CN-121998951-A - Pathological recognition analysis method and system based on blood detection sample

CN121998951ACN 121998951 ACN121998951 ACN 121998951ACN-121998951-A

Abstract

The application relates to the field of intelligent medical treatment, and discloses a pathology recognition analysis method and system based on a blood detection sample. The method comprises the steps of obtaining a blood smear microscopic image and extracting cell outlines, carrying out texture analysis on the insides of the outlines to obtain morphological characteristics, judging potential abnormal areas based on the characteristics, carrying out local enhancement on the abnormal areas and extracting fine difference characteristics, obtaining quantitative pathological descriptors based on the difference characteristics through a deep learning model, identifying abnormal cells according to the matching results of the descriptors and a feature library, and comprehensively analyzing according to the identification results to obtain pathological quantitative indexes. The application solves the problems of low recognition precision and poor efficiency of abnormal cells of blood smear by combining a local enhancement and deep learning model with feature library matching, and improves the accuracy and automation level of pathological analysis.

Inventors

YIN JUNJIE
XU GUOCHAO
ZHANG CHAO
Kong Yakun

Assignees

新乡市中心医院(新乡中原医院管理中心)

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (9)

1. The pathological recognition analysis method based on the blood detection sample is characterized by comprising the following steps of: S101, acquiring a blood smear microscopic image of a blood detection sample, and extracting cell area boundary information from the blood smear microscopic image to obtain a preliminary cell morphology contour; Step S102, performing texture analysis on the internal area of the outline of the primary cell morphology to obtain cell morphology distribution characteristics; step S103, comprehensively analyzing the morphological distribution characteristics of the cells, and judging whether a potential abnormal cell area exists or not; Step S104, carrying out local image enhancement processing on the potential abnormal cell area, and extracting a fine difference feature set; Step S105, carrying out deep analysis on the potential abnormal cell area based on the fine difference feature set and combining a deep learning model to obtain a quantized pathological feature descriptor; Step S106, judging whether a pathological abnormal signal exists or not according to the matching degree of the quantized pathological feature descriptor and a preset feature library, and outputting an abnormal cell identification result; and step S107, comprehensively analyzing the blood smear microscopic image according to the abnormal cell identification result to obtain a comprehensive pathology quantification index.
2. The method according to claim 1, wherein the step S101 includes: shooting the blood smear in multiple areas and multiple magnifications through microscopic imaging equipment to obtain a microscopic image of the blood smear; Dividing the blood smear microscopic image by an image processing algorithm, and extracting boundary information of a cell area; And carrying out edge strengthening and connection processing by using a Canny edge detection algorithm based on the extracted boundary information, and generating a continuous and complete preliminary cell morphology profile.
3. The method according to claim 2, wherein the step S102 includes: Performing texture analysis on the internal area of the outline of the primary cell morphology, and extracting the shape distribution characteristics and the color uniformity distribution characteristics of the central light dyeing area of the red blood cells in the outline; performing quantization processing on the shape distribution characteristics and the color uniformity distribution characteristics, and then applying principal component analysis to reduce the dimension to generate a low-dimension feature vector; Calculating the Euclidean distance between the low-dimensional feature vector and a preset standard feature vector, and quantifying the abnormal degree of the cell morphology according to the Euclidean distance; Identifying the nuclear boundary of the white blood cells by adopting a contour tracking algorithm, analyzing and counting the number of nuclear leaf branches of each white blood cell by using a connecting component, and generating a distribution characteristic of the number of nuclear leaf branches; And carrying out weighted fusion on the abnormality degree and the leaf number distribution characteristics of the nuclear leaves to obtain cell morphology distribution characteristics.
4. The method according to claim 1, wherein the step S103 includes: Extracting shape distribution characteristics, color uniformity distribution characteristics and nuclear leaf division number distribution characteristics from the cell morphology distribution characteristics; Calculating KL divergence of the shape distribution characteristics and normal cell morphology distribution, performing entropy calculation on the color uniformity distribution characteristics, adopting Poisson distribution fit to the nuclear leaf division quantity distribution characteristics, and calculating chi-square statistics; the method comprises the steps of presetting an anomaly judgment threshold, and judging that a preset anomaly judgment condition is met if the KL divergence is larger than the preset divergence threshold, the entropy is smaller than the preset entropy threshold and the chi-square statistic is larger than the preset chi-square threshold; and determining the area where the cells meeting the preset abnormality judgment conditions are located as a potential abnormal cell area.
5. The method according to claim 1, wherein the step S104 includes: Carrying out local image enhancement processing on the potential abnormal cell area by adopting a self-adaptive histogram equalization method, and highlighting cell detail characteristics; Based on the enhanced potential abnormal cell area image, identifying the boundary of the central light dyeing area and the boundary of cytoplasm of the red blood cells, and calculating the vertical distance from each boundary point to the adjacent boundary to form a width sequence; Carrying out statistical analysis on the width sequence, calculating average width and standard deviation, and constructing a distribution histogram to obtain width distribution characteristics of a central light dyeing region and a cytoplasmic transition region of the red blood cells; denoising the leucocyte nuclear leaf areas in the potential abnormal cell areas and marking each nuclear leaf; Positioning nuclear leaf connection points, calculating the positions, curvatures and connection angle distribution of the connection points, and generating the connection morphological characteristics of the nuclear leaf division of the leucocytes; and splicing and integrating the width distribution characteristics and the nuclear leaf splitting connection morphological characteristics to form a fine difference characteristic set.
