A prostate cancer early screening model and a method for constructing the same

Abstract

The application relates to the technical field of medical image intelligent analysis, and discloses a prostate cancer early screening model and a construction method thereof. The method receives original image data from multiple sources in parallel, and obtains a candidate image set after quality evaluation and screening. The candidate image is subjected to layer-by-layer structure decomposition, and a background layer, a tissue layer and a blood vessel structure layer are separated. Structure features of the tissue layer and the blood vessel structure layer are extracted, and are compared with corresponding standard tissue feature libraries and standard blood vessel feature libraries, and a primary risk feature group is formed according to the differences. The feature group is subjected to time sequence simulation evolution, a simulation feature state at a future time point is generated, and is coupled with a current state to construct a dynamic risk feature map. The dynamic risk feature map is used to search for a path with the highest matching degree in a preset risk screening path, and a final risk decision mark is generated. The application realizes decoupled analysis of image components and dynamic risk evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis technology, specifically to an early prostate cancer screening model and its construction method. Background Technology

[0002] Currently, early prostate cancer screening based on medical images mainly relies on the overall analysis of images by artificial intelligence models. Common technical solutions include directly inputting the entire medical image into a deep convolutional neural network for end-to-end classification, or first identifying suspected lesion areas through a segmentation network, and then extracting features from those areas for classification decisions. These methods treat complex images containing background, normal tissue, blood vessels, and potential lesions as a whole or a region of interest, and the extracted features are a mixture of various anatomical structures and signal information. Features of different tissue structures interfere with and mask each other in the model, especially the information of minute vascular structures, which is easily diluted by more significant tissue morphological features. This results in limited ability of the model to identify early deviations in tissue texture that depend on subtle changes in vascular morphology, affecting the sensitivity and specificity of screening.

[0003] Existing screening models are essentially static analyses, with risk assessments based entirely on imaging characteristics at a single point in time. These models can only answer the question, "Does the current image show signs of risk?", but cannot judge the dynamic development trend of those signs. Early malignant lesions or high-risk conditions in clinical practice may have characteristics overlapping with benign conditions at a single point in time, but their evolutionary trajectories may show different trends. Static models cannot capture and quantify this evolutionary potential, making it difficult to identify cases with atypical current characteristics but rapidly evolving towards malignancy, thus limiting the timeliness of screening and early warning. Summary of the Invention

[0004] The purpose of this invention is to provide an early screening model for prostate cancer and a method for constructing the model, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for constructing an early prostate cancer screening model, the method comprising:

[0006] Raw image data of the subjects to be screened are received in parallel from multiple independent medical data sources, and the received raw image data is input into the image quality assessment network.

[0007] A quality score is generated for the original image data based on the image quality assessment network. The original image data is then filtered based on the quality score, and a candidate image set is extracted from the filtered original image data.

[0008] Each image in the candidate image set is subjected to layer-by-layer structural decomposition to identify and separate the background layer, tissue layer and vascular structure layer in the image;

[0009] Structural features were extracted for the tissue layer and the vascular structure layer respectively. The structural features of the tissue layer were aligned and compared with the standard tissue feature library, and the structural features of the vascular structure layer were compared with the standard vascular feature library. Based on the differences between the two comparisons, a primary risk feature group was formed.

[0010] The features in the primary risk feature group are simulated and evolved over time to generate simulated feature states at multiple future time points. The simulated feature states are then coupled with the current feature states, and a dynamic risk feature map is constructed based on the coupling results.

[0011] Based on the dynamic risk feature map, a path search is performed among multiple preset risk screening paths. The target risk screening path with the highest matching degree with the dynamic risk feature map is selected, and the final risk decision label of the object to be screened is generated according to the target risk screening path.

[0012] Preferably, the step of receiving raw image data of the subjects to be screened from multiple independent medical data sources in parallel and inputting the received raw image data into the image quality assessment network includes:

[0013] The multiple independent medical data sources include medical image archiving and communication systems, digital pathology slide databases, and wearable medical device data streams;

[0014] Each medical data source is configured with an independent data receiving interface. Raw image data is acquired in parallel through each data receiving interface, and the acquired raw image data is uniformly packaged into an initial data packet in a preset format.

[0015] The image quality assessment network is started. The image quality assessment network includes parallel sharpness analysis branches, noise level analysis branches, and integrity analysis branches. The original image data in the initial data packet is simultaneously input into the sharpness analysis branch, noise level analysis branch, and integrity analysis branch.

[0016] The sharpness analysis branch outputs the image sharpness value, the noise level analysis branch outputs the image noise level value, and the integrity analysis branch outputs the image structural integrity value.

[0017] The image sharpness value, image noise level value, and image structure integrity value are fused and calculated to obtain the quality score of the original image data.

[0018] Preferably, the step of filtering the original image data based on quality scores and extracting a candidate image set from the filtered original image data includes:

[0019] A preset quality score threshold is set, and the quality score of the original image data is compared with the preset quality score threshold one by one;

[0020] If the quality score of a certain original image data is greater than or equal to the quality score threshold, the original image data is marked as a qualified image and stored in the intermediate image set.

[0021] If the quality score of a certain original image data is less than the quality score threshold, the original image data is marked as a low-quality image and stored in the recycling queue.

