A method for analyzing a tumor region of a pathological image

Abstract

This invention relates to a method for tumor region analysis in pathological images, belonging to the fields of medical image analysis and computational pathology. Addressing the problem of coarse localization caused by existing methods relying solely on block-level appearance features and ignoring kernel structure patterns, this invention explicitly models kernel-level spatial proximity and directional consistency by constructing an orientation-weighted kernel map. Combined with graph neural networks and a multi-instance learning framework, it achieves high-precision tumor region localization using only slice-level labels. The method includes the following steps: spatially consistent whole-slice image decomposition, orientation-prior-based kernel map construction, orientation-constrained graph representation learning, and kernel map-guided multi-instance learning aggregation. This invention can be used for weakly supervised localization and classification of tumor regions in whole-slice images.

Description

Technical Field

[0001] This invention relates to the field of computational pathology, and in particular to a method for tumor region analysis of pathological images. Background Technology

[0002] With the development of deep learning technology and computer-aided diagnosis, artificial intelligence is gradually breaking through the bottlenecks of traditional diagnosis. However, automatic tumor localization in gigapixel-level whole-slide images (WSIs) still faces significant challenges, especially under weak supervision conditions with only slice-level labels and a lack of region-level annotations. In routine pathological diagnosis, malignant progression is manifested not only in changes in cell nuclear morphology but also in synergistic alterations in cell morphology, spatial arrangement, and orientation patterns. Structural phenomena such as nuclear pleomorphism, anisotropic arrangement, and anomalous aggregation reflect the disordered structure of tissues rather than isolated visual cues. Due to the scarcity of pixel-level annotations in gigapixel WSIs, most methods employ Multiple Instance Learning (MIL), representing slices as packets of image patches and obtaining predictions through instance aggregation. Early MIL methods relied on simple pooling strategies, while attention-based MIL improves interpretability by learning adaptive instance weights. Subsequent variants such as CLAM (Clustering-constrained Attention MIL), DSMIL (Deep Set MIL), TransMIL (Transformer-based MIL), and HMIL (Hierarchical MIL) have further improved slice-level classification performance and are widely used in weakly supervised WSI analysis. However, these methods primarily encode block-level appearance features and implicitly assume that instances are independent of each other, leading to coarse localization results or spatial instability. In recent years, advances in computational pathology have enabled cell-level structural analysis. Deep learning-based kernel segmentation methods (such as Hover-Net and Cellpose) can provide reliable kernel detection in different tissues. The kernel representations obtained from these segmentations have shown value in grading, prognostic prediction, and biomarker discovery. Graph Neural Networks (GNNs) further allow for modeling the relational structures between cells, and kernel maps have shown advantages in capturing tissue structure. However, most existing graph constructions rely solely on spatial proximity and fail to adequately integrate directional alignment patterns characterizing the growth of invasive tumors. Therefore, there is an urgent need for a weakly supervised WSI analysis method that can explicitly integrate prior information on nuclear structure and simultaneously model spatial proximity and directional consistency, in order to improve the accuracy and interpretability of tumor localization, provide interpretable artificial intelligence-assisted tools for pathological diagnosis, and has significant clinical application value and promising prospects for promotion. Summary of the Invention The purpose of this invention is to overcome the shortcomings of existing technologies that rely solely on block-level appearance features and ignore nuclear structural patterns, and to propose a tumor region analysis method for pathological images. This method explicitly models the nuclear-level spatial tissue structure by constructing a Direction-Weighted Nuclear Graph (DWNG), and combines graph neural networks and a multi-instance learning framework to achieve high-precision tumor region localization using only slice-level labels. The technical solution is as follows: Step 1: Spatially consistent full-slice image decomposition: After coloring and normalizing the full-slice image, it is discretized into a set of non-overlapping image blocks, and the spatial anchor point position of each image block in the original slice coordinate system is recorded to establish a one-to-one correspondence between the image block index and the slice position. Step 2: Kernel Graph Construction Based on Orientation Prior: Within each image patch, cell nuclei are extracted using a pre-trained kernel segmentation model. Each cell nucleus serves as a kernel node, and each kernel node contains a segmentation mask, centroid coordinates, and principal axis orientation vectors calculated based on second-order central moments. Each kernel node is mapped to a multi-dimensional feature vector through a feature embedding function to represent morphological, staining, texture, and positional information. A kernel interaction graph is constructed based on the k-nearest neighbor strategy, where edge weights are jointly determined by spatial proximity weights and orientation consistency weights, forming an orientation-weighted kernel