6. The method according to claim 1, wherein the step S105 includes: the fine difference feature set is arranged into a standardized vector form, and a pre-trained deep learning model is input; Dividing the potential abnormal cell area into a plurality of subarea images of focused cell local structures through the deep learning model, extracting edge and texture information through a convolutional neural network, pooling and dimension reduction, and generating a preliminary feature map; extracting the distribution characteristics of the leaf separation angles of the white cell nuclear leaves and the morphological characteristics of vacuoles in the central light dyeing area of the red cell from the preliminary characteristic diagram; splicing and integrating the leaf-dividing angle distribution characteristics of the nuclear leaves and the morphological characteristics of the vacuoles in the light dyeing region to form an initial quantized pathological characteristic descriptor; applying principal component analysis to the initial quantized pathological feature descriptors to extract main dimensions, calculating variance contribution rates of the dimensions, and generating feature weight distribution; And carrying out weighted optimization on each component of the initial quantized pathological feature descriptor according to the feature weight distribution to obtain a final quantized pathological feature descriptor.
7. The method according to claim 1, wherein the step S106 includes: calculating the matching degree of the quantized pathological feature descriptors and corresponding feature vectors in a preset feature library by adopting a cosine similarity algorithm; if the matching degree is lower than a preset matching threshold, judging that a pathological abnormal signal exists; generating a preliminary abnormal cell identification result based on the pathological abnormal signal associated with the position coordinates of the potential abnormal cell area; Based on the primary abnormal cell identification result, extracting curvature and smoothness parameters of the leaf-separating edge of the nuclear leaf from the white cell area by utilizing an edge detection algorithm to obtain morphological characteristics of the leaf-separating edge of the nuclear leaf; Dividing the boundary of the central light dyeing region of the red blood cells by a region dividing method, and calculating the proportion of the area of the light dyeing region to the area of the whole cells to obtain the proportion distribution characteristic of the area of the light dyeing region; And carrying out quantitative description on the morphological characteristics of the leaf separation edge of the nuclear leaf and the area proportion distribution characteristics of the light dyeing region, and integrating to generate a complete abnormal cell identification result.
8. The method according to claim 1, wherein the step S107 includes: According to the abnormal cell identification result, extracting the area distribution characteristics of the central light staining area of the red blood cells in the blood smear microscopic image, and extracting the arrangement morphological characteristics of the white blood cell nuclear leaves; converting the area distribution characteristics into normalized vectors, calculating average distance and standard deviation of the arrangement morphological characteristics, and carrying out fusion treatment on the two types of characteristics by adopting a weighted summation mode; based on the fusion processing result, calculating the distribution characteristics of the gap between the nuclear leaves and the width distribution characteristics of the light dyeing region and the cytoplasmic transition region, and generating a comprehensive characteristic set; After the comprehensive feature set is subjected to standardization processing, a preset linear regression model is input, pathological correlation coefficients of all features are calculated, and a feature weight distribution scheme is generated; and carrying out weighted summation on all the features in the comprehensive feature set according to the feature weight distribution scheme to generate a comprehensive pathology quantification index.
9. A blood test sample-based pathology recognition analysis system for implementing the blood test sample-based pathology recognition analysis method of any one of claims 1 to 8, wherein the blood test sample-based pathology recognition analysis system comprises: the image acquisition module is used for acquiring a blood smear microscopic image of a blood detection sample, extracting cell area boundary information from the blood smear microscopic image and acquiring a preliminary cell morphology contour; The texture analysis module is used for carrying out texture analysis on the internal area of the preliminary cell morphology outline to obtain cell morphology distribution characteristics; The abnormality identification module is used for comprehensively analyzing the cell morphology distribution characteristics and judging whether a potential abnormal cell area exists or not; The local enhancement module is used for carrying out local image enhancement processing on the potential abnormal cell area and extracting a fine difference feature set; the depth analysis module is used for carrying out depth analysis on the potential abnormal cell area based on the fine difference feature set and combining a depth learning model to obtain a quantized pathological feature descriptor; The matching judgment module is used for judging whether a pathological abnormal signal exists or not and outputting an abnormal cell identification result according to the matching degree of the quantized pathological feature descriptor and a preset feature library; and the comprehensive evaluation module is used for comprehensively analyzing the blood smear microscopic image according to the abnormal cell identification result to obtain a comprehensive pathology quantification index.