[0022] Redundancy removal is performed on all qualified images in the intermediate image set. The redundancy removal process includes: calculating the structural similarity between qualified images; if the structural similarity exceeds a preset redundancy threshold, the qualified images with lower quality scores are removed, and the qualified images with the highest quality scores are retained.

[0023] The remaining qualified images after redundancy removal are determined as the candidate image set.

[0024] Preferably, the step of performing layer-by-layer structural decomposition on each image in the candidate image set to identify and separate the background layer, tissue layer, and vascular structure layer in the image includes:

[0025] For each image in the candidate image set, apply a multi-scale edge detection operator to generate an edge intensity map of the image;

[0026] Based on the edge intensity map, a method combining region growing and morphological closing operation is used to initially divide the foreground region and background region in the image, and the background region constitutes the background layer.

[0027] Within the foreground region, a pixel classifier based on texture features is used to distinguish pixel regions with glandular tissue structure textures from pixel regions with tubular or linear textures.

[0028] Pixel regions with glandular tissue structure textures are labeled as tissue layers, and pixel regions with tubular and linear textures are labeled as vascular structure layers.

[0029] The separated tissue layer and vascular structure layer are respectively subjected to cavity filling and smoothing processes to generate the final separated background layer image, tissue layer image and vascular structure layer image.

[0030] Preferably, the step of extracting structural features for the tissue layer and the vascular structure layer respectively, aligning and comparing the structural features of the tissue layer with a standard tissue feature library, and comparing the structural features of the vascular structure layer with a standard vascular feature library, and forming a primary risk feature group based on the differences between the two comparisons, including:

[0031] The morphological features of glandular structures, the regularity features of glandular boundaries, and the texture uniformity features inside glands are extracted from tissue layer images to form tissue feature vectors.

[0032] The complexity features of vascular branches, the disorder features of vascular orientation, and the distribution features of vascular density are extracted from the vascular structure layer image to form a vascular feature vector.

[0033] Retrieve standard organizational feature vectors from the standard organizational feature library that match the group characteristics of the target group, and calculate the cosine similarity between the organizational feature vector and the standard organizational feature vector as the organizational difference.

[0034] Retrieve standard vascular feature vectors that match the population characteristics of the subjects to be screened from the standard vascular feature database, and calculate the Euclidean distance between the vascular feature vectors and the standard vascular feature vectors as the vascular difference degree;

[0035] The tissue difference, vascular difference, tissue feature vector, and vascular feature vector are concatenated and combined to form the primary risk feature group.

[0036] Preferably, the step of performing time-series simulation evolution on the features in the primary risk feature group to generate simulated feature states at multiple future time points, coupling the simulated feature states with the current feature states, and constructing a dynamic risk feature map based on the coupling result includes:

[0037] A time-series simulation network is constructed, which takes a primary risk feature set as input.

[0038] Using a time-series simulation network, the evolution of the primary risk feature group at multiple preset future time points is simulated, generating a simulated feature vector corresponding to each future time point. Each simulated feature vector contains simulated tissue features and simulated vascular features at the future time point.

[0039] Extract current tissue and vascular features at the current time point from the primary risk feature set;

[0040] The simulated organizational features at each future time point are fused with the current organizational features to generate the fused organizational features at each future time point.

[0041] The simulated vascular features at each future time point are fused with the current vascular features to generate fused vascular features at each future time point.

[0042] All fused tissue features and fused vascular features at future time points are spatially arranged according to their corresponding time order. In the arranged two-dimensional space, the feature intensity value is used as the pixel gray value to generate a two-dimensional image, which is the dynamic risk feature map.

[0043] Preferably, the step of searching for a path among multiple preset risk screening paths based on a dynamic risk feature map and selecting the target risk screening path with the highest matching degree to the dynamic risk feature map includes:

[0044] Multiple risk screening paths are preset, and each risk screening path is defined as an ordered sequence of a series of feature nodes and their state transition probabilities;

[0045] Extract key feature node sequences from dynamic risk feature maps;

[0046] The key feature node sequence is dynamically time-warped with the feature node sequence in each risk screening path to obtain the path matching degree between the key feature node sequence and each risk screening path.

[0047] Select the risk screening path with the highest path matching degree and determine it as the target risk screening path.

[0048] Preferably, the step of generating the final risk decision marker for the object to be screened based on the target risk screening path includes:

[0049] Obtain the endpoint node of the target risk screening path, and the endpoint node is associated with a risk level label;

[0050] Based on the risk level labels associated with the endpoint nodes, generate initial risk decision labels for the objects to be screened;

[0051] Obtain node attribute information for several preset nodes before the endpoint node on the target risk screening path;

[0052] The confidence level of the initial risk decision label is corrected based on the node attribute information to generate the corrected final risk decision label.

[0053] Preferably, the step of correcting the confidence level of the initial risk decision marker based on the node attribute information includes:

[0054] The node attribute information includes node risk contribution weight and node state transition determinism;

[0055] The initial risk decision label is weighted and adjusted using the node risk contribution weight to generate a weighted adjustment label;

[0056] The weighted adjustment label is fuzzified using the deterministic nature of node state transitions to generate a fuzzy label;

[0057] The obfuscated label is deobfuscated to generate the final risk decision label.

