CN-122019817-A - Image retrieval comparison system and method for inducing sputum cell auxiliary identification

CN122019817ACN 122019817 ACN122019817 ACN 122019817ACN-122019817-A

Abstract

The invention provides an image retrieval comparison system and method for inducing sputum cell auxiliary identification, which relate to the technical field of image retrieval and are characterized in that texture feature vectors and morphological feature vectors of sample images in a standard sample library are subjected to feature fusion to obtain fusion feature vectors of the sample images, multi-window feature refinement is carried out on all the fusion feature vectors to obtain a plurality of window feature class clusters of induced sputum cells, a feature index library of the induced sputum cells is determined through all the window feature class clusters, feature similarity between query feature vectors of the sputum cell images to be queried and all the fusion feature vectors in the feature index library is calculated, neighbor search is carried out on the sputum cell images to be queried through all the feature similarity to obtain target cell images and query tags, and feature increment learning is carried out on the feature index library by using the query tags. Based on the scheme, the on-line searching and dynamic evolution of the sputum cell image feature library can be induced.

Inventors

LIU XINGHUI
ZHANG YAOGANG
LI ZHILI
LIU YIYANG
LI YUAN
SUN MING
LIU BIN

Assignees

山东纬横数据科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. An image retrieval comparison method for inducing sputum cell auxiliary identification, which is characterized by comprising the following steps: Collecting and labeling a plurality of sample images of induced sputum cells to obtain a standard sample library, and extracting texture feature vectors and morphological feature vectors of each sample image in the standard sample library; Feature fusion is carried out on each texture feature vector and the corresponding morphological feature vector, so that fusion feature vectors of each sample image are obtained, multi-window feature refining is carried out on all fusion feature vectors, a plurality of window feature clusters of induced sputum cells are obtained, and then a feature index library of the induced sputum cells is determined through all window feature clusters; Acquiring a sputum cell image to be queried, and executing the same feature extraction and fusion feature extraction to obtain a query feature vector, so as to calculate feature similarity between the query feature vector and each fusion feature vector in the feature index library; And performing neighbor search on the sputum cell image to be queried through each feature similarity to obtain a target cell image and a query tag, and further performing feature increment learning on a feature index library by using the query tag to realize cluster evolution of the feature index library.
2. The method according to claim 1, wherein extracting texture feature vectors and morphological feature vectors of each sample image in the standard sample library specifically comprises: For each sample image in the standard sample library, carrying out texture response coding on the sample image, constructing a gray level co-occurrence matrix, calculating each gray level statistic from the gray level co-occurrence matrix, and then splicing and normalizing through all gray level statistic to obtain a texture feature vector; Performing contour detection on the binarized sample image, and calculating a plurality of basic shape parameters and nuclear characteristics to obtain morphological characteristic vectors of the sample image; And further obtaining texture feature vectors and morphological feature vectors of all sample images in the standard sample library.
3. The method according to claim 1, wherein feature fusion is performed on each texture feature vector and a corresponding morphological feature vector, and further obtaining a fused feature vector of each sample image specifically includes: For each sample image, carrying out normalization processing on the texture feature vector and the morphological feature vector of each sample image; vector stitching is carried out on the normalized texture feature vector and the morphological feature vector, and a primary fusion vector of the sample image is obtained; And performing feature dimension reduction on the primary fusion vector to obtain fusion feature vectors of the sample images, and further obtaining fusion feature vectors of all the sample images.
4. The method of claim 1, wherein performing multi-window feature refinement on all the fusion feature vectors to obtain a plurality of window feature clusters of induced sputum cells comprises: Dividing the feature space after dimension reduction into a plurality of feature windows with equal size; Feature refining is carried out on the fusion feature vector falling into each feature window, and window features of each feature window are obtained; and carrying out local clustering on all window features to obtain a plurality of window feature clusters of the induced sputum cells.
5. The method of claim 1, wherein determining the signature index library of induced sputum cells from all window signature clusters specifically comprises: integrating window characteristics and metadata of each window characteristic class cluster, and distributing a class cluster identifier to each window characteristic class cluster; Each feature window is used as a first-level index key, and a class cluster identifier in the corresponding feature window is used as a second-level index key to construct a hierarchical inverted index structure; And carrying out serialization storage based on the inverted index structure and the related sample images to obtain a feature index library of the induced sputum cells.
6. The method of claim 1, wherein performing neighbor search on the sputum cell image to be queried through each feature similarity to obtain the target cell image and the query tag specifically comprises: initializing an approximate nearest neighbor retrieval model based on a hierarchical navigable small world network; Taking each feature similarity as the edge weight among the graph nodes in the approximate nearest neighbor retrieval model; And searching the sputum cell image to be queried by using the approximate nearest neighbor search model to obtain a target cell image and a query tag.
7. The method according to claim 1, wherein the feature increment learning of the feature index library by using the query tag specifically comprises: Taking the current query image and the query tag as a new image sample, acquiring a fusion feature vector of the new image sample, and adding the fusion feature vector into an original data set of a feature index library; according to the fusion feature vector of the new image sample, determining a feature window to which the new image sample belongs and the nearest existing class cluster, updating the center and boundary range of the adjacent class cluster, and recalculating the representative feature subset; if the new image sample causes the scale growth of the existing class cluster to which the new image sample belongs to exceed a preset threshold value or the feature distribution is changed obviously, the class cluster in the feature window is triggered to be re-divided, and then the self-adaptive class cluster evolution of the feature index library is realized.
8. An image retrieval comparison system for inducing sputum auxiliary recognition for performing the image retrieval comparison method for inducing sputum auxiliary recognition according to any one of claims 1 to 7, the image retrieval comparison system for inducing sputum auxiliary recognition comprising an image retrieval unit, characterized in that the image retrieval unit comprises: The acquisition module is used for acquiring and labeling a plurality of sample images of induced sputum cells to obtain a standard sample library, and extracting texture feature vectors and morphological feature vectors of each sample image in the standard sample library; The processing module is used for carrying out feature fusion on each texture feature vector and the corresponding morphological feature vector so as to obtain fusion feature vectors of each sample image, carrying out multi-window feature refining on all the fusion feature vectors so as to obtain a plurality of window feature clusters of induced sputum cells, and further determining a feature index library of the induced sputum cells through all the window feature clusters; the processing module is further used for acquiring a sputum cell image to be queried and executing the same feature extraction and fusion feature extraction to obtain a query feature vector, and further calculating feature similarity between the query feature vector and each fusion feature vector in the feature index library; And the execution module is used for carrying out neighbor search on the sputum cell image to be queried through each feature similarity to obtain a target cell image and a query tag, and further carrying out feature increment learning on the feature index library by using the query tag to realize cluster evolution of the feature index library.
9. A computer device comprising a memory storing code and a processor configured to obtain the code and to perform the image retrieval comparison method for inducing sputum-aided identification of any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the image retrieval comparison method for inducing sputum cell assisted identification according to any one of claims 1 to 7.