graph. Step 3: Learning the graph representation with orientation constraints: After linearly projecting the kernel node features, the node representation is updated through a multi-layer orientation-weighted message passing mechanism, where the message aggregation process is guided by the orientation-weighted adjacency matrix constructed in Step 2; an attention-based permutation-invariant readout mechanism is used to obtain the image patch-level embedding representation; Step 4: Kernel-guided multi-instance learning aggregation: Each slice is formalized as a package of image patch instances, and supervision is performed using only slice-level labels; the image patch embeddings are aggregated using an attention-based multi-instance learning operator to obtain slice-level predictions; based on the preserved spatial anchor information, the learned attention scores are reprojected onto the original slice space to generate a weakly supervised tumor probability map. The beneficial effects of this invention are: By constructing a direction-weighted kernel map, explicitly integrating kernel-level spatial proximity and directional consistency information, the limitations of existing methods that rely solely on block-level appearance features can be overcome, enabling more accurate capture of cellular and tissue structure patterns relevant to pathological diagnosis. A direction-weighted message passing mechanism is employed to learn structure-aware image block representations, combined with an attention-based multi-instance learning framework. Under weak supervision using only slice-level labels, superior classification performance and tumor localization quality compared to existing baseline methods can be achieved. Tumor probability maps are generated through reprojection of learned attention scores, providing interpretable localization results with spatial coherence and structural consistency without region-level supervision, achieving a high degree of consistency with pathology expert assessments. Attached Figure Description Figure 1 This is an overall flowchart of the multi-instance learning framework based on direction-weighted kernel graph guidance of the present invention; Detailed Implementation To make the technical solution of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings. The present invention is implemented in specific steps: The proposed orientation-weighted kernel graph-guided multiple instance learning (NGG-MIL) framework comprises three main modules: a spatially consistent full-slice image decomposition module, an orientation-prior-based kernel graph construction and graph representation learning module, and a kernel graph-guided multiple instance learning aggregation module. Step 1: Spatially Consistent Full-Slice Image Decomposition Full slice image Represented as a continuous color field, where Slice The spatial domain. After coloring normalization, the sampling operator... Discretize the WSI into a set of non-overlapping image patches. ,in Indicates the first Image blocks, This represents the spatial anchor point in the original slice coordinate system. This representation maintains a one-to-one correspondence between image patch indices and slice positions, enabling spatial reprojection of instance-level responses. The image patch size is set to 256×256 pixels, extracted at a magnification of 40x. Color normalization employs a sparse nonnegative matrix factorization-based method to reduce color variation between different slices. Step 2: Kernel graph construction based on orientation prior In each image block First, kernel nodes are extracted using a pre-trained HoVer-Net model. Each kernel node is represented as... ,in To segment the mask, For the coordinates of the nucleus and centroid, This is the unit direction vector calculated from the principal axis of the mask pixel covariance matrix. Through feature embedding function Map each core node to A 3D feature vector encodes morphological, coloration, texture, and positional information. Let... For image blocks The set of kernel features within. Constructing a direction-weighted kernel graph ,in For the set of core nodes and edge set Defined by the k-nearest neighbor strategy in the spatial domain: )} For each edge Define two affinity terms: Spatial proximity weight: Directional consistency weight: in The bandwidth hyperparameter is used to control spatial attenuation. The non-normalized edge weights are... By node Normalization within the neighborhood yields a row random adjacency matrix. . Step 3: Learning the graphical representation of orientation constraints Given a kernel map First, the node features are linearly projected as... ,in This is a learnable weight matrix. Then, through... Layer-wise weighted message passing update graph representation: in For layer-specific transformation matrices, It is a non-linear activation function. To obtain image patch-level embeddings, an attention-based permutation-invariant readout mechanism is applied: in For learnable attention vectors, For the first Layer nodes Feature representation, This represents the number of kernel nodes within an image patch. This is for image patch-level embedding. The graph neural network encoder uses two graph convolutional layers with a hidden dimension of 128. Step 4: Kernel Graph-Guided Multi-Instance Learning Aggregation Each slice Packages formalized as image patch instances: Use only slice level tags To conduct oversight. Aggregating image patch embeddings using an attention-based multi-instance learning operator: Where V and w are learnable parameters, For image patch-level embedding, This represents the number of image blocks in the slice. Slice-level prediction ,in This is the sigmoid function. Furthermore, it preserves the spatial anchor point. Attention score Reprojection to Above, a weakly supervised tumor localization map is generated.

