Hierarchical prior learning and constraint clustering based pancreatic tumor sub-region segmentation method

Abstract

The present application relates to the technical field of medical image segmentation, in particular to a pancreatic tumor sub-region segmentation method based on hierarchical prior learning and constraint clustering, which introduces expert-annotated sparse prior knowledge to guide the clustering process, forms a two-stage algorithm framework of hierarchical prior learning-constraint clustering, combines the advantages of data-driven and knowledge-driven, realizes that even if the expert annotation is sparse, the algorithm can also learn the distribution rule and generalize to the whole tumor for robust segmentation, and the constraint condition avoids the problems of meaningless small fragments or inconsistent class numbers, ensures the stability and consistency of the segmentation result on a large-scale data set, is a key prerequisite for subsequent quantitative calculation of tumor heterogeneity index and construction of accurate preoperative grading model, ensures that the extracted heterogeneity features truly reflect the biological nature of the tumor from the source, thereby significantly improving the performance of downstream tasks.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a method for pancreatic tumor subregion segmentation based on hierarchical prior learning and constrained clustering. Background Technology

[0002] The significant biological heterogeneity of pancreatic neuroendocrine tumors (PNENs) poses a challenge to precision treatment. Currently, the World Health Organization (WHO) classification based on the Ki-67 proliferation index is the gold standard for prognosis and treatment. However, this standard relies on invasive tissue biopsies, which are subject to problems such as sampling bias and inability to repeat monitoring, often leading to inaccurate preoperative classification and delaying treatment for high-risk patients.

[0003] Dual-phase enhanced computed tomography (CT) can non-invasively capture tumor angiogenesis and perfusion dynamics, providing a possibility for assessing tumor heterogeneity. However, existing medical image-based analysis methods have significant limitations: 1) Macroscopic analysis: Traditional radiomics methods usually extract "averaged" features of the entire tumor, masking key differences in different biological behavior regions within the tumor (such as high-proliferation areas and necrotic areas), which contradicts the pathological principle of assessing Ki-67 in "hotspot areas"; 2) Unsupervised segmentation: Clustering segmentation methods that rely entirely on image data (such as traditional K-means) can divide subregions, but the generated regions often lack clear clinical or biological interpretations, making it difficult for radiologists to effectively intervene or verify, resulting in low reliability and practicality; 3) Static description: Most studies only focus on the volume or proportion of subregions, failing to effectively quantify the dynamic changes in enhancement patterns (i.e., temporal features) of the same region during the arterial and venous phases, while this spatiotemporal heterogeneity is considered an important imaging characteristic of tumor biological invasiveness.

[0004] Therefore, current technologies cannot perform pancreatic tumor image subregion segmentation under non-invasive conditions that combines high biological interpretability, high clinical relevance, and effective quantification of intratumor spatiotemporal heterogeneity. Summary of the Invention

[0005] The purpose of this invention is to provide a pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering, so as to solve the technical problem that existing technologies cannot generate pancreatic tumor image subregion segmentation results with high biological interpretability, high clinical relevance, and effective quantification of intratumor spatiotemporal heterogeneity under non-invasive conditions.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A method for pancreatic tumor subregion segmentation based on hierarchical prior learning and constrained clustering includes the following steps: Step S1: Obtain dual-phase enhanced CT images of pancreatic tumor patients and obtain segmentation masks for the tumor region; Step S2: Preprocess the dual-phase enhanced CT images, including resampling the images into isotropic voxels and registering the venous phase images to the arterial phase images; Step S3: Extract the multidimensional feature vector of each voxel within the tumor region, and perform robust scaling and normalization on the multidimensional feature vector. The multidimensional feature vector combines spatial features and arteriovenous time features. Step S4: Perform hierarchical prior learning using sparse point labels annotated with expert experience, where the sparse point labels correspond to three enhancement categories that guide sub-region segmentation; Step S5: Integrate the prior knowledge obtained from hierarchical prior learning into the constrained clustering framework, cluster voxels within the tumor region, forcibly divide them into three sub-regions, and apply minimum cluster size constraints during the clustering process. Step S6: Post-process and quality verification of the clustering results, and output a three-dimensional sub-region segmentation label map.

[0007] As a preferred embodiment of the present invention, the dual-phase enhanced CT images include venous phase images and arterial phase images. The pancreas and tumor are segmented in the venous phase images and arterial phase images using a pre-trained nnU-Net model to obtain the segmentation results of the pancreatic region and the tumor region, and a segmentation mask for the tumor region is formed.

[0008] As a preferred embodiment of the present invention, the preprocessing resamples both the venous phase image and the arterial phase image. The voxels were calculated, and the venous and arterial phase images were cropped according to the pancreatic region segmentation results of the nnU-Net model into a shape centered on the pancreas and with edges of voxels. The bounding box is then used to precisely deform and register the venous phase image to the spatial coordinates of the arterial phase image, achieving alignment of the anatomical structures in the pancreatic and tumor regions of the two phase images.

[0009] As a preferred embodiment of the present invention, the method for extracting the multidimensional feature vector includes: For each tumor voxel, extract 3D arterial phase features: weighted arterial phase intensity. , Local texture Gradient magnitude ,in, The pixel HU value during the arterial phase. for Weight, For local three-dimensional standard deviation operators; Calculating intratumoral venous phase enhancement characteristics: enhancement differences ;ratio Absolutely enhanced differences Intensity during the venous phase Average strength ;Use an enhanced directional ternary indicator with a threshold of ±0.05 Where V is the pixel HU value during the venous phase; Will ~ The combination forms the multidimensional feature vector.

[0010] As a preferred embodiment of the present invention, robust scaling of multidimensional feature vectors is performed using the median and interquartile range RobustScaler.

[0011] As a preferred embodiment of the present invention, each tumor region is labeled with 5-50 sparse point tags, corresponding to three enhancement categories: low arterial intensity, medium arterial intensity, and high arterial intensity. The low arterial intensity is labeled as tag 1, the medium arterial intensity as tag 2, and the high arterial intensity as tag 3.

[0012] As a preferred embodiment of the present invention, the prior learning method includes: Features of sparse point voxels annotated by expert experience , and Learning prior knowledge, the feature vectors of sparse points labeled with category labels y∈1,2,3 are used. Perform cross-tumor aggregation; in: For each label category, calculate the inter-tumor summary statistic for arterial phase HU at all labeled sparse points, including mean, standard deviation, and median; Within each tumor, pairwise differences between label category-specific mean arterial phase HUs and intra-category HU dispersion were calculated to capture intratumoral heterogeneity patterns. At the cross-tumor level, the median and interquartile range are used to summarize the inter-tumor summary statistics, the pairwise differences, and the intra-category HU dispersion to assess whether the predicted clustering exhibits reasonable intra-tumor HU separation. Prior likelihoods of new voxels are obtained by estimating the class conditional feature distribution for each label category and arterial phase feature using Gaussian kernel density estimation (KDE) or Gaussian fitting. For each label category, the centroid is calculated based on the median feature vector, and the centroids are sorted from low to high according to the arterial phase intensity feature to standardize the semantics among patients, resulting in the centroids of label 1, label 2, and label 3. The importance of prior-guided features is estimated by the inter-class / intra-class variance ratio, normalized to a sum of 1 among features, and used to guide the selection of arterial phase features in prior-guided clustering. The centroid, distribution, intratumoral statistics, intertumoral statistics, and feature importance are used as learned prior knowledge and are all serialized and stored in pickle format on disk for reuse. When the number of sparse points that meet the prior learning conditions is insufficient, in order to avoid the prior constraints from having an unstable impact on the clustering results, this invention sets a minimum sparse point threshold. If, after validity screening, there are fewer than 10 sparse points located within the tumor region of interest and possessing valid grayscale information in both arterial and venous phase images, the prior learning step is skipped, and clustering segmentation without prior constraints is performed directly.