[0058] Preferably, the present invention also includes an early prostate cancer screening model, which is constructed by the method for constructing an early prostate cancer screening model as described above.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] By performing layer-by-layer structural decomposition on candidate images, independent tissue and vascular structures are separated and aligned with corresponding standard tissue and vascular feature libraries, respectively. This technique enables decoupled analysis of different biologically significant components within images. The separation and alignment avoids interference and confusion between tissue morphology and vascular network features within the model, allowing subtle abnormal changes in blood vessels to be independently and clearly identified and quantified. The comparison process with the standard feature library essentially measures individual features within a reference system of "normal range" or "typical pattern," enhancing the model's sensitivity to atypical but pathologically significant subtle feature patterns that deviate from normal patterns and reducing reliance on large-scale labeled data.

[0061] The extracted primary risk features are subjected to time-series simulation evolution, generating simulated feature states at multiple future time points. These simulated future states are then coupled with the current state to construct a dynamic risk feature map. The static feature space is expanded into a dynamic feature space that incorporates temporal evolution trends. Risk assessment no longer relies solely on the instantaneous values ​​of features but integrates their potential change paths and rates over a simulated period. This enables the model to identify features whose absolute values ​​at the current time point may not yet have reached the high-risk threshold, but whose evolutionary trajectory already exhibits dynamic characteristics of rapid progression towards malignancy. Through simulation and coupling, screening decisions incorporate forward-looking predictions of disease progression, enabling earlier warnings of potential risks on a malignant evolutionary trajectory. Attached Figure Description

[0062] Figure 1 This is a schematic diagram illustrating the working principle of the method for constructing an early prostate cancer screening model according to the present invention.

[0063] Figure 2 A flowchart for receiving raw image data and performing quality assessment;

[0064] Figure 3 This is a flowchart for selecting and constructing a candidate image set based on quality scores;

[0065] Figure 4 A bar chart showing the distribution of fused feature vectors at future time points in early prostate cancer screening;

[0066] Figure 5 A bar chart comparing the results of different unfuzzy resolution methods in prostate cancer risk decision-making. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Please see Figure 1 This invention provides a method for constructing an early prostate cancer screening model. The method includes: receiving raw image data of the subject to be screened in parallel from multiple independent medical data sources, and inputting the received raw image data into an image quality assessment network. Based on the image quality assessment network, a quality score is generated for the raw image data; the raw image data is then filtered according to the quality score, and a candidate image set is extracted from the filtered raw image data. Each image in the candidate image set undergoes layer-by-layer structural decomposition to identify and separate the background layer, tissue layer, and vascular structure layer. Structural features are extracted for the tissue layer and vascular structure layer respectively; the structural features of the tissue layer are aligned and compared with a standard tissue feature library, and the structural features of the vascular structure layer are compared with a standard vascular feature library. A primary risk feature group is formed based on the differences between the two comparisons. The features in the primary risk feature group undergo temporal simulation evolution to generate simulated feature states at multiple future time points, and the simulated feature states are coupled with the current feature states. A dynamic risk feature map is constructed based on the coupling result. Based on the dynamic risk feature map, a path search is performed among multiple preset risk screening paths. The target risk screening path with the highest matching degree with the dynamic risk feature map is selected, and the final risk decision label of the object to be screened is generated according to the target risk screening path.

[0069] In one embodiment of the present invention, see [reference] Figure 2 Multiple independent medical data sources are used, including medical image archiving and communication systems, digital pathology slide databases, and wearable medical device data streams. Each medical data source is configured with an independent data receiving interface, through which raw image data is acquired in parallel. The acquired raw image data is then uniformly packaged into an initial data packet of a preset format. An image quality assessment network is activated, which includes parallel sharpness analysis, noise level analysis, and integrity analysis branches. The raw image data from the initial data packet is simultaneously input into these branches. The sharpness analysis branch outputs an image sharpness value, the noise level analysis branch outputs an image noise level value, and the integrity analysis branch outputs an image structural integrity value. The image sharpness value, image noise level value, and image structural integrity value are then fused and calculated to obtain a quality score for the raw image data.

[0070] In the implementation, multiple independent medical data sources are configured, including a medical image archiving and communication system, a digital pathology slide database, and a wearable medical device data stream. Specifically, a separate data receiving interface is configured for the medical image archiving and communication system, another for the digital pathology slide database, and a third for the wearable medical device data stream. Raw image data is acquired in parallel through these interfaces. Since the raw image data acquired from different data sources differs in format and encapsulation, the acquired raw image data is uniformly encapsulated into an initial data packet with a preset format. This preset format can be a standardized structure containing image matrix data, acquisition timestamps, and a data source identifier.

[0071] In some embodiments, the image quality assessment network is launched immediately after encapsulation. The image quality assessment network comprises three parallel analysis branches: a sharpness analysis branch, a noise level analysis branch, and a structural integrity analysis branch. The original image data from the initial data package is simultaneously copied and input to these three branches. Specifically, the sharpness analysis branch calculates the edge gradient magnitude on the input image data, the noise level analysis branch calculates the standard deviation within smooth regions on the input image data, and the structural integrity analysis branch performs contour integrity detection based on prior anatomical structures on the input image data. The sharpness analysis branch outputs a quantified image sharpness value, the noise level analysis branch outputs a quantified image noise level value, and the structural integrity analysis branch outputs a quantified image structural integrity value.

