal vector is calculated to construct the abnormal components to be removed. If the determination result is compliant, all corresponding correction terms are set to zero vectors to generate feature correction vector parameters.

[0029] The physiological alignment execution submodule calls the feature correction vector parameters, performs a linear vector subtraction operation in the multidimensional feature space, removes the normal projection components that violate geometric constraints in the original features, retains the tangential feature components that conform to the benign subspace distribution law, performs numerical stability verification and dimensional consistency check on the corrected vector data, and generates physiological alignment feature vectors.

[0030] As a further aspect of the present invention, the risk assessment module includes:

[0031] The prototype vector retrieval submodule establishes a data transmission channel with the cloud-based medical database, traverses the stored image feature data of confirmed cases, identifies the spatial centroid of malignant sample groups, extracts a set of multidimensional parameters characterizing the core pathological attributes of prostate cancer, and generates prototype vectors of typical carcinogenesis features.

[0032] The similarity calculation submodule calls the physiological alignment feature vector and the typical cancer feature prototype vector, performs the dot product operation and modulus calculation operation of the multidimensional space vector, solves the cosine value of the angle between the two feature vectors, quantifies the consistency between the feature distribution of the test sample and the malignant prototype in the geometric direction, and generates the feature distribution overlap coefficient.

[0033] The risk index generation submodule loads a preset risk classification mapping table for the feature distribution overlap coefficient, converts the cosine similarity value into a standardized probability score, and performs weighted aggregation processing on the contribution weight of the differentiated feature dimension to the malignancy judgment, calculates the quantitative value representing the occurrence of malignant lesions in the tissue under test, and generates a malignancy risk quantitative index.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0035] In this invention, biochemical antigen concentration and gland volume parameters are converted into density benchmark values. A geometric subspace perpendicular to the azimuth axis is constructed in the feature space to generate a constrained normal vector. Based on this, the gray-level dispersion of lesions is extracted and the feature deviation distance is calculated. The inner product operation is used to perform subtraction correction on feature vectors that exceed the tolerance threshold to eliminate texture noise interference caused by benign hyperplasia and ensure that the image features are accurately aligned in the physiological dimension. By comparing the cosine value of the angle between the calibrated feature vector and the typical cancerous prototype in the cloud, the distribution overlap between the sample and the confirmed case is quantified in all aspects, thereby improving the specificity and accuracy of malignancy risk assessment. Attached Figure Description

[0036] Figure 1 This is a system flowchart of the present invention;

[0037] Figure 2 This is a flowchart of the biochemical image joint acquisition module in this invention;

[0038] Figure 3 This is a flowchart of the constrained hyperplane space construction module in this invention;

[0039] Figure 4 This is a flowchart of the lesion texture feature extraction module in this invention;

[0040] Figure 5 This is a flowchart of the projection correction module in this invention;

[0041] Figure 6 This is a flowchart of the risk assessment module in this invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0044] Please see Figure 1 The cloud-based collaborative prostate cancer multimodal data analysis system includes:

[0045] The biochemical imaging combined acquisition module acquires T2-weighted magnetic resonance imaging data and tPSA concentration values ​​of the patient to be tested, scans the pixels in the prostate capsule and counts the total number, calculates the total volume of the prostate by combining the imaging slice thickness parameter and pixel spacing parameter, performs a division operation between the tPSA concentration value and the total volume of the prostate, and generates a PSA density baseline value.

[0046] The constrained hyperplane space construction module defines the origin of the coordinate system and the axis pointing to a specific direction in the preset multi-dimensional feature space based on the PSA density benchmark value, constructs a geometric subspace perpendicular to the axis and representing the benign physiological distribution area, and generates a geometric constraint normal vector.

[0047] The lesion texture feature extraction module performs pixel-level scanning on the lesion region in T2-weighted magnetic resonance imaging data according to the feature space dimension order defined by the geometric constraint normal vector, extracts the gray-level distribution dispersion value and texture arrangement rule value, and generates the original image feature vector.

[0048] The projection correction module calculates the inner product of the original image feature vector and the geometric constraint normal vector, divides it by the vector magnitude to obtain the feature deviation distance value, compares the feature deviation distance value with the preset tolerance threshold, performs vector subtraction correction operation for cases that exceed the preset tolerance threshold, retains the original value in cases that do not exceed the limit, and generates physiological alignment feature vector.