[0013] As a preferred embodiment of the present invention, the sub-region clustering method includes: Each tumor region was forced to generate three clusters, and the labels were reordered according to the mean weighted arterial phase intensity after clustering to meet the requirements. ; To prevent degenerate clustering, an improved constrained K-means algorithm is used, setting a minimum clustering ratio p=5%. For N tumor voxels, the minimum cluster size is... Iterative updates of allocation and center, employing a two-phase constraint allocation: One-stage constraint assignment: The constraint satisfaction process greedily assigns voxels based on the distance from the point to the center in a global ranking, ensuring that each cluster obtains at least m samples; Two-phase constraint assignment: The process completes by assigning all remaining voxels to the nearest center, recalculating the mean of the assigned points as the center, and iterating up to 100 steps. Convergence is evaluated by center stability. Clustering employs a prior knowledge-guided initialization and feature fusion strategy: arterial phase feature blocks consist of arterial phase features whose importance exceeds the average importance threshold and are robustly scaled; venous phase derived feature blocks consist of venous phase features with sufficient variance and are robustly scaled; the final fused feature representation is a weighted concatenation of arterial phase features and venous phase derived features, with weights of 60% and 40%, respectively. The centroid of prior knowledge is transformed into the scaled arterial phase space to initialize cluster centers, aligning the clusters with the prior semantics. When the constrained K-means clustering algorithm fails, a Gaussian Mixture Model (GMM) is used for clustering, and... Posterior probability of fused voxel-level Gaussian mixture model (GMM) Prior probabilities derived from the learning distribution and centroid similarity ;pass Assign labels; where, This represents the posterior probability that a sample belongs to cluster label k given a voxel feature vector x, where k is the cluster category label, typically {1, 2, 3} corresponding to three sub-regions, and x is the multi-dimensional feature vector after robust scaling and feature fusion. In a Gaussian mixture model, given x, the posterior probability that a sample belongs to cluster k. This represents the prior probability derived from the learned distribution and centroid similarity. The label is assigned by taking the maximum posterior probability over all possible k. If both constrained K-means and Gaussian mixture model (GMM) fail, then standard K-means is used for clustering.

[0014] As a preferred embodiment of the present invention, the post-processing and quality verification method includes: After initial clustering, verify whether the clustering results meet the constraint that "there are exactly three clusters and the smallest cluster size is m". If there are any discrepancies, enforce these constraints through post-processing. When the number of clusters is less than three, use K-means to split the largest cluster; When there are more than three clusters, the smallest cluster is merged into the nearest neighboring cluster in the feature space, forcing the formation of exactly three clusters; By iteratively reassigning voxels from clusters that do not meet the size requirement to the nearest other cluster centers, the minimum cluster size requirement is enforced. Then, the constraint of "exactly three clusters exist" is reapplied, and the labels are reordered according to the average weighted arterial phase intensity to obtain the one-dimensional cluster labels of the tumor region segmentation mask. Write the one-dimensional clustering labels within the tumor region segmentation mask back into the three-dimensional label map to obtain the three-dimensional sub-region segmentation label map, where background = 0; tumor voxels are labeled as 1-3, and saved as an NIFTI file.

[0015] As a preferred embodiment of the present invention, the tumor region is divided into low-perfusion subregions, medium-perfusion subregions, and high-perfusion subregions based on a three-dimensional subregion segmentation label map, wherein: The low-perfusion subregion is the region of tumor tissue with the lowest degree of enhancement in both phases, and very little perfusion in the arterial and venous phases. It reflects the tumor tissue region where necrosis or matrix is ​​dominant and the number of surviving tumor cells is limited. The mid-perfusion subregion is characterized by moderate enhancement in both the arterial and venous phases, representing the tumor tissue region with balanced perfusion dynamics. The high-perfusion subregion is characterized by the highest enhancement during the arterial phase and rapid washout during the venous phase, corresponding to metabolically active tumor tissue regions with rich blood vessels and high proliferative potential.

[0016] Compared with the prior art, the present invention has the following advantages: This invention introduces sparse prior knowledge annotated by experts to guide the clustering process, forming a two-stage algorithm framework of hierarchical prior learning and constrained clustering. It combines the advantages of data-driven and knowledge-driven approaches, enabling the algorithm to learn the distribution patterns of expert annotations and generalize to the entire tumor for robust segmentation even when expert annotations are sparse. Furthermore, the constraints avoid generating meaningless small fragments or inconsistent class numbers, ensuring the stability and consistency of the segmentation results on large-scale datasets. Attached Figure Description

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0018] Figure 1 Flowchart of a pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering provided in an embodiment of the present invention; Figure 2 A comparison chart showing the subregion segmentation results of pancreatic neuroendocrine tumor lesions with and without expert-guided prior learning, provided for embodiments of the present invention. Detailed Implementation

[0019] 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.

[0020] like Figure 1 As shown, this invention provides a method for pancreatic tumor subregion segmentation based on hierarchical prior learning and constrained clustering, comprising the following steps: Step S1: Obtain dual-phase enhanced CT images of pancreatic tumor patients and obtain segmentation masks for the tumor region; Step S2: Preprocess the dual-phase enhanced CT images, including resampling the images into isotropic voxels and registering the venous phase images to the arterial phase images; Step S3: Extract the multidimensional feature vector of each voxel within the tumor region, and perform robust scaling and normalization on the multidimensional feature vector. The multidimensional feature vector combines spatial features and arteriovenous time features. Step S4: Perform hierarchical prior learning using sparse point labels annotated by expert experience. The sparse point labels correspond to the three enhancement categories that guide sub-region segmentation. Step S5: Integrate the prior knowledge obtained from hierarchical prior learning into the constrained clustering framework, cluster voxels within the tumor region, forcibly divide them into three sub-regions, and apply minimum cluster size constraints during the clustering process. Step S6: Post-process and quality verification of the clustering results, and output a three-dimensional sub-region segmentation label map.