[0072] In some embodiments, image sharpness, image noise level, and image structural integrity are fused to obtain a quality score for a single representative original image data. In specific implementations, the fusion calculation employs a weighted linear combination method. The quality score fusion formula can be understood as follows:

[0073] ;

[0074] in: Represents the final quality score. Represents image sharpness value, Represents the image noise level value. Represents the image structure integrity value. , and These are the positive weighting coefficients assigned to the sharpness value, noise level value, and structural integrity value, respectively, and the sum of these weighting coefficients is one. The sharpness value contributes a positive portion, the noise level value contributes a negative portion, and the structural integrity value contributes a positive portion. Optional, the weighting coefficients... , and This can be determined by performing a grid search on a historical labeled dataset. Alternatively, in another implementation, the fusion computation can also employ a multilayer perceptron model to map the three input values ​​to a single score. In an alternative implementation using a multilayer perceptron model for fusion computation, the multilayer perceptron model can be constructed as a feedforward neural network structure containing an input layer, hidden layers, and an output layer. The input layer has three neurons that receive image sharpness, image noise level, and image structural integrity as input features, respectively. The hidden layer can contain several neurons, introducing a nonlinear transformation through an activation function to capture the complex interactions between the three input features. The output layer is designed as a single neuron, whose output value is processed by a linear or sigmoid function and mapped to a normalized quality score ranging from 0 to 1. During model training, the network weights and bias parameters are optimized using a backpropagation algorithm on a historical labeled dataset to minimize the mean squared error or cross-entropy loss between the model's predicted quality score and the true quality label.

[0075] In one embodiment of the present invention, see [reference] Figure 3A preset quality score threshold is set, and the quality score of each original image data is compared with this threshold. If the quality score of an original image data is greater than or equal to the quality score threshold, it is marked as a qualified image and stored in the intermediate image set. If the quality score of an original image data is less than the quality score threshold, it is marked as a low-quality image and stored in the recycling queue. Redundancy removal is performed on all qualified images in the intermediate image set. This redundancy removal process includes calculating the structural similarity between qualified images. If the structural similarity exceeds a preset redundancy threshold, qualified images with lower quality scores are removed, and qualified images with the highest quality scores are retained. The remaining qualified images after redundancy removal are determined as a candidate image set. For each image in the candidate image set, a multi-scale edge detection operator is applied to generate an edge intensity map. Based on the edge intensity map, a method combining region growing and morphological closing operations is used to initially divide the foreground and background regions in the image. This background region constitutes the background layer. Within the foreground region, a pixel classifier based on texture features is used to distinguish pixel regions with glandular tissue structure textures and pixel regions with tubular or linear textures. Pixel regions with glandular tissue structure textures are labeled as tissue layers, and pixel regions with tubular and linear textures are labeled as vascular structure layers. The separated tissue and vascular structure layers are then subjected to hole filling and smoothing processes to generate the final separated background, tissue, and vascular structure layer images.

[0076] In practice, a fixed quality score threshold is preset, and the quality score of each piece of raw image data is compared with this threshold one by one. In some embodiments, if the quality score of a piece of raw image data is greater than or equal to the preset quality score threshold, the raw image data is marked as a qualified image and stored in a temporary storage structure called the intermediate image set. If the quality score of a piece of raw image data is less than the preset quality score threshold, the raw image data is marked as a low-quality image and stored in a separate retrieval queue for subsequent reprocessing or analysis.

[0077] In specific implementations, all qualified images in the intermediate image set undergo redundancy removal processing. The core of this redundancy removal process is calculating the structural similarity between qualified images. In some embodiments, the structural similarity is calculated in the grayscale space of the image. For any two qualified images in the intermediate image set, their structural similarity index is calculated. If the calculated structural similarity exceeds a preset redundancy threshold, the two images are considered highly similar and constitute redundancy. In specific implementations, at this point, qualified images with lower quality scores need to be removed, and the qualified image with the highest quality score is retained. In specific implementations, each image in the candidate image set undergoes layer-by-layer structural decomposition to identify and separate the background layer, tissue layer, and vascular structure layer. For each image in the candidate image set, a multi-scale edge detection operator is applied. The multi-scale edge detection operator uses Gaussian kernels with different standard deviations for filtering and calculates gradient magnitudes to generate an edge intensity map that comprehensively reflects the edge information of the image. Based on the generated edge intensity map, a method combining region growing and morphological closing operation is used to divide the region. The seed point of region growing is selected from the connected region with the lowest gradient value in the edge intensity map. Morphological closing operation is used to fill the small holes and gaps inside the foreground region after region growing, thereby initially dividing the foreground region and background region in the image. The set of all pixels marked as background region constitutes the background layer.

[0078] Within the foreground region, a pixel classifier analyzes each pixel or pixel block within the foreground region, outputting the probability of it belonging to different texture categories. Pixel regions with glandular tissue structure texture features are labeled as tissue layers, characterized by regular and repetitive alveolar or tubular texture patterns. Pixel regions with tubular or linear texture features are labeled as vascular structure layers, characterized by elongated, branching linear structures. It can be understood that semantic separation of tissue and vascular structures is achieved through texture classification. The separated tissue and vascular structure layer images are then subjected to morphological hole filling and smoothing based on anisotropic diffusion equations, respectively. Hole filling is used to fill in small regions missing due to classification errors, while smoothing suppresses noise while preserving the main structural boundaries, ultimately generating clearly separated background, tissue, and vascular structure layer images. Optionally, smoothing can use anisotropic diffusion filters based on partial differential equations. In specific implementations, structural similarity calculation can be quantified using the following formula:

[0079] ;

[0080] in: Representative image With images The structural similarity index between them and Representing images and images The average pixel value, and Representing images and images The standard deviation of pixels, Representative image and images covariance, and This is a small constant introduced to stabilize division operations. Redundancy threshold. It is a preset value between 0 and 1, when At that time, the two images were determined to be redundant.

