CN-121982018-A - Image recognition analysis-based pathology detection device and method

CN121982018ACN 121982018 ACN121982018 ACN 121982018ACN-121982018-A

Abstract

The invention relates to the technical field of image data processing, and discloses a pathology detection device and a pathology detection method based on image recognition analysis. The method comprises the steps of obtaining a pathological full-slice image, identifying cell nucleus and gland structures, constructing a cross-scale heterogram comprising cell nucleus nodes, gland nodes, same-scale sides and cross-scale sides based on identification results, inputting the heterogram into a graph neural network for coding to obtain node representation fused with cross-scale context information, randomly shielding part of gland node characteristics in a training stage, carrying out self-supervision learning by predicting shielded characteristics, calculating prediction errors of each gland node as structural abnormality indexes in a detection stage, and generating a structural abnormality thermodynamic diagram. Realizing the accurate quantification of the abnormal shape of the tissue structure, solving the technical problem that the prior method ignores the semantic meaning of the tissue structure and relies on expensive labeling, the method has the advantages of high detection accuracy, strong interpretability and good clinical applicability.

Inventors

ZHANG HONGLAN
YANG CONGYING
JIANG YANTING

Assignees

连云港市第一人民医院

Dates

Publication Date: 20260505
Application Date: 20260401

Claims (10)