[0021] To achieve non-invasive, refined, and interpretable characterization and quantification of the internal heterogeneity of pancreatic neuroendocrine tumors (PNENs), this invention divides the tumor region into three subregions. This division is not simply based on image grayscale, but rather on the spatiotemporal perfusion dynamics and biological functional differences exhibited by the tumor in dual-phase enhanced CT. The aim is to construct an imaging biomarker system with a clear correspondence to the tumor's internal microenvironment (such as necrosis, stroma, and highly vascularized proliferative areas), thereby providing a reliable imaging structural basis for subsequent precise quantitative analysis, pathological grading prediction, and individualized treatment decisions.

[0022] This invention, when dividing a tumor into three subregions, first introduces sparse expert annotations from radiologists as hierarchical prior knowledge, transforming medical experience into quantifiable data guidance. Then, it constructs a multidimensional feature vector integrating spatial texture features and arteriovenous temporal dynamics features, establishing a comprehensive imaging description for each voxel. Next, the prior knowledge and multidimensional features are input into an improved constrained clustering framework. Through constraints such as the mandatory number of categories and minimum cluster size, a stable and interpretable three-region division is achieved. Finally, the clustering results are post-processed and consistency verified to ensure that the segmentation results conform to anatomical and pathophysiological principles. In this way, a biologically meaningful and clinically interpretable detailed deconstruction of tumor internal heterogeneity is achieved on non-invasive imaging, improving the accuracy of preoperative assessment and providing reliable imaging evidence for individualized treatment.

[0023] This invention employs a two-stage unsupervised / semi-supervised process to quantify intratumoral heterogeneity of pancreatic tumors using enhanced CT. The process (1) learns prior knowledge from clinically labeled intratumoral points; (2) applies a prior-guided clustering model with hard constraints on tri-clustering and HU gradient label sorting. The clustering process utilizes a built-in minimum cluster size constraint (5%), and is further enhanced through post-processing. The specific steps are as follows: Dual-phase enhanced CT images include venous phase images and arterial phase images. The pancreas and tumor are segmented in the venous phase images and arterial phase images using a pre-trained nnU-Net model to obtain segmentation results for the pancreatic region and the tumor region, and to form a segmentation mask for the tumor region.

[0024] Preprocessing involves resampling both the venous phase and arterial phase images. The voxels were calculated, and the venous and arterial phase images were cropped according to the pancreatic region segmentation results of the nnU-Net model into a shape centered on the pancreas and with edges of voxels. The bounding box is then used to precisely deform and register the venous phase image to the spatial coordinates of the arterial phase image, achieving alignment of the anatomical structures in the pancreatic and tumor regions of the two phase images.

[0025] This embodiment, based on nnU-Net, implements an independent segmentation, cropping, and registration process for dual-phase (venous and arterial phase) CT images on publicly available datasets such as PanTS. The two phases of data are processed by separate encoders to extract features, followed by deep cross-channel fusion. The decoder upsamples the fused features layer by layer and outputs the pancreatic mask (Mpan) and the tumor mask (Mtumor). If a sub-region branch is enabled, three types of sub-region probability maps are output. Furthermore, the network incorporates a prior learning branch that maps sparse point information to a prior distribution and linearly fuses it with data-driven posterior information, forming an end-to-end multi-task optimization framework. Five-fold cross-validation is used during training, ensuring consistent geometric scale and alignment between input and output to guarantee the effectiveness of pixel-level supervision. If finer-grained sub-region constraints or prior fusion are required in subsequent stages, the network can be further expanded on the sub-region and prior branches.

[0026] In the 3DFull-ResolutionnnU-Net experiment with five-fold cross-validation, the average Dice coefficient of the pancreas was approximately 0.92, and the average Dice coefficient of the tumor was approximately 0.78. This indicates that the method achieves a high degree of overlap in pancreatic boundary localization and is stable in the tumor region, providing reliable segmentation results under diverse lesion morphologies.

[0027] Methods for extracting multidimensional feature vectors include: For each tumor voxel, extract 3D arterial phase features (for prior learning): weighted arterial phase intensity. Local texture Gradient magnitude , The pixel HU value during the arterial phase. for Weight, For each voxel in the arterial phase image A, the standard deviation of the HU values ​​in its surrounding 3×3×3 neighborhood is calculated and used as a local texture feature. Calculate intratumoral venous phase enhancement features (for clustering): Enhancement differences ;ratio Absolutely enhanced differences Intensity during the venous phase Average strength ;Use an enhanced directional ternary indicator with a threshold of ±0.05 , The pixel HU value during the venous phase; Will ~ The combination forms a multidimensional feature vector.

[0028] In this case, the contributions of period A and period V to the heterogeneity weights are different, therefore These are adjustable parameters, and the present invention preferably sets them to 2.0.

[0029] Robust scaling of multidimensional feature vectors is performed using the median and interquartile range (IMR). Before clustering, features are scaled using robust scalers (median and IMR) to reduce sensitivity to outliers. In a priori-guided mode, feature blocks derived from the arterial and venous phases are scaled separately and then concatenated.

[0030] Each tumor region is labeled with 5-50 sparse point labels, corresponding to three enhancement categories: low arterial intensity, medium arterial intensity, and high arterial intensity. Low arterial intensity is labeled as label 1, medium arterial intensity as label 2, and high arterial intensity as label 3.

[0031] Prior learning methods include: Features of sparse point voxels annotated by expert experience , and Learning prior knowledge in middle school, categorized by tags Feature vectors of labeled sparse points Perform cross-tumor aggregation; If the total number of available sparse points is less than 10, then prior learning is skipped. For each label category, calculate the inter-tumor summary statistic for arterial phase HU at all labeled sparse points, including mean, standard deviation, and median; Within each tumor, pairwise differences between label category-specific mean arterial phase HUs and intra-category HU dispersion were calculated to capture intratumoral heterogeneity patterns. At the cross-tumor level, the median and interquartile range were used to summarize the inter-tumor summary statistics, pairwise differences, and intra-category HU dispersion to assess whether the predictive clustering showed reasonable intra-tumor HU separation. Prior likelihoods of new voxels are obtained by estimating the class conditional feature distribution for each label category and arterial phase feature using Gaussian kernel density estimation (KDE) or Gaussian fitting. For each label category, the centroid is calculated based on the median feature vector, and the centroids are sorted from low to high according to the arterial phase intensity feature to standardize the semantics among patients, resulting in the centroids of label 1, label 2, and label 3. The importance of prior-guided features is estimated by the inter-class / intra-class variance ratio, normalized to a sum of 1 among features, and used to guide the selection of arterial phase features in prior-guided clustering. Among them, the formula for calculating feature importance , For the first The feature importance of a feature is used to measure its ability to distinguish between different enhancement categories (low, medium, and high arterial intensity). For the first The between-class variance of a feature reflects the degree of difference in the mean of that feature between different classes (e.g., labels 1, 2, and 3). The larger the value, the better the feature can distinguish between different classes. For the first The within-class variance of a feature reflects the degree of dispersion of samples within the same class for that feature. The smaller the value, the more consistent the features are within the same class.

[0032] The centroid, distribution, intratumoral statistics, intertumoral statistics, and feature importance are used as learned prior knowledge and are all serialized and stored in pickle format on disk for reuse.