[0081] In one embodiment of the present invention, morphological features of glandular structures, regularity features of glandular boundaries, and texture uniformity features within glands are extracted from tissue layer images to form tissue feature vectors. Complexity features of vascular branches, disorder features of vascular orientation, and distribution features of vascular density are extracted from vascular structure layer images to form vascular feature vectors. Standard tissue feature vectors matching the group characteristics of the target population are retrieved from a standard tissue feature library, and the cosine similarity between the tissue feature vector and the standard tissue feature vector is calculated as the tissue difference. Standard vascular feature vectors matching the group characteristics of the target population are retrieved from a standard vascular feature library, and the Euclidean distance between the vascular feature vector and the standard vascular feature vector is calculated as the vascular difference. The tissue difference, vascular difference, tissue feature vectors, and vascular feature vectors are concatenated and combined to form the primary risk feature group.

[0082] In the specific implementation, structural features are extracted for the tissue layer and the vascular structure layer respectively. The structural features of the tissue layer are aligned and compared with the standard tissue feature library, and the structural features of the vascular structure layer are compared with the standard vascular feature library. The steps of forming a primary risk feature group based on the difference results of the two comparisons are as follows: morphological features of glandular structures, regularity features of glandular boundaries, and texture uniformity features inside the gland are extracted from the separated tissue layer images. In the specific implementation, the morphological features of glandular structures can be characterized by calculating the area, perimeter, compactness, and Fourier descriptor of the segmented glandular region. The regularity features of glandular boundaries are quantified by calculating the standard deviation of curvature and fractal dimension of the glandular contour. The texture uniformity features inside the gland are obtained by calculating the entropy and contrast of the gray-level co-occurrence matrix within the glandular region. These calculated feature values ​​together constitute a multidimensional tissue feature vector. The complexity features of vascular branches, the disorder features of vascular orientation, and the distribution features of vascular density are extracted from the separated vascular structure layer image. In specific implementation, the complexity features of vascular branches are obtained by counting branch points and measuring the average branch length on the skeletonized vascular structure map. The disorder features of vascular orientation are evaluated by calculating the standard deviation of the orientation angle of vascular segments and the orientation autocorrelation length. The distribution features of vascular density are described by dividing the image into grids and calculating the statistical distribution of the proportion of vascular pixels in each grid cell. The feature values ​​obtained from these calculations together constitute a multidimensional vascular feature vector.

[0083] In some embodiments, the standard tissue feature library uses age and prostate-specific antigen (PSA) level as index keys to retrieve a set of standard tissue feature vectors matching the current object. This set of vectors represents the typical tissue characteristics of a healthy population at that age and PSA level. The cosine similarity between the tissue feature vector extracted from the tissue layer image and the retrieved standard tissue feature vector is calculated, and this cosine similarity value is used as the tissue dissimilarity. Standard vascular feature vectors matching the same population characteristics of the object to be screened are retrieved from the standard vascular feature library. The Euclidean distance between the vascular feature vector extracted from the vascular structure layer image and the retrieved standard vascular feature vector is calculated, and this Euclidean distance value is used as the vascular dissimilarity.

[0084] In specific implementations, the calculated tissue difference, vascular difference, and the original tissue feature vector and vascular feature vector are concatenated to form a primary risk feature set. In some embodiments, the concatenation operation involves joining the tissue difference scalar, the vascular difference scalar, all dimensional elements of the tissue feature vector, and all dimensional elements of the vascular feature vector in a predefined order to form a longer one-dimensional feature array, which is the primary risk feature set. Optionally, the concatenation order can be: first place the tissue difference, then the vascular difference, then all elements of the tissue feature vector, and finally all elements of the vascular feature vector. Optionally, in another implementation, the tissue feature vector and vascular feature vector can be dimensionality reduced first, and then the dimensionality-reduced features can be concatenated with the two difference scalars. In specific implementations, the cosine similarity of the tissue difference can be expressed using the formula:

[0085] ;

[0086] in: Represents organizational differences. This represents the tissue feature vector extracted from the current tissue layer image. This represents the standard organization feature vector retrieved from the standard organization feature library. The dimension representing the organizational feature vector. and These represent the first two eigenvectors respectively. Each dimension component.

[0087] In one embodiment of the present invention, a temporal simulation network is constructed, which takes a primary risk feature set as input. Using the temporal simulation network, the evolution of the primary risk feature set at multiple preset future time points is simulated, generating a simulated feature vector corresponding to each future time point. Each simulated feature vector contains simulated tissue features and simulated vascular features for the future time point. Current tissue features and current vascular features for the current time point are extracted from the primary risk feature set. The simulated tissue features of each future time point are fused with the current tissue features to generate a fused tissue feature for each future time point. The simulated vascular features of each future time point are fused with the current vascular features to generate a fused vascular feature for each future time point. All fused tissue features and fused vascular features for the future time points are spatially arranged according to their corresponding time order, and a two-dimensional image is generated in the arranged two-dimensional space, using the feature intensity value as the pixel grayscale value. This two-dimensional image is the dynamic risk feature map. Multiple risk screening paths are preset, each risk screening path being defined as an ordered sequence of a series of feature nodes and their state transition probabilities. A sequence of key feature nodes is extracted from the dynamic risk feature map. The key feature node sequence is dynamically time-warped with the feature node sequence in each risk screening path to obtain the path matching degree between the key feature node sequence and each risk screening path. The risk screening path with the highest path matching degree is selected as the target risk screening path.