[0049] The risk assessment module calls the typical prostate cancer feature prototype vectors pre-stored on the cloud server, calculates the cosine value of the angle between the physiological alignment feature vector and the typical prostate cancer feature prototype vector, quantifies the degree of overlap between the feature distribution of the test sample and the confirmed case, and generates a malignancy risk quantification index.

[0050] The PSA density benchmark includes biochemical antigen concentration parameters and gland volume space parameters. The geometric constraint normal vector includes spatial orientation guidance components and hyperplane normal coefficients. The original image feature vector includes pixel grayscale distribution statistics and lesion texture topology parameters. The physiological alignment feature vector includes feature projection correction components and spatial geometric calibration terms. The malignancy risk quantification indicators include feature similarity probability values ​​and lesion carcinogenesis tendency scores.

[0051] Please see Figure 2 The biochemical imaging joint acquisition module includes:

[0052] The image biochemical data analysis submodule acquires T2-weighted magnetic resonance imaging data and tPSA concentration values, identifies the prostate capsule edge trajectory within the imaging data, traverses the internal region of the edge trajectory, accumulates the number of discrete pixels, and generates prostate capsule pixel statistics.

[0053] The submodule directly connects to a medical 3.0T MRI scanner workstation via a hardwired interface or DICOM network protocol to read T2-weighted MRI data sequences. These sequences contain 20 to 30 transverse images, with a single-slice resolution set to a 512×512 pixel matrix. The submodule synchronously receives serum tPSA (total prostate-specific antigen) concentration values ​​from the laboratory's LIS system, in floating-point format, with units uniformly converted to ng / mL. For the imaging data, the submodule incorporates an edge detection operator, employing a combination of Canny edge detection logic and morphological closing operations to identify high-contrast edge trajectories of the prostate capsule in each image layer. The processing logic first performs Gaussian filtering on the original DICOM images for noise reduction, with a convolution kernel size of 5×5. Then, it calculates the magnitude and direction of the image gradient, retaining local maxima. For non-closed edge segments, it uses the Euclidean distance minimization principle to connect and close them, forming a complete prostate capsule outline. The submodule iterates through all pixel coordinates within the contour, counts the total number of pixels labeled "glandular tissue," and converts the voxel count in three-dimensional space into discretized prostate capsule pixel statistics. For example, in an actual acquisition process, the system received a tPSA concentration of 8.5 ng / mL, and the total number of pixels within the capsule in all layered images was found to be 45,000 after image processing algorithms were used.

[0054] The gland volume space calculation submodule calls the pixel statistics of the prostate capsule, introduces the scanning slice thickness parameter, slice spacing parameter, and physical area of ​​a single pixel plane, and combines the morphological calibration factor and boundary convergence coefficient, using the following formula:

[0055] ;

[0056] The total volume of the prostate is calculated by identifying the three-dimensional volume information of the gland after spatial geometry correction.

[0057] in, This represents the total volume of the prostate gland. Represents the pixel count of the prostate capsule. Represents the physical area of ​​a single pixel plane. Represents the scanned slice thickness parameter. Represents the interlayer spacing parameter. Representative morphological calibration factor, Represents the boundary convergence coefficient;

[0058] The scan slice thickness and interslice spacing parameters are extracted from the DICOM header file. The physical area of ​​a single pixel plane is calculated by dividing the field of view (FOV) by the matrix dimension. To eliminate geometric distortion caused by phase encoding direction and volume estimation errors due to slicing effect during magnetic resonance imaging, the submodule introduces a morphological calibration factor and a boundary convergence coefficient. These two parameters are not fixed constants but are correction values ​​obtained through calibration tests on the phantom. The morphological calibration factor is used to compensate for the geometric approximation error of non-spherical glands, and the boundary convergence coefficient is used to correct edge blurring caused by part of the volume effect.

[0059] The specific value logic of each parameter in the formula is as follows:

[0060] Morphological calibration factor ( This value is determined by analyzing the eccentricity of the prostate. In this embodiment, based on the inverted cone morphology of most prostates, the factor is set to 0.98 by comparing the standard volume measured by transrectal ultrasound and combining the weighted correction of morphological statistics of the same race using cloud big data. This factor is used to correct the step-like error generated by the MRI slice stacking method.

[0061] Boundary convergence coefficient ( ): Considering that edge pixels of MRI images contain some signals from surrounding adipose tissue, this coefficient is set to 0.95 to tighten the volume calculation boundary and eliminate some interference from volume effects.