1. The pathology detection method based on image recognition analysis is characterized by comprising the following steps of: S1, obtaining a pathological full-slice image of a tissue to be detected, and respectively identifying cell nuclei and gland structures in the image under at least two amplification factors; s2, constructing a cross-scale heterogram based on the identified cell nuclei and glands, wherein the cross-scale heterogram comprises two types of nodes, the first type of nodes correspond to each cell nucleus and are associated with first feature vectors, the second type of nodes correspond to each gland and are associated with second feature vectors, and the cross-scale heterogram further comprises a first type of edges constructed based on the spatial proximity relation between the cell nuclei, a second type of edges constructed based on the spatial proximity relation between the glands and cross-scale edges constructed between the corresponding nodes when the cell nuclei are positioned in the glands; S3, inputting the cross-scale heterogeneous graph into a graph neural network for coding, and enabling each node to aggregate information of neighbor nodes through an attention mechanism to obtain a node representation vector fusing cross-scale context information, wherein the neighbor nodes comprise nodes with the same scale and nodes with different scales connected through cross-scale edges; S4, in a training stage, randomly masking a second feature vector of part of gland nodes, coding the second feature vector of the masked nodes through the graph neural network by utilizing the unmasked nodes, and inputting a prediction network to predict the second feature vector of the masked nodes, wherein the graph neural network and the prediction network are trained by taking the error minimization between a predicted value and a true value as a target; s5, in a detection stage, inputting a cross-scale heterogeneous graph of a sample to be detected into a trained graph neural network and a prediction network, and calculating a prediction error of each gland node as a structural abnormality index of the gland; s6, generating a structural abnormality thermodynamic diagram of the pathological image based on the structural abnormality index, and outputting a detection result.
2. The method for detecting pathology based on image recognition analysis according to claim 1, wherein the recognition of the cell nuclei comprises segmenting the outline of each cell nucleus on a high-magnification pathology image through a deep learning segmentation model, extracting nuclear morphology features and texture features to form a first feature vector, and the recognition of the glands comprises segmenting the outline of each gland on a medium-magnification pathology image through a deep learning segmentation model, extracting glandular morphology features and cell distribution features to form a second feature vector.
3. The pathology detection method based on image recognition analysis according to claim 1, wherein the construction of the cross-scale heterogram specifically comprises taking all cell nuclei as first class nodes, taking all glands as second class nodes, taking all the glands as second class nodes, setting up a first class edge when the spatial Euclidean distance of each pair of cell nuclei nodes is smaller than a first preset threshold value, setting up a second class edge when the spatial Euclidean distance of each pair of gland nodes is smaller than a second preset threshold value, and setting up a cross-scale edge between each pair of cell nuclei nodes if the coordinates of each pair of cell nuclei nodes are located in a partitioned area of any gland node.
4. The pathology detection method based on image recognition analysis according to claim 1, wherein the graph neural network is a graph annotation force network, and the coding process comprises the steps of calculating attention coefficients of each node and all neighbor nodes, wherein the attention coefficients are obtained based on similarity of node self characteristics and neighbor node characteristics, weighting and aggregating the neighbor node characteristics by using the attention coefficients, updating a representation vector of a current node through nonlinear transformation, and carrying out multi-layer iteration to enable a final representation vector of each node to contain context information of a local structure of the final representation vector, wherein the context information comprises distribution of cells or glands with the same scale and integral characteristics of individual characteristics or glands of the administered cells.
5. The pathology detection method based on image recognition analysis according to claim 1, wherein the second feature vector of the part of glandular nodes is randomly selected according to a preset masking proportion in each training iteration, the second feature vector is replaced by a preset masking mark vector, the prediction network is a multi-layer perceptron, a final representation vector obtained by encoding the masking nodes through a graph neural network is mapped into a prediction feature vector, and the error minimization is achieved by calculating Euclidean distances between the prediction feature vectors of all the masking nodes and the true second feature vector of all the masking nodes and averaging the Euclidean distances as a loss function.
6. The method for detecting pathology based on image recognition analysis according to claim 1, wherein the structural abnormality index is calculated by inputting a final representation vector obtained by encoding each gland node in a sample to be detected through a graph neural network into a prediction network with training completed to obtain a prediction feature vector of the gland node, calculating the Euclidean distance between the prediction feature vector and an original second feature vector of the gland node, and taking the Euclidean distance as the structural abnormality index of the gland, wherein the index reflects the coordination degree between the gland and a context environment formed by surrounding cells and other glands, and the greater the index is, the higher the possibility of indicating structural abnormality is.
7. The method for detecting pathology based on image recognition analysis according to claim 1, wherein the generating a structural anomaly thermodynamic diagram comprises dividing a whole image of pathology into regular grids, counting structural anomaly indexes of all gland nodes in each grid for each grid unit, calculating statistics as anomaly scores of the grid units, wherein the statistics comprise average values, maximum values or median values, converting the anomaly scores of each grid unit into pixel values through color mapping, and generating a thermodynamic diagram aligned with an original image space, wherein a region with darker color in the thermodynamic diagram represents a structural anomaly more remarkable.
8. The method for detecting pathology based on image recognition analysis according to claim 1, further comprising a diagnosis classification step of extracting global features based on the structural anomaly thermodynamic diagram, wherein the global features comprise the number of anomaly regions, total area, average anomaly index and anomaly region distribution density, inputting the global features into a pre-trained classifier, and outputting diagnosis categories of tissues to be detected, wherein the diagnosis categories comprise normal, benign lesions and malignant lesions.
9. The method for detecting pathology based on image recognition analysis according to any one of claims 1 to 8, wherein the training phase is performed by using only the pathology image of normal tissue without labeling the lesion area, and the model is made to grasp the spatial organization rule between cells and glands in the normal tissue structure by self-supervised learning, so that the abnormal structural area deviating from the normal rule is recognized by prediction error in the detection phase.
10. A pathology detection apparatus based on image recognition analysis for implementing a pathology detection method based on image recognition analysis according to any one of claims 1 to 9, comprising: the image acquisition module is used for acquiring a pathological full-slice image of the tissue to be detected; The structure identification module is used for identifying cell nucleus and gland structures from the pathology full-film image under at least two amplification factors respectively, and extracting a first feature vector of each cell nucleus and a second feature vector of each gland; A diagram construction module for constructing a cross-scale heterogram based on the identified cell nuclei and glands, the cross-scale heterogram comprising cell nuclei nodes, gland nodes, co-scale edges based on spatial proximity relations, and cross-scale edges based on containment relations; the graph coding module is used for coding the cross-scale heterograph through a graph neural network to obtain a representation vector of fused cross-scale context information of each node; the prediction module is used for predicting the second feature vector of the gland node and outputting a prediction feature vector; The anomaly scoring module is used for calculating the error between the predicted characteristic vector of each gland node and the true second characteristic vector of each gland node to be used as the structural anomaly index of the gland; the visual output module is used for generating a structural abnormality thermodynamic diagram based on the structural abnormality index and outputting the structural abnormality thermodynamic diagram; The image coding module and the prediction module are obtained through pre-training.