[0088] In specific implementation, the steps involve performing time-series simulation evolution of the features in the primary risk feature group to generate simulated feature states at multiple future time points, coupling the simulated feature states with the current feature states, and constructing a dynamic risk feature map based on the coupling results. The specific execution flow is as follows: First, a time-series simulation network is constructed. This network takes the primary risk feature group as input and employs a recurrent neural network or temporal convolutional network structure. This network has been trained on a historical dataset containing prostate image sequences at multiple time points and is able to learn the mapping pattern from the current feature state to future feature states. Second, the time-series simulation network is used to simulate the evolution of the primary risk feature group at multiple preset future time points. These preset future time points can include three months later, six months later, one year later, etc., generating simulated feature vectors corresponding to each future time point. Each simulated feature vector contains simulated tissue features and simulated vascular features for the corresponding future time point. Third, the current tissue features and current vascular features for the current time point are extracted from the primary risk feature group. These current tissue features and current vascular features are the concatenated tissue feature vector and vascular feature vector portions from the primary risk feature group.

[0089] In practice, the simulated tissue features at each future time point are fused with the current tissue features. This feature fusion operation can be vector concatenation, weighted summation, or performed through a small fusion neural network to generate fused tissue features for each future time point. Similarly, the simulated vascular features at each future time point are fused with the current vascular features to generate fused vascular features for each future time point. This fusion operation ensures that each simulated feature at a future time point incorporates contextual information about the current state. In practice, all fused tissue and fused vascular features from future time points are spatially arranged according to their corresponding chronological order, typically linearly from most recent to oldest. Within this arranged two-dimensional space, the intensity value of each fused feature is used as the pixel grayscale value to generate a two-dimensional image, which is the dynamic risk feature map. In some embodiments, assuming there are three future time points, and both the fused tissue features and fused vascular features are three-dimensional vectors, the fused tissue features at the three time points can be arranged in chronological order in one row, and the fused vascular features at the three time points can be arranged in the next row in the same chronological order, thus forming a 2x9 image matrix. The value of each element in the matrix is ​​normalized and mapped to a grayscale pixel value. See Table 1.

[0090] Table 1: Fusion Features at Future Time Points

[0091] Time point Fusion of organizational feature vectors (3 dimensions) Fusion of vascular feature vectors (3 dimensions) t+1 (3 months later) [0.85,1.20,-0.33] [0.62,0.91,0.15] t+2 (6 months later) [0.92,1.35,-0.41] [0.58,1.03,0.22] t+3 (12 months later) [1.10,1.50,-0.52] [0.51,1.15,0.30]

[0092] In some embodiments, based on a dynamic risk feature map, path searching is performed among multiple preset risk screening paths to select the target risk screening path with the highest matching degree to the dynamic risk feature map. Multiple risk screening paths are preset, each defined as an ordered sequence of feature nodes and their state transition probabilities, where each feature node corresponds to a feature pattern or state. Key feature node sequences are extracted from the dynamic risk feature map using peak detection or feature clustering methods. These sequences reflect the trajectory of feature evolution over time. The extracted key feature node sequences are then dynamically time-warped with the feature node sequences in each risk screening path. This dynamic time-warping calculation aligns two sequences of different lengths and calculates their minimum cumulative distance. The path matching degree between the key feature node sequences and each risk screening path is obtained. The risk screening path with the highest path matching degree is selected as the target risk screening path. In specific implementations, the path matching degree... It can be calculated using the formula:

[0093] ;

[0094] in: Represents path matching degree. This represents the sequence of key feature nodes calculated using the dynamic time warping algorithm. Feature node sequence of the preset risk screening path The minimum cumulative alignment distance between them.

[0095] See Figure 4 This is a bar chart showing the distribution of fused feature vectors at future time points in early prostate cancer screening. It's a multi-dimensional grouped bar chart, displaying the vector values ​​of each dimension of tissue and vascular features at three future time points (3 / 6 / 12 months later), corresponding to the multi-dimensional feature distribution analysis at the feature fusion stage. The core dimension (d2) of both tissue and vascular features shows an increasing trend over time, reflecting the evolution direction of risk features. It visually presents the changes in risk features at different time points, providing fundamental data support for constructing dynamic risk feature maps and selecting risk screening paths, helping to assess the temporal evolution trend of prostate cancer risk. In early prostate cancer screening models, this type of chart is a key tool for quantifying the evolution of risk features, clearly showing the changing patterns of each dimension of features, and providing multi-dimensional basis for subsequent risk decisions.