[0062] Physical area of ​​a single pixel plane ( Based on a FOV of 200mm×200mm and a matrix of 512×512, the calculated side length of a single pixel is approximately 0.39mm, and the area is approximately 0.1521mm². In the actual calculation, the previously obtained pixel statistics of 45000 were used, and the scanning layer thickness parameter was set to 3.0mm, and the interlayer spacing parameter was set to 0mm (no-interval scanning).

[0063] Substitute the above values ​​into the formula:

[0064] ;

[0065] The calculated total prostate volume was approximately 19114.7 mm³, or 19.11 mL. This result demonstrates that by introducing cloud-based collaboratively optimized calibration and convergence parameters, the error in slice-based volume estimation can be controlled within a clinically acceptable range, providing an accurate denominator for subsequent density calculations.

[0066] The biochemical density benchmark mapping submodule calls the total prostate volume value and tPSA concentration value to construct the ratio relationship between antigen mass and gland volume, quantifies the distribution density of biochemical indicators per unit volume, and generates a PSA density benchmark value.

[0067] The system retrieves the generated total prostate volume (19.11 mL) and the received tPSA concentration (8.5 ng / mL). The submodule performs a division operation to construct the ratio of antigen mass to glandular volume. This ratio directly reflects the ability of prostate tissue per unit volume to secrete PSA, and is an important physiological indicator for distinguishing benign prostatic hyperplasia (BPH) from prostate cancer. The calculation divides 8.5 by 19.11, retaining three decimal places, generating a baseline PSA density of 0.445 ng / mL². This baseline value is then stored in the system's temporary register, along with a timestamp and patient ID tag, serving as the core input parameter for subsequently constructing the constrained hyperplane space.

[0068] Please see Figure 3 The constrained hyperplane space construction module includes:

[0069] The feature space benchmark definition submodule filters the benign sample set based on the PSA density benchmark value, calculates the distribution centroid of the benign sample set in the multidimensional feature space to set the origin of the coordinate system, plans and identifies the principal component directions of the feature distribution, and generates the feature space orientation axis parameters.

[0070] The system initializes the coordinate system of the multidimensional feature space by pre-loading a high-dimensional feature dataset containing 1000 pathologically confirmed benign prostatic hyperplasia (BPH) patients. The submodule first filters out a subset of benign samples with PSA densities within ±10% of the baseline value and processes this subset using Principal Component Analysis (PCA). The processing includes calculating the covariance matrix of the feature data, performing eigenvalue decomposition on the covariance matrix, and extracting the three largest eigenvectors, which are defined as the principal component directions of the feature distribution. These directions represent the dimensions where the benign tissue characteristics change most drastically at that PSA density level. The submodule calculates the geometric centroid (mean vector) of this subset in the feature space and sets this centroid as the origin of the local coordinate system for the current analysis. The final generated feature space orientation axis parameters contain three unit orthogonal vectors, which together define the baseline reference frame for observing the samples. For example, under a PSA density of 0.445 ng / mL², the calculated centroid coordinates of the benign distribution are (0.5, 0.3, 0.2), and the principal component directions... The result is (0.8, 0.6, 0).

[0071] The benign subspace construction submodule calls the feature space orientation axis parameters, performs spatial orthogonal decomposition operation, constructs a geometric plane perpendicular to the principal component direction, defines the projection coverage of benign samples on the geometric plane, and generates the boundary of the benign geometric subspace.

[0072] A geometric plane perpendicular to the direction of the first principal component (i.e., the direction of maximum variation) is constructed, and this plane is considered a "safe cross-section" for benign feature fluctuations. Subsequently, the submodule projects all selected benign samples onto this plane and uses the Convex Hull algorithm to calculate the minimum enclosing geometry of the projected points. To improve generalization ability, the submodule extends the convex hull boundary outward by a buffer distance, which is determined by twice the standard deviation of the global sample projection points fed back by the cloud-based collaborative computing engine based on the federated learning strategy. This defines the projection coverage of benign samples on the geometric plane, generating the boundary of the benign geometric subspace. Mathematically, this boundary is represented by a set of inequality constraints, used to subsequently determine whether the feature vector of the test sample deviates from the normal fluctuation range of the benign population.