Description

Image recognition analysis-based pathology detection device and method Technical Field The invention relates to the technical field of image data processing, in particular to a pathology detection device and a pathology detection method based on image recognition analysis. Background In the existing pathological image analysis method based on deep learning, a pathological full-film image is generally segmented into a plurality of image blocks, the visual characteristics such as textures, forms and the like of the image blocks are extracted through a convolutional neural network, and then classification or segmentation of a tumor area is realized, however, the method has a fundamental method theory defect that the pathological image is treated as a natural image, and a model is learned in a pixel mode instead of structural semantics. The essence of pathological diagnosis is structural analysis problem, cancer cells are identified, the cancer cells are not abnormal in morphology of single cells, but destroy normal tissue structures, doctors judge whether glands are malignant or not according to structural semantic information such as whether gland outlines are complete, whether cell arrangement is disordered, whether gland cavity morphology is regular or not, but the prior model does not really understand the structural semantic of how cells are organized into glands and how glands are organized into tissues, but learns the correlation between pixel statistical distribution of local image blocks and lesion labels. This results in a model that is extremely sensitive to "spurious structural disturbances" caused by the production artefacts, but it is difficult to identify early lesions where structural damage is not yet significant, and it is also prone to false positives due to benign lesions just appearing similar to the textural features of malignant lesions. Therefore, there is a need for a pathology detection method that enables models to understand tissue structure semantics and make diagnostic reasoning based on the spatial tissue relationship of cells and glands. Disclosure of Invention The invention aims to solve the technical problem that a pathology image analysis and diagnosis system based on deep learning in the prior art is difficult to accurately identify pathological changes, and therefore, the invention provides a pathology detection device and a pathology detection method based on image identification and analysis. In order to achieve the purpose, the application adopts the following technical scheme that the pathology detection method based on image recognition analysis comprises the following steps: S1, obtaining a pathological full-slice image of a tissue to be detected, and respectively identifying cell nuclei and gland structures in the image under at least two amplification factors; S2, constructing a cross-scale heterogram based on the identified cell nuclei and glands, wherein the cross-scale heterogram comprises two types of nodes, the first type of nodes correspond to each cell nucleus and are associated with a first feature vector, the second type of nodes correspond to each gland and are associated with a second feature vector, and the cross-scale heterogram further comprises a first type of edges constructed based on the spatial proximity relation between the cell nuclei, a second type of edges constructed based on the spatial proximity relation between the glands and cross-scale edges constructed between the corresponding nodes when the cell nuclei are positioned in the glands; s3, inputting the cross-scale heterogeneous graph into a graph neural network to encode, so that each node aggregates the information of neighbor nodes through an attention mechanism to obtain a node representation vector fusing the cross-scale context information, wherein the neighbor nodes comprise nodes with the same scale and nodes with different scales connected through cross-scale edges; S4, in the training stage, randomly masking a second feature vector of part of gland nodes, coding the second feature vector of the masked nodes through a graph neural network by utilizing the unmasked nodes, and inputting the second feature vector of the masked nodes into a prediction network to predict the second feature vector of the masked nodes, wherein the error between a predicted value and a true value is minimized to be a target training graph neural network and a prediction network; s5, in a detection stage, inputting a cross-scale heterogeneous graph of a sample to be detected into a trained graph neural network and a prediction network, and calculating a prediction error of each gland node as a structural abnormality index of the gland; s6, generating a structural abnormality thermodynamic diagram of the pathological image based on the structural abnormality index, and outputting a detection result. Preferably, the identification of the cell nuclei comprises the steps of dividing the outline of each cell n