[0033] When the number of sparse points that meet the prior learning conditions is insufficient, in order to avoid the instability of the prior constraints on the clustering results, this invention sets a minimum sparse point threshold. When the number of sparse points located within the tumor region of interest and having effective grayscale information in both arterial and venous phase images is less than 10 after validity screening, the prior learning step is skipped, and clustering segmentation without prior constraints is performed directly.

[0034] Cluster centers in clustering splits without prior constraints can be determined using random initialization or heuristic methods based on data distribution (such as selecting quantile points as initial centers).

[0035] Subregion clustering methods include: Each tumor region was forced to generate three clusters, and the labels were reordered according to the mean weighted arterial phase intensity after clustering to meet the requirements. Standardized patient-to-patient explanations; To prevent degenerate clustering, an improved constrained K-means algorithm is used, setting a minimum clustering ratio p, preferably p=0.05. For N tumor voxels, the minimum cluster size is... Iterative updates of allocation and center, employing a two-phase constraint allocation: One-stage constraint assignment: The constraint satisfaction process greedily assigns voxels based on the distance from the point to the center in a global ranking, ensuring that each cluster obtains at least m samples; Two-stage constraint assignment: The process completes by assigning all remaining voxels to the nearest center; the mean of the assigned points is recalculated as the center, and the iteration continues until a preset number of steps is reached, such as a maximum of 100 steps. The convergence is evaluated by the center stability. Clustering employs a prior knowledge-guided initialization and feature fusion strategy: arterial phase feature blocks consist of arterial phase features with importance exceeding the average importance threshold and are robustly scaled; venous phase derived feature blocks consist of venous phase features with sufficient variance and are robustly scaled; the most fused feature is represented as a weighted concatenation of arterial phase features and venous phase derived features, with weights of 60% and 40%, respectively. The centroid of prior knowledge is transformed into the scaled arterial phase space to initialize cluster centers, aligning the clusters with the prior semantics. When the constrained K-means clustering algorithm fails, a Gaussian Mixture Model (GMM) is used for clustering, and... Posterior probability of fused voxel-level Gaussian mixture model (GMM) Prior probabilities derived from the learning distribution and centroid similarity ;pass Assign labels; where, This represents the posterior probability that a sample belongs to cluster label k given a voxel feature vector x, where k is the cluster category label, typically {1, 2, 3} corresponding to three sub-regions, and x is the multi-dimensional feature vector after robust scaling and feature fusion. In a Gaussian mixture model, given x, the posterior probability that a sample belongs to cluster k. This represents the prior probability derived from the learned distribution and centroid similarity. The label is assigned by taking the maximum posterior probability over all possible k. If both constrained K-means and Gaussian mixture model (GMM) fail, then standard K-means is used for clustering.

[0036] In practical use, this invention will If set to 0.65, then , and The two are linearly fused to obtain the final posterior probability P(k|x), where the weights 0.65 and 0.35 represent the balance between the data-driven GMM and prior information.

[0037] The constrained K-means algorithm in this invention refers to introducing a constraint on the minimum clustering ratio during clustering (e.g., setting p=0.05, ...). (where N is the current number of samples) to ensure that each cluster obtains at least a certain number of samples, thereby avoiding degradation; if constrained K-means clustering fails, it enters the Gaussian Mixture Model (GMM) stage to calculate the posterior probability and assign labels. If GMM also fails, it reverts to using standard K-means clustering. The intention of this design is that when data is sufficient and the feature distribution is relatively clear, GMM and prior information jointly provide robust posterior judgments; when data is insufficient or the distribution is complex, prior information guides the stabilization of clustering results, while setting a fallback path to ensure that an executable clustering output can still be obtained in the end.

[0038] Post-processing and quality verification methods include: After initial clustering, verify whether the clustering results meet the constraint that "there are exactly three clusters and the smallest cluster size is m". If there are any discrepancies, enforce these constraints through post-processing. When the number of clusters is less than three, use K-means to split the largest cluster; When there are more than three clusters, the smallest cluster is merged into the nearest neighboring cluster in the feature space, forcing the formation of exactly three clusters; By iteratively reassigning voxels from clusters that do not meet the size requirement to the nearest other cluster centers, the minimum cluster size is enforced. Then, the constraint of "exactly three clusters exist" is reapplied, and the labels are reordered according to the average weighted arterial phase intensity to obtain the one-dimensional cluster labels of the tumor region segmentation mask. Write the one-dimensional clustering labels within the tumor region segmentation mask back into the three-dimensional label map to obtain the three-dimensional sub-region segmentation label map, where background = 0; tumor voxels are labeled as 1-3, and saved as an NIFTI file.

[0039] Based on the three-dimensional subregion segmentation label map, the tumor region was divided into a low-perfusion subregion (label 1), a medium-perfusion subregion (label 2), and a high-perfusion subregion (label 3), where: The low-perfusion subregion is the region of tumor tissue with the lowest degree of enhancement in both phases, and very little perfusion in the arterial and venous phases. It reflects the tumor tissue region where necrosis or matrix is ​​dominant and the number of surviving tumor cells is limited. The mid-perfusion subregion is characterized by moderate enhancement in both the arterial and venous phases, representing the tumor tissue region with balanced perfusion dynamics. The high-perfusion subregion is characterized by the highest enhancement during the arterial phase and rapid washout during the venous phase, corresponding to metabolically active tumor tissue regions with rich blood vessels and high proliferative potential. Specifically, in terms of implementation: Phase 1: Learning prior knowledge from annotations.

[0040] For cases with effective intratumoral physician annotation, prior knowledge is learned from the arterial phase 3 feature representation of the annotated voxels. (By category labeling) The labeled feature vectors are aggregated across tumors; prior learning is skipped if the total number of available labeled points is less than 10. For each class, the inter-tumor aggregate statistics for the arterial phase HU (first feature) of all labeled points are calculated, including the mean, standard deviation, and median. To capture intra-tumor heterogeneity patterns, pairwise differences between class-specific mean arterial phase HU (e.g., class 1 vs 2, 1 vs 3, 2 vs 3) and intra-class HU dispersion (standard deviation when ≥ 2 points) are calculated within each tumor; at the cross-tumor level, these quantities are aggregated using the median and interquartile range (25%–75%), and then used to assess whether the predicted clustering exhibits reasonable intra-tumor HU separation.