[0073] The constraint normal generation submodule calls the PSA density baseline value, feature space orientation axis parameters, and benign geometric subspace boundary limits to analyze the deviation of antigen density from benign distribution, using the following formula:

[0074] ;

[0075] Calculate the hyperplane normal deflection weight, combine it with the azimuth axis parameters to perform vector synthesis, and generate the geometric constraint normal vector;

[0076] in, Represents the hyperplane normal deflection weight. Represents the PSA density benchmark value. This represents a reference value for the average density of a benign population. Represents the variance of the density distribution within a benign geometric subspace. Represents the background noise variance term. The local geometric curvature representing the characteristic manifold. The effective coverage radius representing the clustering of benign samples. Represents the basic bias coefficient;

[0077] The increments of each parameter in the formula and the basis for their values ​​are as follows:

[0078] Reference value for average density of benign population ( ): Mean PSA density of benign patients of the same age group, obtained from real-time synchronization with a cloud database. In this embodiment, for the 60-70 age group, it is set to 0.15 ng / mL².

[0079] The variance of the density distribution within a benign geometric subspace ( ): This quantifies the dispersion of the density of benign samples. Statistically, this value is 0.02.

[0080] Background noise variance term ( ): To prevent the denominator from being zero and to smooth out the extremely small positive numbers introduced by numerical fluctuations, it is set to 0.001.

[0081] Local geometric curvature of characteristic manifolds The mean curvature () is obtained by calculating the average curvature of the benign sample manifold at its centroid, and is used to characterize the degree of curvature of the benign distribution boundary. High curvature indicates a complex boundary between benign and malignant samples, requiring enhanced constraints. Calculated using a manifold learning algorithm, this value is 1.2.

[0082] Effective coverage radius of benign sample clustering ( ): The normalized radius of the region covered by the benign sample set in the feature space, which is measured to be 2.5 in this example.

[0083] Basic bias coefficient ( ): A constant used to ensure minimum constraint strength, empirically set to 0.5. In actual calculations, the PSA density benchmark value of 0.445 is substituted into the formula;

[0084] The molecular part is ;

[0085] The denominator is ;

[0086] The ratio of the first term is approximately 2.034;

[0087] The second product is ;

[0088] Calculation results:

[0089] ;

[0090] The results indicate that, due to the significantly higher PSA density of the test samples compared to the benign mean (0.445 > 0.15) and their location in a region with complex boundaries, the system generates a large deflection weight of 6.602. The submodule applies this weight to the principal orientation axis of the feature space, causing the constraint normal to deflect towards the high-probability malignant region, ultimately generating a geometric constraint normal vector. This vector guides the subsequent texture extraction and feature correction, effectively improving the feature sensitivity for high-risk samples.

[0091] Please see Figure 4 The lesion texture feature extraction module includes:

[0092] The scanning region guidance and definition submodule loads T2-weighted magnetic resonance imaging data based on the geometric constraint normal vector, establishes a scanning index along the spatial gradient direction indicated by the geometric constraint normal vector, identifies the tissue signal abrupt change edge in the cross section orthogonal to the geometric constraint normal vector, identifies the lesion voxel set through gradient determination, and generates the spatial boundary of the lesion region.

[0093] The signal gradient of each voxel within the orthogonal section of the normal vector is calculated. When a sudden change in tissue signal is detected (i.e., the gradient magnitude exceeds a preset noise threshold of 20.0), the region growing algorithm is triggered. The algorithm uses high-gradient points as seeds to aggregate voxels with similar signal intensities, forming a set of candidate lesion voxels. To eliminate artifacts from vascular pulsation or rectal gas, the submodule calculates the volume of the connected components of the set, eliminating tiny discrete regions with a volume less than 50 mm³, ultimately locking in and generating the spatial boundary of the region of interest (ROI). This boundary consists of a series of three-dimensional coordinate points that precisely enclose the core region of the suspected lesion.

[0094] The grayscale texture feature extraction submodule calls the spatial boundary of the lesion region, extracts the brightness intensity value of the pixel, calculates the second moment statistics of the grayscale histogram, quantifies the tissue density dispersion, constructs the grayscale co-occurrence matrix to analyze the topological structure of the pixel spatial arrangement, calculates and obtains the energy spectrum, contrast and inverse moment attributes, and generates grayscale dispersion and texture rule parameters.

[0095] The brightness values ​​are normalized to a grayscale level of 0-255. Then, the second-order moment statistics of the grayscale histogram are calculated, including variance and skewness, to quantify the overall dispersion of the organization density. The core operation is constructing the Gray-Level Co-occurrence Matrix (GLCM) to analyze the topological structure of the pixel spatial arrangement. The calculation distance is set. Pixels were selected at four angles: 0°, 45°, 90°, and 135°. Based on GLCM, energy spectrum, contrast, and inverse difference moment attributes were obtained through computation. Table 1 shows some texture feature data extracted during a specific execution.