[0096] In one embodiment of the present invention, the endpoint node of the target risk screening path is obtained, and this endpoint node is associated with a risk level label. Based on the risk level label associated with the endpoint node, an initial risk decision label for the object to be screened is generated. Node attribute information of several preset nodes preceding the endpoint node on the target risk screening path is obtained, including node risk contribution weights and node state transition determinism. The initial risk decision label is weighted and adjusted using the node risk contribution weights to generate a weighted adjustment label. The weighted adjustment label is then fuzzified using the node state transition determinism to generate a fuzzy label. Finally, a defuzzification decision is made on the fuzzy label to generate the final risk decision label.

[0097] In specific implementation, the step of generating the final risk decision label for the object to be screened based on the target risk screening path is implemented as follows: Obtain the endpoint node of the target risk screening path. Each endpoint node is associated with a predefined risk level label, which can be "low risk," "medium risk," "high risk," or "medium-high risk." Based on the risk level label associated with the endpoint node, generate the initial risk decision label for the object to be screened. In some embodiments, the initial risk decision label can be a discrete symbol corresponding to the risk level label or an initial risk probability value.

[0098] In practical implementation, node attribute information of several preset nodes before the endpoint on the target risk screening path is obtained. The number of preset nodes can be two or three. The node attribute information includes node risk contribution weight and node state transition determinism. The node risk contribution weight represents the relative influence of the node's characteristics on the final risk outcome, and the node state transition determinism represents the confidence level of the probability of transitioning from that node to the next node. In some embodiments, this node attribute information has been defined and stored through historical data statistics or expert knowledge when constructing the risk screening path. The initial risk decision marker is weighted and adjusted using the node risk contribution weight to generate a weighted adjustment marker. The weighted adjustment can reflect the influence of key intermediate nodes on the path, rather than relying solely on the endpoint node.

[0099] In practical implementation, the deterministic nature of node state transitions is used to fuzzify the weighted adjustment label, generating a fuzzified label. This fuzzification introduces the uncertainty inherent in the deterministic nature of node state transitions into the decision label. The fuzzified label is then defuzzified to generate the final risk decision label. Defuzzification can employ defuzzification methods such as the centroid method or the maximum membership method to convert the fuzzy quantity into a deterministic decision output. Optionally, the final risk decision label can be a corrected risk level label. Alternatively, the final risk decision label can also be a corrected risk probability value ranging from 0 to 1. The formula exemplarily describes one method for calculating weighted adjustment and defuzzification decisions:

[0100] ;

[0101] in: The numerical value representing the final risk decision marker. The numerical value representing the initial risk decision marker. This represents the preset number of nodes used for correction. Representing the Deterministic node state transitions of a set of preset nodes. Representative and the The number of node risk contribution weights associated with each node. Representing the Risk contribution weight of each node It is an exponential parameter used to adjust the first... The risk contribution weight of each node in the th... The strength of influence in the context of each node.

[0102] See Figure 5This is a bar chart comparing the results of different unambiguous methods in prostate cancer risk decision-making. It's a grouped bar chart showing the final risk decision values ​​for low, medium, and high-risk cases under four unambiguous methods, corresponding to the method effectiveness evaluation in the deambiguation stage. The decision values ​​for each risk category remain stable across different methods, indicating that the unambiguous methods have minimal impact on the differentiation of risk levels. This chart is used to evaluate the decision stability of different unambiguous methods, assisting in selecting the optimal deambiguity strategy for prostate cancer screening, and ensuring the reliability and consistency of risk decision results. In early prostate cancer screening models, this type of chart is a key tool for verifying the reliability of the decision correction stage, providing a direct comparison of the output differences between different methods and offering data support for method selection in clinical applications.

[0103] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing an early screening model for prostate cancer, characterized in that, The method includes: Raw image data of the subjects to be screened are received in parallel from multiple independent medical data sources, and the received raw image data is input into the image quality assessment network. A quality score is generated for the original image data based on the image quality assessment network. The original image data is then filtered based on the quality score, and a candidate image set is extracted from the filtered original image data. Each image in the candidate image set undergoes layer-by-layer structural decomposition to identify and separate the background layer, tissue layer, and vascular structure layer, including: For each image in the candidate image set, apply a multi-scale edge detection operator to generate an edge intensity map of the image; Based on the edge intensity map, a method combining region growing and morphological closing operation is used to initially divide the foreground region and background region in the image, and the background region constitutes the background layer. Within the foreground region, a pixel classifier based on texture features is used to distinguish pixel regions with glandular tissue structure textures from pixel regions with tubular or linear textures. Pixel regions with glandular tissue structure textures are labeled as tissue layers, and pixel regions with tubular and linear textures are labeled as vascular structure layers. The separated tissue layer and vascular structure layer are respectively filled with holes and smoothed to generate the final separated background layer image, tissue layer image and vascular structure layer image; Structural features were extracted for both the tissue layer and the vascular structure layer. The structural features of the tissue layer were aligned and compared with a standard tissue feature library, and the structural features of the vascular structure layer were compared with a standard vascular feature library. Based on the differences between the two comparisons, a primary risk feature group was formed, including: The morphological features of glandular structures, the regularity features of glandular boundaries, and the texture uniformity features inside glands are extracted from tissue layer images to form tissue feature vectors. The complexity features of vascular branches, the disorder features of vascular orientation, and the distribution features of vascular density are extracted from the vascular structure layer image to form a vascular feature vector. Retrieve standard organizational feature vectors from the standard organizational feature library that match the group characteristics of the target group, and calculate the cosine similarity between the organizational feature vector and the standard organizational feature vector as the organizational difference. Retrieve standard vascular feature vectors that match the population characteristics of the subjects to be screened from the standard vascular feature database, and calculate the Euclidean distance between the vascular feature vectors and the standard vascular feature vectors as the vascular difference degree; The tissue difference, vascular difference, tissue feature vector, and vascular feature vector are concatenated and combined to form the primary risk feature group; The features in the primary risk feature group undergo time-series simulation evolution to generate simulated feature states at multiple future time points. These simulated feature states are then coupled with the current feature states. Based on the coupling results, a dynamic risk feature graph is constructed, including: A time-series simulation network is constructed, which takes a primary risk feature set as input. Using a time-series simulation network, the evolution of the primary risk feature group at multiple preset future time points is simulated, generating a simulated feature vector corresponding to each future time point. Each simulated feature vector contains simulated tissue features and simulated vascular features at the future time point. Extract current tissue and vascular features at the current time point from the primary risk feature set; The simulated organizational features at each future time point are fused with the current organizational features to generate the fused organizational features at each future time point. The simulated vascular features at each future time point are fused with the current vascular features to generate fused vascular features at each future time point. All fused tissue features and fused vascular features at future time points are spatially arranged according to their corresponding time order. In the arranged two-dimensional space, the feature intensity value is used as the pixel gray value to generate a two-dimensional image, which is the dynamic risk feature map. Based on the dynamic risk feature map, a path search is performed among multiple preset risk screening paths. The target risk screening path with the highest matching degree with the dynamic risk feature map is selected, and the final risk decision label of the object to be screened is generated according to the target risk screening path.