[0041] Furthermore, when feasible, the category conditional feature distribution is estimated for each category and arterial phase feature using Gaussian kernel density estimation (KDE); otherwise, Gaussian fitting (mean and standard deviation) is used to calculate the prior likelihood of new voxels. Specifically, when the number of effectively labeled voxels for a category is sufficient (e.g., ≥10) and the data distribution is not uniform, Gaussian kernel density estimation (KDE) is used to estimate the category conditional feature distribution, where the kernel bandwidth is automatically calculated using Scott's rule, with a Gaussian kernel used by default; if the data volume is insufficient or the variance is zero, a fallback to Gaussian fitting based on the mean and standard deviation is used. The feature importance threshold is set to 0.8 times the mean of all feature importance, i.e., features with importance above 80% of the mean are retained by default. For each category, the centroid (median feature vector) is calculated, and the centroids are sorted from low to high according to the arterial phase intensity feature to standardize the semantics among patients, resulting in Label 1 (lowest arterial phase intensity), Label 2 (intermediate arterial phase intensity), and Label 3 (highest arterial phase intensity). Prior-guided feature importance is estimated using the between-class / within-class variance ratio (similar to ANOVA), normalized to a sum of 1 among features, and used to guide arterial phase feature selection in prior-guided clustering. All learned prior knowledge (centroid, distribution, intra-tumor / inter-tumor statistics, and feature importance) is serialized and stored on disk (pickle format) for reuse.

[0042] Phase 2: Prior-guided clustering with minimum cluster size constraints.

[0043] This process forces each tumor into exactly three clusters, and then reorders the labels according to the average weighted arterial phase intensity to meet the requirements. Standardized patient-to-patient interpretations were implemented. To prevent degenerate clustering, a custom constrained K-means algorithm was implemented, with a minimum clustering ratio p=0.05; for N tumor voxels, the minimum cluster size was... This algorithm is similar to the standard K-means algorithm, iteratively updating the assignments and centers, but employs a two-stage constraint assignment: First, the constraint satisfaction process greedily assigns voxels based on the distance from points to the centers in a globally ranked manner, ensuring that each cluster obtains at least m samples; second, the completion process assigns all remaining voxels to the nearest centers; the mean of the assigned points is recalculated as the centers, with a maximum of 100 iterations, and convergence is evaluated through center stability. When prior knowledge is available, clustering adopts a prior-guided initialization and feature fusion strategy: arterial phase feature blocks consist of arterial phase features with importance exceeding the average importance threshold and are robustly scaled, while venous phase derived feature blocks consist of venous phase features with sufficient variance (above the 30th percentile) and are robustly scaled; the final representation is a weighted concatenation of 60% arterial phase features and 40% venous phase derived features. If prior centroids are available, they are transformed to the scaled arterial phase space to initialize cluster centers (venous phase derived dimensions are initialized to zero), aligning the clustering with the prior semantics. If the constraint K-means fails, a Gaussian mixture model (GMM; maximum 3 initializations) is used, and then... Posterior probability of fused voxel-level Gaussian mixture model (GMM) Prior probabilities derived from the learning distribution and centroid similarity ;pass Assign labels. If both the constraint K-means and GMM fail, the standard K-means is used as a last resort. The weighting coefficients 0.65 / 0.35 are determined by grid search optimization on the validation set and are fixed hyperparameters used to balance data-driven and prior-guided approaches, ensuring the stability and reproducibility of the fallback mechanism.

[0044] Post-processing guarantees (triple guarantees) and quality checks. After initial clustering, it is verified that there are exactly three clusters, and each cluster size is at least m. If violations exist, these constraints are enforced through post-processing: when the number of clusters is less than three, the largest cluster is split using K-means (k=2); when the number of clusters is more than three, the smallest cluster is merged into the nearest neighboring cluster in the feature space, forcing the formation of exactly three clusters; voxels are iteratively reassigned from clusters that do not meet the size requirement to the nearest other cluster centers (up to 5 iterations), forcing the minimum cluster size to be met; then the "exactly three clusters" constraint is reapplied, followed by HU gradient label reordering. When prior knowledge of intratumoral HU differences is available, a multiplicative consistency score is calculated by comparing the observed HU separation between predicted clusters with the expected interquartile range learned from labeled tumors. This score serves as a diagnostic quality indicator, not an optimization objective.

[0045] Output generation. For each case, the one-dimensional cluster labels within the tumor mask are written back to a three-dimensional label map (background = 0; tumor voxels labeled 1-3), and saved as an NIFTI file. The summary CSV report for each case includes cluster counts, minimum cluster proportion and balance index, whether the 5% size constraint and HU gradient ordination are met, tumor descriptors (including tumor volume, proportion of each cluster, mean strength of each cluster, and standard deviation of enhancement heterogeneity (VA)), and overall quality indicators. .

[0046] In practical applications, the training set can be configured using methods such as grid search or cross-validation. , The optimal values ​​of parameters such as minimum clustering ratio p are determined. This invention, approved by the ethics committee and in accordance with the Declaration of Helsinki, included 461 pathologically confirmed PNENs patients from three medical centers. The study included a retrospective cohort of 398 patients recruited from May 2012 to November 2024, and a prospective cohort of 63 patients recruited from Hospital 1 from December 2024 to April 2025. Data processing complied with relevant laws and regulations such as the Personal Information Protection Law and the Measures for the Management of Personal Information Used in Scientific Research by Medical and Health Institutions, and security measures such as de-identification were implemented. The purpose of this invention is to assist in medical diagnosis and scientific research, aiming to improve diagnostic and treatment accuracy and patient well-being, and is in accordance with medical ethics.

[0047] Baseline clinical information, including age, sex, relevant laboratory parameters, and imaging features, was collected from all patients. Tumors were classified as grade 1 (G1), grade 2 (G2), grade 3 (G3), or neuroendocrine carcinoma (NEC) according to the 2019 WHO classification criteria. Following the WHO (2019) and European Society for Neuroendocrine Oncology (ENETS, 2023) guidelines, G3NETs and NECs were merged into a “high-grade” group to reflect their similar proliferative activity and enhancement patterns, and to ensure comparability with previous clinical and imaging studies. The final classification framework comprised three categories: G1, G2, and G3 / carcinoma.

[0048] All patients underwent dual-phase enhanced CT examination using a standardized protocol (120kVp, automatic tube current modulation, slice thickness 1-1.5mm). A standardized preprocessing procedure was adopted: (1) Arterial and venous phase images were resampled to isotropic 1×1×1mm. 3 (2) The pancreas and tumor were automatically segmented on the arterial phase image using a pre-trained nnU-Net model; (3) The two phase images were cropped into bounding boxes centered on the pancreas with 20mm edges; (4) The venous phase image was deformed and registered to the arterial phase using Elastix software, with the registration range limited to the pancreas and tumor regions; (5) Two experienced abdominal radiologists independently reviewed and manually corrected all segmentation and registration results under blind pathological grading, and reached a consensus on the discrepancies.