[0096] Table 1: Texture Feature Parameters of Lesion Area

[0097]

[0098] The submodule averages the feature values ​​in four directions and combines this with second-order moment statistics to generate grayscale dispersion and texture regularity parameters. High contrast and low inverse moment indicate disordered tissue structure, consistent with the microscopic pathological characteristics of malignant tumors.

[0099] The feature vector multidimensional aggregation submodule performs numerical standardization operations on the gray-level statistical values ​​and texture topology attributes of the differentiated physical dimensions for gray-level dispersion and texture rule parameters, eliminates the differences in numerical distribution scale between feature components, and performs data serial stitching operation according to the preset feature dimension order to generate the original image feature vector.

[0100] Because grayscale statistical values ​​(such as variance) and texture topological properties (such as energy spectrum) have different physical dimensions, direct combination will lead to features with larger values ​​dominating the classifier. Therefore, the submodule adopts the Z-Score normalization method, using the formula... Each feature dimension is dimensionless. After eliminating the differences in numerical distribution scales between feature components, the submodule performs a serial data concatenation operation according to the preset feature dimension order (e.g., [energy, contrast, inverse moment, grayscale variance, grayscale skewness]). Furthermore, spatial weights are assigned based on the importance scores of each feature in the previous training set. For example, the contrast feature is assigned a weight of 1.2, while grayscale skewness is assigned a weight of 0.8. Finally, a dimension is constructed. The numerical array is used to generate the original image feature vector.

[0101] Please see Figure 5 The projection correction module includes:

[0102] The projection deviation quantization submodule performs a dot product operation on the original image feature vector and geometric constraint normal vector to extract the projection intensity of the feature on the normal constraint axis. At the same time, it analyzes the Euclidean norm of the geometric constraint normal vector, divides the dot product result by the Euclidean norm, performs numerical normalization processing, and generates the feature deviation distance value.

[0103] Based on the feature vector of the original image (denoted as vector) ) and geometric constraint normal vector (denoted as vector) This aims to quantify the "spurious" components implicit in feature vectors, caused by non-pathological factors (such as individual anatomical differences). The submodule performs a dot product multiplication of two multidimensional vectors. This submodule extracts the projection intensity of the features onto the normal constraint axis. This projection intensity represents the magnitude of the feature vector's component along the constraint direction. Simultaneously, the submodule analyzes the Euclidean norm of the geometric constraint normal vector. Divide the result of the dot product by the Euclidean norm, i.e. Numerical normalization is then performed. The geometric meaning of this calculation is clear: it calculates the projection length of the feature vector onto the constraint normal, generating the feature deviation distance value. Assume the original image feature vector is... The constrained normal vector, after normalization, is The dot product is The value 1.68 is the feature deviation distance, which represents the vertical distance of the feature point from the benign hyperplane.

[0104] The tolerance discrimination decision submodule calls the feature deviation distance value and the geometric constraint normal vector, reads the preset tolerance threshold parameter, performs a numerical comparison, and determines whether the feature deviation distance value exceeds the allowable range of benign physiological fluctuations. If the determination result is out of bounds, the product of the deviation distance and the geometric constraint normal vector is calculated to construct the abnormal components to be removed. If the determination result is compliant, all corresponding correction item values ​​are set to zero vectors to generate feature correction vector parameters.

[0105] Read the preset tolerance threshold parameter from the memory. The threshold setting logic is based on 3 benign sample libraries. The principle (3 standard deviations) is set to 0.5 in this embodiment. The submodule performs numerical comparisons to determine whether the feature deviation distance (1.68) exceeds the allowable range (0.5) for benign physiological fluctuations. Because... The judgment result is out of bounds, indicating that the feature vector contains too many deviation components perpendicular to the benign manifold caused by individual differences or noise. At this point, the submodule calculates the product of the deviation distance and the geometric constraint normal vector to construct the vector of outlier components to be removed. If the judgment result is compliant (e.g., distance is 0.2), then all corresponding correction terms are set to zero vectors. Finally, feature correction vector parameters are generated, which indicate the specific numerical direction to be subtracted from the original features.