Description

Image retrieval comparison system and method for inducing sputum cell auxiliary identification Technical Field The invention relates to the technical field of image retrieval, in particular to an image retrieval comparison system and method for inducing sputum cell auxiliary identification. Background The induced sputum cells are cell components in sputum samples obtained by stimulating respiratory tract secretion through atomizing and inhaling an inducer such as hypertonic saline, the method can obtain deep sputum from the lower respiratory tract, the cell components are richer and more typical than natural expectoration, and the induced sputum cells comprise alveolar macrophages, ciliated columnar epithelial cells and possible abnormal cells, and are important cytological examination bases for early screening of lung cancer and diagnosis of respiratory tract diseases. The existing sputum cell auxiliary identification generally builds a retrieval model based on a static feature library, the feature library keeps fixed parameters and sample sets after deployment, when new case data is faced, the system cannot effectively integrate new sample features into an original feature space structure, internal characterization of the model cannot be readjusted according to correction feedback of experts on retrieval results, coverage of the fixed feature library on cell morphological changes is gradually reduced along with accumulation of clinical sample diversity and evolution of pathological diagnosis standards, and feature distribution and real data distribution are offset, so that characteristic mismatch problems occur in the retrieval model. The technical defect is particularly characterized in that the sample retrieval precision of the system on the emerging pathological morphology or staining difference is reduced, namely the model performance is gradually degraded along with the drift of data distribution, and finally the long-term effectiveness and clinical applicability of auxiliary diagnosis are influenced. Therefore, how to realize the on-line retrieval and dynamic evolution of the induced sputum cell image feature library, thereby improving the diagnosis precision of the sputum cell auxiliary identification becomes a difficult problem in the industry. Disclosure of Invention The invention provides an image retrieval comparison system and method for inducing sputum cell auxiliary identification, which can realize the on-line retrieval and dynamic evolution of an image feature library of the induced sputum cell, thereby improving the diagnosis precision of the sputum cell auxiliary identification. In a first aspect, the present invention provides an image retrieval comparison method for inducing sputum cell assisted identification, comprising the steps of: Collecting and labeling a plurality of sample images of induced sputum cells to obtain a standard sample library, and extracting texture feature vectors and morphological feature vectors of each sample image in the standard sample library; Feature fusion is carried out on each texture feature vector and the corresponding morphological feature vector, so that fusion feature vectors of each sample image are obtained, multi-window feature refining is carried out on all fusion feature vectors, a plurality of window feature clusters of induced sputum cells are obtained, and then a feature index library of the induced sputum cells is determined through all window feature clusters; Acquiring a sputum cell image to be queried, and executing the same feature extraction and fusion feature extraction to obtain a query feature vector, so as to calculate feature similarity between the query feature vector and each fusion feature vector in the feature index library; And performing neighbor search on the sputum cell image to be queried through each feature similarity to obtain a target cell image and a query tag, and further performing feature increment learning on a feature index library by using the query tag to realize cluster evolution of the feature index library. In some embodiments, extracting the texture feature vector and the morphological feature vector of each sample image in the standard sample library specifically includes: For each sample image in the standard sample library, carrying out texture response coding on the sample image, constructing a gray level co-occurrence matrix, calculating each gray level statistic from the gray level co-occurrence matrix, and then splicing and normalizing through all gray level statistic to obtain a texture feature vector; Performing contour detection on the binarized sample image, and calculating a plurality of basic shape parameters and nuclear characteristics to obtain morphological characteristic vectors of the sample image; And further obtaining texture feature vectors and morphological feature vectors of all sample images in the standard sample library. In some embodiments, feature fusion is performed on each tex