[0049] This invention extracts a 9-dimensional feature vector for each tumor voxel, combining spatial features (weighted enhancement degree, local texture, gradient) and arteriovenous temporal features (enhancement dynamics, washout rate, pattern encoding), and standardizes it through robust scaling. Two radiologists performed sparse point annotation (5-50 points per tumor) on 118 PNENs, covering three enhancement categories, to guide sub-region segmentation. These annotations were used during training to construct hierarchical prior knowledge and integrated into a constrained clustering framework, combining data-driven features with expert-guided distribution features. The algorithm initializes the centroid of subregions using statistical information derived by experts and adaptively adjusts voxel allocation through mixed posterior probability, dividing each tumor into three regions: (1) Low-perfusion subregion – with the lowest enhancement in both phases, and very little perfusion in the arterial and venous phases, possibly reflecting necrotic or matrix-dominated areas with a limited number of surviving tumor cells; (2) Medium-perfusion subregion – showing moderate enhancement in both the arterial and venous phases, representing tumor tissue with balanced perfusion dynamics; (3) High-perfusion subregion – showing the highest enhancement in the arterial phase and rapid washout in the venous phase, corresponding to metabolically active tumor regions with rich blood vessels and high proliferation potential. Consistency scoring is used. Assess intratumoral homogeneity and inter-subregional differentiation; a score ≥1.0 indicates good segmentation quality.

[0050] To evaluate the influence of expert-guided prior learning on clustering performance, a comparative evaluation is performed on models with and without expert annotation, recorded as Score1 and Score0 respectively. Under the two configurations with and without expert-guided prior learning, the same heterogeneity index derived from the sub-region results obtained by their respective clustering partitions is used as the input of the same classification model to predict tumor grading.

[0051] The predicted probability scores are divided into three equal-interval intervals, which correspond to the classification of three-level tumor grading prediction results: G1 grade (lower tertile), G2 grade (middle tertile) and G3 grade or neuroendocrine carcinoma (upper tertile) respectively. Six complementary indicators are used to quantify the classification performance: Accuracy, Weighted F1-score, Cohen's Kappa, Quadratic Weighted Kappa (QWK), Macro-AUC and Concordance Index (C-index). The relative improvement rate (%) is calculated as follows: Improvement rate (%) = (Score1-Score0)×100 / Score0.

[0052] The classification model specifically adopts an ordinal three-class classifier to perform three types of ordered prediction of tumor grades from low to high, and the output is a three-class probability distribution p G =[p G1 ,p G2 ,p G3 , where G1<G2<G3, corresponding to low-grade, intermediate-grade and high-grade respectively (for example, G3 may represent neuroendocrine carcinoma). During training, the ordinal classifier introduces grade relationship constraints, and adopts an ordinal target encoding strategy or ordinal losses (such as ordinal cross-entropy, serialized contrastive loss or thresholded boundary constraints) to strengthen the discrimination and order consistency between adjacent grades. To ensure end-to-end trainability, the input is the concatenated feature of sub-region heterogeneity indices from the two configurations Score1 and Score0, which passes through a unified classification head to output p_G, and optimization is performed within a five-fold cross-validation framework. The evaluation indicators remain accuracy, weighted F1, Cohen’s Kappa, Quadratic Weighted Kappa, Macro-AUC and C-index in the multi-classification scenario. The core of the ordinal three-class classifier lies in explicitly capturing the ordinal relationship between grades, so as to improve the stability and clinical interpretability of tumor grading.

[0053] The heterogeneity index discussed here is a unified tumor heterogeneity index (THI) developed in this invention to comprehensively capture multidimensional intratumoral heterogeneity by integrating dynamic enhancement patterns, spatial distribution, and temporal evolution features. This index is constructed using a ridge regression classifier, selecting features based on clinical relevance and statistical significance. All input features are quantile-normalized to ensure robust performance and mitigate the impact of outliers. Q(X) = Quantile converter (X); Q(X) transforms the feature X into a uniform distribution in the range [0,1], resulting in standardized feature values ​​that are insensitive to extreme values.

[0054] The THI score is calculated as a weighted linear combination of 10 carefully selected features: Score = 2.128347 × Q (amplitude diversity) - 1.285431 × Q (arterial phase standardized separation) + 1.134319 × Q (arterial phase kurtosis heterogeneity standardized value) - 0.656003 × Q (radial shape diversity) - 0.482377 × Q (mean signal attenuation ratio) + 0.367101 × Q (venous phase intergroup sum of squares) + 0.287913 × Q (mean stable ratio) - 0.099048 × Q (venous phase interquartile range coefficient of variation standardized value) - 0.096684 × Q (venous phase standardized separation) - 0.060550 × Q (arterial phase interquartile range coefficient of variation standardized value) - 0.333333; Standardize the raw scores to a 0-100 scale to improve clinical interpretability: THI = (Score - Lowest Score) / (Highest Score - Lowest Score) × 100; The lowest score was -1.808639 and the highest score was 2.892440, both derived from the training cohort distribution. A higher THI value indicates an increased degree of overall tumor heterogeneity.

[0055] After expert-guided prior learning, each pancreatic neuroendocrine tumor was automatically divided into three distinct subregions based on voxel-level dynamic enhancement features from arterial and venous phase CT images. Compared to unsupervised clustering, the expert-guided model generated more refined subregion boundary delineations that better matched subtle enhancement transitions. Quantitative evaluation showed that the tumor region subregion delineation incorporating expert-guided prior knowledge significantly improved the classification performance of tumor grading (see Table 1 and...). Figure 2As shown in Table 1, under the same experimental conditions, the introduction of prior learning based on expert knowledge and constrained clustering resulted in a stable improvement in multiple classification performance indicators, including accuracy, weighted F1 score, and macro-average AUC. Simultaneously, the improvements in consistency indicators such as the Coen-Kappa coefficient and weighted Kappa coefficient were 18.36% and 12.39%, respectively, indicating that this approach enhances the consistency between predicted results and actual pathological grading in multi-class scenarios, effectively reducing the randomness and instability of model judgments. Further combining... Figure 2 The technical advantages of this solution can be intuitively verified.

[0056] Figure 2 This paper presents tumor subregion segmentation results based on arterial and venous phase CT images, using methods without prior learning and with expert-guided prior learning, respectively. It can be seen that without expert prior learning, the subregion segmentation results generally suffer from fragmented boundaries and discontinuous spatial distribution, making it difficult to accurately reflect the true enhancement heterogeneity characteristics within the tumor. However, after introducing prior constraints based on expert sparse annotations, the resulting subregions are more spatially coherent, and the distribution patterns of different subregions are consistent with the dual-phase enhancement characteristics, better aligning with medical experts' understanding of tumor perfusion differences and tissue composition. In other words, the pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering used in this invention has significant advantages and can improve the accuracy of subsequent tumor grading.

[0057] Table 1. Evaluation Results of Expert-Guided Prior Learning

[0058] Figure 2 The visualization results of two representative cases of clustering PNENs lesion regions with high blood supply (ae) and low blood supply (fl) are presented, including (a)(f) arterial phase CT images; (b)(g) venous phase CT images; (c)(h) automatic clustering subregions before expert-guided learning (unsupervised results); (d)(i) sparse expert annotation visualization (sparse point annotations for the three enhancement categories); and (e)(j) optimized subregion maps after expert-guided prior learning, showing the optimized division of necrotic areas (red: low perfusion subregions), intermediate areas (green: medium perfusion subregions), and high blood supply areas (blue: high perfusion subregions).