[0106] The physiological alignment execution submodule calls the feature correction vector parameters, performs a vector linear subtraction operation in the multidimensional feature space, removes the normal projection components that violate geometric constraints in the original features, retains the tangential feature components that conform to the distribution law of the benign subspace, performs numerical stability verification and dimensional consistency check on the corrected vector data, and generates physiological alignment feature vectors.

[0107] Perform a linear vector subtraction operation in a multidimensional feature space: The physical meaning of this operation is to remove the normal projection components that violate geometric constraints from the original features, retaining the tangential feature components that conform to the distribution law of the benign subspace. This orthogonal correction ensures that the retained feature differences mainly originate from pathological changes, rather than background noise. The submodule performs numerical stability checks on the corrected vector data, checking for outliers caused by floating-point overflows (such as NaN or Infinity), and reconfirming the consistency of vector dimensions. After this series of rigorous mathematical transformations, a physiologically aligned feature vector is generated. This vector is "pulled back" to the vicinity of the tangent plane of the benign manifold in the feature space, eliminating systematic bias and providing cleaner data input for subsequent risk assessment.

[0108] Please see Figure 6 The risk assessment module includes:

[0109] The prototype vector retrieval submodule establishes a data transmission channel with the cloud-based medical database, traverses the stored image feature data of confirmed cases, identifies the spatial centroid of malignant sample groups, extracts a set of multidimensional parameters characterizing the core pathological attributes of prostate cancer, and generates prototype vectors of typical carcinogenesis features.

[0110] This submodule serves as the core interaction port of the entire cloud-based collaborative ecosystem, establishing an encrypted data transmission channel with a cloud-based medical database that stores radiomics features of over 100,000 confirmed prostate cancer cases. The submodule iterates through the stored historical confirmed case image feature data, using the K-Means++ clustering algorithm to identify the spatial centroids of malignant sample groups with different Gleason scores (e.g., 3+3, 3+4, 4+3, etc.). Based on the feature dimension configuration of the current sample, it extracts a multidimensional parameter set (i.e., the coordinate vectors of each cluster center) that characterizes the core pathological attributes of prostate cancer. The submodule then performs a weighted average of these center vectors, with the weights depending on the sample density of each cluster, thereby generating a prototype vector of the most representative typical cancerous features. For example, this prototype vector is... This represents the "standard" characteristic pattern exhibited by typical prostate cancer within this feature space.

[0111] The similarity calculation submodule calls the physiological alignment feature vector and the typical cancer feature prototype vector, performs the dot product operation and modulus calculation operation of the multidimensional space vector, solves the cosine value of the angle between the two feature vectors, quantifies the degree of consistency between the feature distribution of the test sample and the malignant prototype in the geometric direction, and generates the feature distribution overlap coefficient.

[0112] Call the physiological alignment feature vector (denoted as ) and typical cancer feature prototype vector (denoted as The submodule performs the dot product operation on multidimensional vectors. ) and module length calculation operation ( Based on the cosine similarity formula:

[0113] ;

[0114] Calculate the cosine of the angle between two feature vectors. This value ranges from -1 to 1. The closer the value is to 1, the more parallel the feature direction of the test sample is to the typical cancer feature direction, i.e., the more similar the feature pattern. Quantify the degree of consistency between the feature distribution of the test sample and the malignant prototype in the geometric direction. Assuming the calculated cosine value is 0.85, generate the feature distribution overlap coefficient. This coefficient directly reflects the projection of the sample's "malignancy" onto its geometric shape.

[0115] The risk index generation submodule loads a preset risk classification mapping table for the feature distribution overlap coefficient, converts the cosine similarity value into a standardized probability score, and performs weighted aggregation processing on the contribution weight of the differentiated feature dimension to the malignancy judgment, calculates the quantitative value representing the occurrence of malignant lesions in the tissue under test, and generates a malignancy risk quantitative index.

[0116] The pre-defined risk grading mapping table is loaded, which is a probability transformation table trained based on a large-sample logistic regression model. The submodule converts the cosine similarity values ​​into standardized probability scores (0-100 points). For example, a coefficient of 0.85 corresponds to a high-risk interval with a mapping score of 88. Furthermore, the submodule combines the contribution weights of differentiated feature dimensions to malignancy detection (e.g., texture feature weight 0.6, morphological feature weight 0.4) and performs weighted aggregation to fine-tune the final score. The quantitative value representing the occurrence of malignant lesions in the tested tissue is calculated, generating a malignancy risk quantification index. Table 2 shows an example of risk assessment output:

[0117] Table 2: Risk Assessment Result Mapping Table

[0118]

[0119] In this example, the coefficient of 0.85 falls within the range of [0.7, 1.0]. The final output of the malignancy risk quantification index by the system is 88, and a "high risk" prompt is highlighted on the physician's operation interface to assist the physician in making a decision to perform a puncture biopsy.