2. The method for constructing an early prostate cancer screening model according to claim 1, characterized in that, The process of receiving raw image data of the subjects to be screened in parallel from multiple independent medical data sources and inputting the received raw image data into the image quality assessment network includes: The multiple independent medical data sources include medical image archiving and communication systems, digital pathology slide databases, and wearable medical device data streams; Each medical data source is configured with an independent data receiving interface. Raw image data is acquired in parallel through each data receiving interface, and the acquired raw image data is uniformly packaged into an initial data packet in a preset format. The image quality assessment network is started. The image quality assessment network includes parallel sharpness analysis branches, noise level analysis branches, and integrity analysis branches. The original image data in the initial data packet is simultaneously input into the sharpness analysis branch, noise level analysis branch, and integrity analysis branch. The sharpness analysis branch outputs the image sharpness value, the noise level analysis branch outputs the image noise level value, and the integrity analysis branch outputs the image structural integrity value. The image sharpness value, image noise level value, and image structure integrity value are fused and calculated to obtain the quality score of the original image data.

3. The method for constructing an early prostate cancer screening model according to claim 2, characterized in that, The process of filtering the original image data based on quality scores and extracting a candidate image set from the filtered original image data includes: A preset quality score threshold is set, and the quality score of the original image data is compared with the preset quality score threshold one by one; If the quality score of a certain original image data is greater than or equal to the quality score threshold, the original image data is marked as a qualified image and stored in the intermediate image set. If the quality score of a certain original image data is less than the quality score threshold, the original image data is marked as a low-quality image and stored in the recycling queue. Redundancy removal is performed on all qualified images in the intermediate image set. The redundancy removal process includes: calculating the structural similarity between qualified images; if the structural similarity exceeds a preset redundancy threshold, the qualified images with lower quality scores are removed, and the qualified images with the highest quality scores are retained. The remaining qualified images after redundancy removal are determined as the candidate image set.

4. The method for constructing an early prostate cancer screening model according to claim 3, characterized in that, The process, based on a dynamic risk feature map, involves searching for a path among multiple preset risk screening paths and selecting the target risk screening path with the highest matching degree to the dynamic risk feature map, including: Multiple risk screening paths are preset, and each risk screening path is defined as an ordered sequence of a series of feature nodes and their state transition probabilities; Extract key feature node sequences from dynamic risk feature maps; The key feature node sequence is dynamically time-warped with the feature node sequence in each risk screening path to obtain the path matching degree between the key feature node sequence and each risk screening path. Select the risk screening path with the highest path matching degree and determine it as the target risk screening path.

5. The method for constructing an early prostate cancer screening model according to claim 4, characterized in that, The process of generating the final risk decision marker for the object to be screened based on the target risk screening path includes: Obtain the endpoint node of the target risk screening path, and the endpoint node is associated with a risk level label; Based on the risk level labels associated with the endpoint nodes, generate initial risk decision labels for the objects to be screened; Obtain node attribute information for several preset nodes before the endpoint node on the target risk screening path; The confidence level of the initial risk decision label is corrected based on the node attribute information to generate the corrected final risk decision label.

6. The method for constructing an early prostate cancer screening model according to claim 5, characterized in that, The confidence correction of the initial risk decision label based on the node attribute information includes: The node attribute information includes node risk contribution weight and node state transition determinism; The initial risk decision label is weighted and adjusted using the node risk contribution weight to generate a weighted adjustment label; The weighted adjustment label is fuzzified using the deterministic nature of node state transitions to generate a fuzzy label; The obfuscated label is deobfuscated to generate the final risk decision label.

7. A model for early screening of prostate cancer, characterized in that, It is constructed using a method for constructing an early prostate cancer screening model as described in any one of claims 1 to 6.

Images