[0059] This invention's fundamental difference from existing unsupervised clustering methods lies in its forced division into three sub-regions, imposition of a minimum cluster size constraint (p=0.05), prior knowledge-guided cluster initialization, and a triple guarantee mechanism in post-processing. Specifically: This invention directly corresponds to the three typical perfusion patterns (low perfusion, medium perfusion, and high perfusion) of pancreatic neuroendocrine tumors in dual-phase enhanced CT, avoiding the inconsistent results of traditional methods such as K-means, which output 2 or 4 categories due to data distribution fluctuations, thus solving the problem of inconsistent category numbers.

[0060] This invention mandates that each subregion contains at least 5% tumor voxels during the clustering process, effectively preventing the algorithm from dividing edge noise or local artifacts into meaningless small fragments, thus ensuring the anatomical continuity and clinical interpretability of the segmentation results.

[0061] This invention learns the category centroid and feature importance by using sparse points annotated by experts, so that the cluster center is semantically aligned with the clinical enhancement category (low / medium / high arterial intensity). This avoids the problems of unstable segmentation results and poor cross-patient consistency caused by the randomness of the initial center in pure data-driven methods.

[0062] After clustering is completed, this invention further ensures that the output results strictly meet the preset anatomical and physiological constraints through operations such as cluster number correction, minimum size enforcement, and HU gradient reordering, thus ensuring the stability and consistency of the segmentation results on large-scale datasets.

[0063] Furthermore, firstly, hierarchical prior learning differs from conventional methods that only learn cluster centers. Instead, it systematically learns multi-level prior knowledge from expert sparse annotations, including category centroids, distribution of category conditional features, intratumoral pairwise differences, intertumor summary statistics, and feature importance. This knowledge not only provides semantically aligned initial centers for clustering, but more importantly, it captures the intratumoral heterogeneity patterns of pancreatic neuroendocrine tumors in dual-phase enhanced CT, which is clinical prior information that cannot be obtained by purely data-driven methods.

[0064] Secondly, the deep integration of the constrained clustering framework with the learned priors is reflected in several aspects: (1) using the importance of prior features to guide the selection of arterial phase features, rather than using all features; (2) mapping the prior centroid to the feature space to initialize the clustering center, ensuring that the clustering results directly correspond to the clinical semantics of low, medium, and high perfusion; (3) in the Gaussian mixture model (GMM) backoff strategy, through The linear fusion method organically combines data likelihood with prior probabilities from prior distribution and centroid similarity, achieving an adaptive balance between data-driven and knowledge-driven approaches.

[0065] Furthermore, the mandatory tri-clustering, minimum cluster size (e.g., 5%), and post-processing triple safeguards set forth in this invention constitute a complete constraint system. Its purpose is not merely to restrict algorithmic behavior, but to fundamentally solve two specific technical problems commonly encountered in existing unsupervised segmentation methods for pancreatic tumor subregion division: the generation of clinically insignificant small fragments and inconsistent category numbers with data fluctuations. These constraints directly correspond to the clinicopathological understanding of pancreatic neuroendocrine tumors—that is, the tumor typically exhibits three typical perfusion patterns, and each pattern should occupy a sufficient tissue volume to have biological and clinical interpretable significance.

[0066] The synergistic effects of the above mechanisms are shown in Table 1 and Figure 2 This has been verified. As shown in Table 1, after introducing the expert-guided prior learning described in this invention, the Cohen's Kappa coefficient, which reflects the consistency between the classification results and the actual pathological grading, increased from 0.2307 to 0.2731, a relative improvement of 18.36%. This significant improvement is not achieved by the improvement of the clustering algorithm itself, but because the overall scheme described in this invention ensures the stability and consistency of the segmentation results across patient data, guaranteeing from the source that the extracted heterogeneous features truly reflect the biological nature of the tumor. Figure 2 The visual comparison further demonstrates that, guided by prior knowledge and constraints, the sub-region boundaries are more continuous and their distribution is more in line with the laws of enhancement dynamics, effectively avoiding the boundary fragmentation and spatial discontinuity phenomena commonly found in unsupervised segmentation.

[0067] In summary, this invention introduces sparse prior knowledge annotated by experts to guide the clustering process, forming a two-stage algorithm framework of hierarchical prior learning and constrained clustering. This framework combines the advantages of data-driven and knowledge-driven approaches, enabling the algorithm to learn the distribution patterns of expert annotations and generalize to the entire tumor for robust segmentation even when expert annotations are sparse. Furthermore, the constraints (such as mandatory 3-class division and minimum cluster size) avoid the generation of meaningless small fragments or inconsistent class numbers, ensuring the stability and consistency of the segmentation results on large-scale datasets. This is a key prerequisite for subsequent quantitative calculation of the tumor heterogeneity index (THI) and construction of accurate preoperative grading models. It ensures from the source that the extracted heterogeneity features truly reflect the biological nature of the tumor, thereby significantly improving the performance of downstream tasks.

[0068] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for pancreatic tumor subregion segmentation based on hierarchical prior learning and constrained clustering, characterized in that, Includes the following steps: Step S1: Obtain dual-phase enhanced CT images of pancreatic tumor patients and obtain segmentation masks for the tumor region; Step S2: Preprocess the dual-phase enhanced CT images, including resampling the images into isotropic voxels and registering the venous phase images to the arterial phase images; Step S3: Extract the multidimensional feature vector of each voxel within the tumor region, and perform robust scaling and normalization on the multidimensional feature vector. The multidimensional feature vector combines spatial features and arteriovenous time features. Step S4: Perform hierarchical prior learning using sparse point labels annotated by expert experience. The sparse point labels correspond to three enhancement categories that guide sub-region segmentation. The hierarchical prior learning refers to learning the feature centroids and feature distributions of the three enhancement categories from the labeled voxels to form prior knowledge. Step S5: Integrate the prior knowledge obtained from hierarchical prior learning into the constrained clustering framework, cluster voxels within the tumor region, forcibly divide them into three sub-regions, and apply a minimum cluster size constraint during the clustering process. The constrained clustering framework initializes the cluster centers using the category centroids in the prior knowledge, performs clustering after weighted fusion of arterial phase features and venous phase derived features, forces the number of voxels in each sub-region to be no less than a preset ratio, and assigns the three sub-regions to low-perfusion sub-regions, medium-perfusion sub-regions, and high-perfusion sub-regions respectively according to the weighted arterial phase intensity from low to high. Step S6: Post-process and quality verification of the clustering results, and output a three-dimensional sub-region segmentation label map.

2. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: The dual-phase enhanced CT images include venous phase images and arterial phase images. The pancreas and tumor are segmented in the venous phase images and arterial phase images using a pre-trained nnU-Net model to obtain the segmentation results of the pancreatic region and the tumor region, and to form a segmentation mask for the tumor region.

3. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: Preprocessing involves resampling both the venous phase and arterial phase images. The voxels were calculated, and the venous and arterial phase images were cropped according to the pancreatic region segmentation results of the nnU-Net model into a shape centered on the pancreas and with edges of voxels. The bounding box is then used to precisely deform and register the venous phase image to the spatial coordinates of the arterial phase image, achieving alignment of the anatomical structures in the pancreatic and tumor regions of the two phase images.

4. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: The method for extracting the multidimensional feature vector includes: For each tumor voxel, extract 3D arterial phase features: weighted arterial phase intensity. Local texture Gradient magnitude ,in, The pixel HU value during the arterial phase. for The weight, For local three-dimensional standard deviation operators; Calculating intratumoral venous phase enhancement characteristics: enhancement differences ;ratio Absolutely enhanced differences Intensity during the venous phase Average strength ;Use an enhanced directional ternary indicator with a threshold of ±0.05 Where V is the pixel HU value during the venous phase; Will ~ The combination forms the multidimensional feature vector.

5. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: Robust scaling of multidimensional feature vectors is achieved using the median and interquartile range.

6. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: Each tumor region is labeled with 5-50 sparse point labels, corresponding to three enhancement categories: low arterial intensity, medium arterial intensity, and high arterial intensity. Low arterial intensity is labeled as label 1, medium arterial intensity as label 2, and high arterial intensity as label 3.

7. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 4, characterized in that: The prior learning methods include: Features of sparse point voxels annotated by expert experience , and Learning prior knowledge, the feature vectors of sparse points labeled with category labels y∈1,2,3 are used. Perform cross-tumor aggregation, which refers to summarizing and statistically analyzing the arterial phase HU values ​​of the same category markers from multiple different tumors to obtain the inter-tumor aggregate statistics; in: For each label category, calculate the intertumor pooled statistic for arterial phase HU at all labeled sparse points, including mean, standard deviation, and median; Within each tumor, pairwise differences between label category-specific mean arterial phase HUs and intra-category HU dispersion were calculated to capture intratumoral heterogeneity patterns. At the cross-tumor level, the median and interquartile range are used to summarize the inter-tumor summary statistics, the pairwise differences, and the intra-category HU dispersion to assess whether the predicted clustering exhibits reasonable intra-tumor HU separation. Prior likelihoods of new voxels are obtained by estimating the class conditional feature distribution for each label category and arterial phase feature using Gaussian kernel density estimation (KDE) or Gaussian fitting. For each label category, the centroid is calculated based on the median feature vector, and the centroids are sorted from low to high according to the arterial phase intensity feature to standardize the semantics among patients, resulting in the centroids of label 1, label 2, and label 3. The importance of prior-guided features is estimated by the inter-class / intra-class variance ratio, normalized to a sum of 1 among features, and used to guide the selection of arterial phase features in prior-guided clustering. The centroid, distribution, intratumoral statistics, intertumoral statistics, and feature importance are used as learned prior knowledge and are all serialized and stored in pickle format on disk for reuse. When the number of sparse points that meet the prior learning conditions is insufficient, in order to avoid the prior constraints from having an unstable impact on the clustering results, this invention sets a minimum sparse point threshold. If, after validity screening, there are fewer than 10 sparse points located within the tumor region of interest and possessing valid grayscale information in both arterial and venous phase images, the prior learning step is skipped, and clustering segmentation without prior constraints is performed directly.

8. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 7, characterized in that: The subregion clustering method includes: Each tumor region was forced to generate three clusters, and the labels were reordered according to the mean weighted arterial phase intensity after clustering to meet the requirements. ; To prevent degenerate clustering, an improved constrained K-means algorithm is used, setting a minimum clustering ratio p. For N tumor voxels, the minimum cluster size is m = p N Iterative updates of allocation and center, employing a two-phase constraint allocation: One-stage constraint assignment: The constraint satisfaction process greedily assigns voxels based on the distance from the point to the center in a global ranking, ensuring that each cluster obtains at least m samples; Two-stage constraint assignment: The process completes by assigning all remaining voxels to the nearest center, recalculating the mean of the assigned points as the center, iterating for a preset number of steps, and evaluating convergence through the center stability. Clustering employs a prior knowledge-guided initialization and feature fusion strategy: arterial phase feature blocks consist of arterial phase features whose importance exceeds the average importance threshold and are robustly scaled; venous phase derived feature blocks consist of venous phase features whose variance meets the threshold and are robustly scaled; and finally, the fused feature representation is a weighted concatenation of arterial phase features and venous phase derived features. The centroid of prior knowledge is transformed into the scaled arterial phase space to initialize cluster centers, aligning the clusters with the prior semantics. When the constrained K-means clustering algorithm fails, a Gaussian Mixture Model (GMM) is used for clustering, and... Posterior probability of fused voxel-level Gaussian mixture model (GMM) Prior probabilities derived from the learning distribution and centroid similarity ; pass Assign tags; in, Let x represent the posterior probability that a sample belongs to cluster label k given a voxel feature vector x, where k is the cluster category label, taking {1, 2, 3} to correspond to the three sub-regions, and x is the multi-dimensional feature vector after robust scaling and feature fusion. In a Gaussian mixture model, given x, the posterior probability that a sample belongs to cluster k. This represents the prior probability derived from the learned distribution and centroid similarity. The label is assigned by taking the maximum posterior probability over all possible k. If both constrained K-means and Gaussian mixture model (GMM) fail, then standard K-means is used for clustering.

9. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: The post-processing and quality verification methods include: After initial clustering, verify whether the clustering results meet the constraint that "there are exactly three clusters and the smallest cluster size is m". If there are any discrepancies, enforce these constraints through post-processing. When the number of clusters is less than three, use K-means to split the largest cluster; When there are more than three clusters, the smallest cluster is merged into the nearest neighboring cluster in the feature space, forcing the formation of exactly three clusters; By iteratively reassigning voxels from clusters that do not meet the size requirement to the nearest other cluster centers, the minimum cluster size requirement is enforced. Then, the constraint of "exactly three clusters exist" is reapplied, and the labels are reordered according to the average weighted arterial phase intensity to obtain the one-dimensional cluster labels of the tumor region segmentation mask; Write the one-dimensional clustering labels within the tumor region segmentation mask back into the three-dimensional label map to obtain the three-dimensional sub-region segmentation label map, where background = 0; tumor voxels are labeled as 1-3, and saved as an NIFTI file.

10. The pancreatic tumor subregion segmentation method based on hierarchical prior learning and constrained clustering according to claim 1, characterized in that: Based on the three-dimensional subregion segmentation label map, the tumor region is divided into low-perfusion, medium-perfusion, and high-perfusion subregions, among which: The low-perfusion subregion is the region of tumor tissue with the lowest degree of enhancement in both phases, and very little perfusion in the arterial and venous phases. It reflects the tumor tissue region where necrosis or matrix is ​​dominant and the number of surviving tumor cells is limited. The mid-perfusion subregion is characterized by moderate enhancement in both the arterial and venous phases, representing the tumor tissue region with balanced perfusion dynamics. The high-perfusion subregion is characterized by the highest enhancement during the arterial phase and rapid washout during the venous phase, corresponding to metabolically active tumor tissue regions with rich blood vessels and high proliferative potential.

Images