[0120] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A cloud-based collaborative multimodal data analysis system for prostate cancer, characterized in that: The system includes: The biochemical imaging combined acquisition module acquires T2-weighted magnetic resonance imaging data and tPSA concentration values, counts prostate capsule pixels and calculates the total prostate volume by combining slice thickness and spacing, and divides the tPSA concentration value by the total prostate volume to generate a PSA density baseline value. The constrained hyperplane space construction module defines the origin and azimuth axis in the feature space based on the PSA density reference value, constructs a benign geometric subspace perpendicular to the azimuth axis, and generates a geometric constraint normal vector. The lesion texture feature extraction module defines the lesion region of sequentially scanned T2-weighted magnetic resonance imaging data based on the geometric constraint normal vector, extracts grayscale dispersion and texture regularity values, and generates the original image feature vector. The projection correction module calculates the inner product of the original image feature vector and the geometric constraint normal vector, divides it by the vector magnitude to obtain the feature deviation distance value, compares the feature deviation distance value with the preset tolerance threshold, and if it exceeds the preset tolerance threshold, performs vector subtraction correction; otherwise, it retains the value and generates a physiological alignment feature vector. The risk assessment module calls the cloud-based typical prostate cancer feature prototype vector, calculates the cosine value of the angle between the physiologically aligned feature vector and the typical prostate cancer feature prototype vector, and generates a malignancy risk quantification index. The projection correction module includes: The projection deviation quantization submodule performs a dot product operation on the original image feature vector and the geometric constraint normal vector to extract the projection intensity of the feature on the normal constraint axis. At the same time, it analyzes the Euclidean norm of the geometric constraint normal vector, divides the dot product result by the Euclidean norm, performs numerical normalization processing, and generates the feature deviation distance value. The tolerance discrimination decision submodule calls the feature deviation distance value and the geometric constraint normal vector, reads the preset tolerance threshold parameter, performs a numerical comparison, and determines whether the feature deviation distance value exceeds the allowable range of benign physiological fluctuations. If the determination result is out of bounds, the product of the deviation distance and the geometric constraint normal vector is calculated to construct the abnormal components to be removed. If the determination result is compliant, all corresponding correction terms are set to zero vectors to generate feature correction vector parameters. The physiological alignment execution submodule calls the feature correction vector parameters, performs a linear vector subtraction operation in the multidimensional feature space, removes the normal projection components that violate geometric constraints in the original features, retains the tangential feature components that conform to the benign subspace distribution law, performs numerical stability verification and dimensional consistency check on the corrected vector data, and generates physiological alignment feature vectors.

2. The cloud-based collaborative prostate cancer multimodal data analysis system according to claim 1, characterized in that, The PSA density benchmark value includes biochemical antigen concentration parameters and glandular volume space parameters; the geometric constraint normal vector includes spatial orientation guiding components and hyperplane normal coefficients; the original image feature vector includes pixel gray-level distribution statistics and lesion texture topology parameters; and the physiological alignment feature vector includes feature projection. The correction components and spatial geometric calibration terms, the malignancy risk quantification indicators include feature similarity probability values ​​and lesion carcinogenesis tendency scores.

3. The cloud-based collaborative prostate cancer multimodal data analysis system according to claim 1, characterized in that, The biochemical image acquisition module includes: The image biochemical data analysis submodule acquires T2-weighted magnetic resonance imaging data and tPSA concentration values, identifies the prostate capsule edge trajectory within the imaging data, traverses the internal region of the edge trajectory, accumulates the number of discrete pixels, and generates prostate capsule pixel statistics. The gland volume space calculation submodule calls the pixel statistics of the prostate capsule, introduces the scanning layer thickness parameter, layer spacing parameter and single pixel plane physical area, and combines the morphological calibration factor and boundary convergence coefficient to identify the three-dimensional volume information of the gland after spatial geometric correction, and calculates the total volume value of the prostate. The biochemical density benchmark mapping submodule calls the total prostate volume value and tPSA concentration value to construct the ratio relationship between antigen mass and gland volume, quantifies the distribution density of biochemical indicators per unit volume, and generates a PSA density benchmark value.