Claims

1. A method for analyzing tumor regions in pathological images, characterized in that, Includes the following steps: Step 1: Spatially consistent full-slice image decomposition: After coloring and normalizing the full-slice image, it is discretized into a set of non-overlapping image blocks, and the spatial anchor point position of each image block in the original slice coordinate system is recorded to establish a one-to-one correspondence between the image block index and the slice position. Step 2: Kernel graph construction based on orientation prior: Within each image patch, kernel nodes are extracted using a pre-trained kernel segmentation model. Each kernel node contains a segmentation mask, centroid coordinates, and principal axis orientation vectors calculated based on second-order central moments. Each kernel node is mapped to a multi-dimensional feature vector through a feature embedding function; a kernel interaction graph is constructed based on the k-nearest neighbor strategy, where the edge weights are jointly determined by spatial proximity weights and directional consistency weights, forming a directional weighted kernel graph. Step 3: Learning the graph representation with orientation constraints: After linearly projecting the kernel node features, the node representation is updated through a multi-layer orientation-weighted message passing mechanism, where the message aggregation process is guided by the orientation-weighted adjacency matrix constructed in Step 2; an attention-based permutation-invariant readout mechanism is used to obtain the image patch-level embedding representation; Step 4: Kernel-guided multi-instance learning aggregation: Each slice is formalized as a package of image patch instances, supervised only by slice-level labels; the image patch embeddings are aggregated using attention-based multi-instance learning operators to obtain slice-level predictions; Based on the preserved spatial anchor information, the learned attention score is reprojected onto the original slice space to generate a weakly supervised tumor probability map.

2. The method for tumor region analysis in pathological images according to claim 1, characterized in that, The spatial proximity weights mentioned in step 2 are calculated using a Gaussian kernel function, and the calculation formula is as follows: in, and Let J and K be the centroid coordinates of the core nodes j and k, respectively. To control the bandwidth hyperparameter of spatial attenuation; the directional consistency weight mentioned in step 2 is calculated by the inner product of the kernel direction vectors, and the calculation formula is: in, and These are the unit direction vectors for kernel nodes j and k, respectively.

3. The method for tumor region analysis in pathological images according to claim 1, characterized in that, The attention-based permutation-invariant readout mechanism described in step 3 calculates the attention weight of each kernel node using a learnable attention vector, and then weights and aggregates them to obtain the image patch-level embedding. The calculation formula is as follows: in For learnable attention vectors, For the first Layer nodes Feature representation, This represents the number of kernel nodes within an image patch. Embedded at the image patch level.

4. The method for tumor region analysis in pathological images according to claim 1, characterized in that, The attention-based multi-instance learning operator in step 4 includes: calculating the attention score for each image patch using a learnable transformation matrix and attention vector; weighting and aggregating the embeddings of all image patches to obtain a slice-level representation; and finally obtaining a classification prediction using the sigmoid function. The formula for calculating the attention score is as follows: Where V and w are learnable parameters, For image patch-level embedding, This represents the number of image blocks in the slice.

Images