Description

Pathological recognition analysis method and system based on blood detection sample Technical Field The application relates to the field of intelligent medical treatment, in particular to a pathology recognition analysis method and system based on a blood detection sample. Background The morphological analysis of blood cells is the basis of clinical examination and pathological diagnosis, and has a key role in screening, diagnosis and curative effect evaluation of blood system diseases, infectious diseases and malignant tumors. Traditionally, manual microscopic examination is relied on for cell morphology observation and abnormality identification, and the method is complex in flow, high in subjectivity and low in efficiency. With the development of digital pathology and artificial intelligence technology, an automatic cell analysis method based on image processing becomes an important research direction for improving objectivity, standardization and scale of pathology diagnosis. Currently, in the field of detection of cytopathological abnormalities based on image processing, the existing methods still have significant limitations. Firstly, in the aspects of cell segmentation and morphological contour extraction, due to the problems of cell adhesion, overlapping, uneven dyeing, complex background noise and the like in blood smear images, boundary fracture, excessive contour simplification or false segmentation phenomena are easily generated in common threshold segmentation, edge detection and other methods, so that the preliminary cell morphological contour extraction is inaccurate, and the accuracy of subsequent feature analysis is directly affected. Second, in morphological feature analysis and abnormality determination, most methods rely on single or shallow features (e.g., cell area, circularity, etc.), and have insufficient capture ability for fine morphological differences such as central light staining of erythrocytes, and leaf separation of leucocytes. Meanwhile, most of the feature extraction and abnormality judgment processes are cutting, a systematic fusion analysis framework from local details to global images is lacking, and distribution modes and correlations among cell populations are difficult to quantify, so that the recognition sensitivity to potential heterogeneous lesions (such as myelodysplastic syndrome) is low. Thirdly, in the aspects of pathology quantification and model generalization, the traditional machine learning classifier or fixed threshold rules are mostly adopted in the existing method, when facing clinical scenes with various cell morphologies and wide lesion lineages, the model has poor self-adaptation capability, the quantitative index has weak relevance with clinical pathology significance, and a comprehensive quantitative report with diagnostic value is difficult to generate. Aiming at the defects, the method solves the technical problems of low cell segmentation precision, weak fine difference capturing capability and incomplete pathology quantification of the traditional method by combining the accurate extraction of the multi-scale morphological contours, the deep learning analysis of the fine structural features and the multi-level feature fusion with quantitative modeling, and improves the accuracy, the automation degree and the clinical auxiliary diagnosis value of the cell pathological abnormality detection. Disclosure of Invention The application provides a pathology recognition analysis method and system based on a blood detection sample, which solve the technical problems of low cell segmentation precision, weak capability of capturing subtle differences and incomplete pathology quantification of the traditional method and improve the accuracy, the degree of automation and the clinical auxiliary diagnosis value of cell pathology abnormality detection. In a first aspect, the present application provides a method for pathology recognition analysis based on a blood test sample, the method comprising: S101, acquiring a blood smear microscopic image of a blood detection sample, and extracting cell area boundary information from the blood smear microscopic image to obtain a preliminary cell morphology contour; Step S102, performing texture analysis on the internal area of the outline of the primary cell morphology to obtain cell morphology distribution characteristics; step S103, comprehensively analyzing the morphological distribution characteristics of the cells, and judging whether a potential abnormal cell area exists or not; Step S104, carrying out local image enhancement processing on the potential abnormal cell area, and extracting a fine difference feature set; Step S105, carrying out deep analysis on the potential abnormal cell area based on the fine difference feature set and combining a deep learning model to obtain a quantized pathological feature descriptor; Step S106, judging whether a pathological abnormal signal exists or not according to the matchi