4. The cloud-based collaborative prostate cancer multimodal data analysis system according to claim 3, characterized in that, The process for obtaining the total prostate volume value is as follows: The pixel statistics of the prostate capsule and the physical area of ​​a single pixel are obtained. The physical area conversion of the two in the planar dimension is performed to extract the true physical area features at the slice level. Then, the physical area features are combined with the scanning slice thickness parameters and the interslice spacing parameters to perform interslice expansion in the depth direction and three-dimensional geometric space integration to construct a preliminary organ volume assessment. Finally, for this preliminary organ volume assessment, a preset morphological calibration factor is introduced to perform three-dimensional morphological distortion compensation processing, and the boundary convergence coefficient is combined simultaneously to perform numerical approximation and error convergence correction calculations on the edge contour. Finally, the total prostate volume value that accurately represents the three-dimensional spatial distribution of the target organ is obtained.

5. The cloud-based collaborative prostate cancer multimodal data analysis system according to claim 3, characterized in that, The constrained hyperplane space construction module includes: The feature space benchmark definition submodule filters the benign sample set based on the PSA density benchmark value, calculates the distribution centroid of the benign sample set in the multidimensional feature space, sets the origin of the coordinate system, plans and identifies the principal component directions of the feature distribution, and generates the feature space orientation axis parameters. The benign subspace construction submodule calls the feature space orientation axis parameters, performs spatial orthogonal decomposition operation, constructs a geometric plane perpendicular to the principal component direction, defines the projection coverage of benign samples on the geometric plane, and generates the boundary of the benign geometric subspace. The constraint normal generation submodule calls the PSA density reference value, feature space orientation axis parameters, and benign geometric subspace boundary limits to analyze the degree of deviation between antigen density and benign distribution, calculate the hyperplane normal deflection weight, and combine the orientation axis parameters to perform vector synthesis to generate the geometric constraint normal vector.

6. The cloud-based collaborative prostate cancer multimodal data analysis system according to claim 5, characterized in that, The lesion texture feature extraction module includes: The scanning area guidance and definition submodule loads T2-weighted magnetic resonance imaging data according to the geometric constraint normal vector, establishes a scanning index along the spatial gradient direction indicated by the geometric constraint normal vector, identifies the tissue signal abrupt change edge in the cross section orthogonal to the geometric constraint normal vector, identifies the lesion voxel set through gradient determination, and generates the spatial boundary of the lesion area. The grayscale texture feature extraction submodule calls the spatial boundary of the lesion region, extracts the brightness intensity value of the pixel, calculates the second moment statistics of the grayscale histogram, quantifies the tissue density dispersion, constructs the grayscale co-occurrence matrix to analyze the topological structure of the pixel spatial arrangement, calculates and obtains the energy spectrum, contrast and inverse moment attributes, and generates grayscale dispersion and texture rule parameters. The feature vector multidimensional aggregation submodule performs numerical standardization operations on the gray-level statistical values ​​and texture topology attributes of the differentiated physical dimensions for the gray-level dispersion and texture rule parameters, eliminates the differences in the numerical distribution scale between feature components, and performs data serial stitching operation according to the preset feature dimension order to generate the original image feature vector.

7. The cloud-based collaborative prostate cancer multimodal data analysis system according to claim 1, characterized in that, The risk assessment module includes: The prototype vector retrieval submodule establishes a data transmission channel with the cloud-based medical database, traverses the stored image feature data of confirmed cases, identifies the spatial centroid of malignant sample groups, extracts a set of multidimensional parameters characterizing the core pathological attributes of prostate cancer, and generates prototype vectors of typical carcinogenesis features. The similarity calculation submodule calls the physiological alignment feature vector and the typical cancer feature prototype vector, performs the dot product operation and modulus calculation operation of the multidimensional space vector, solves the cosine value of the angle between the two feature vectors, quantifies the consistency between the feature distribution of the test sample and the malignant prototype in the geometric direction, and generates the feature distribution overlap coefficient. The risk index generation submodule loads a preset risk classification mapping table for the feature distribution overlap coefficient, converts the cosine similarity value into a standardized probability score, and performs weighted aggregation processing on the contribution weight of the differentiated feature dimension to the malignancy judgment, calculates the quantitative value representing the occurrence of malignant lesions in the tissue under test, and generates a malignancy risk quantitative index.

Images