Breast cancer image data intelligent analysis and grading diagnosis platform based on large language model

Abstract

The invention relates to the technical field of medical image analysis, in particular to a breast cancer image data intelligent analysis and grading diagnosis platform based on a large language model, and is characterized in that an organization level feature segmentation module segments an image into various lesion areas based on breast image data through feature texts extracted from image description by the large language model; and mapping to a hierarchical path of the large language model to obtain an image feature hierarchical mapping result. According to the method, through multi-dimensional means such as feature extraction and hierarchical mapping, the breast image data is subdivided into a plurality of feature hierarchies, and the feature relationship of the focus area is constructed, so that multi-level refinement in a focus diagnosis path is realized. By identifying the change trend and the abnormal point of each level feature in the focus, the potential focus area can be more accurately locked, the focus feature identification process is accelerated, and effective transmission of information and comprehensive recording of the focus structure are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model. Background Art

[0002] Medical image analysis is a technology that uses computer vision, machine learning and artificial intelligence technologies to analyze medical image data. It mainly covers the automated processing and analysis of various medical images, such as X-rays, ultrasound, CT, MRI and PET scans. Through the automated analysis of these image data, medical image analysis technology can assist doctors in making diagnosis and treatment decisions more quickly and accurately.

[0003] Among them, the breast cancer imaging data intelligent analysis and hierarchical diagnosis platform is specifically used for intelligent analysis and hierarchical diagnosis of breast cancer imaging data. The platform aims to conduct in-depth analysis of breast cancer imaging data through machine learning and artificial intelligence technology, so as to automatically identify, locate and grade breast lesions. It can assist doctors in screening and early diagnosis of breast cancer and provide a reference for hierarchical diagnosis.

[0004] Although the existing technology has basic recognition, positioning and grading functions in the automated analysis of breast images, it is difficult to effectively capture the complex feature differences in the lesion area, which leads to limitations in deep feature extraction and lesion abnormality recognition. A single image grading path is often unable to cope with the complexity of multi-lesion and multi-level pathological characteristics, which may lead to the omission of information in some lesion areas. In addition, the existing technology is difficult to adapt to the changing trends of lesion characteristics in real time in lesion classification and abnormality marking, which may cause a lag in the recognition of certain new lesion types. In addition, errors are prone to occur in the judgment of lesion boundaries and risk levels, affecting the overall diagnostic consistency and accuracy, and increasing the doctor's interpretation burden and risk of misjudgment. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and propose an intelligent analysis and hierarchical diagnosis platform for breast cancer image data based on a large language model.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model, the platform comprising:

[0007] The tissue-level feature segmentation module is based on breast imaging data. It uses the feature text extracted from the image description by the large language model to segment the image into multiple lesion areas, map them to the hierarchical path of the large language model, and obtain the image feature hierarchical mapping results.

[0008] The image feature path generation module identifies the feature conversion trend in each level according to the image feature hierarchical mapping result, marks the abnormal feature set where the edge or density abnormality area appears as the lesion, and obtains the pathological diagnosis path information;

[0009] The lesion feature similarity transfer module selects key features of the lesion area in the lesion abnormal feature set based on the pathological diagnosis path information, compares the key features with the unlabeled area item by item, records the newly identified lesion area structural features, and generates lesion structural feature transfer data;

[0010] The lesion structure classification processing module classifies the features of each lesion area based on the lesion structure feature transfer data, dynamically updates the lesion category structure, and obtains a lesion structure update result;

[0011] The regional feature probability analysis module updates the lesion structure result, calculates the feature probability distribution of the lesion, determines the high probability area by referring to the feature occurrence frequency and distribution range, and forms a lesion abnormality probability distribution map;

[0012] The lesion fuzzy grading module identifies the boundary fuzzy features of each lesion area based on the lesion abnormality probability distribution map, analyzes the risk level of each area, and obtains the lesion fuzzy grading result.

[0013] The present invention has the following improvements: the image feature hierarchical mapping result includes an edge smoothness description of a common tissue region, a morphological heterogeneity description of a suspected lesion region, and a density increase feature of a determined lesion region; the pathological diagnosis path information includes a transition feature description from a common tissue to a suspected lesion, a morphological transformation description from a suspected lesion to a lesion, and a density feature path of the lesion; the lesion structure feature transfer data includes an edge roughness parameter, a texture irregularity parameter, and a density contrast analysis of a newly discovered lesion region; the lesion structure update result includes a classification of lesions with clearly defined edges, a classification of lesions with irregular morphology, and a classification of lesions with rough textures; the lesion abnormality probability distribution map includes a division of a density concentration region with a probability value higher than a predetermined threshold, a heterogeneous region with abnormal morphology, and a lesion region with prominent edge roughness; the lesion fuzzy grading result includes a membership evaluation of an unclear edge region, a membership evaluation of a mixed texture region, and a fuzzy membership evaluation of an irregular morphology region.

[0014] The present invention is improved in that the organizational level feature segmentation module includes:

[0015] The tissue region division submodule divides the tissue regions in the image based on breast image data, extracts tissue feature text in the image through a large language model, divides the image into common tissue, suspected lesion area and confirmed lesion area, and obtains the tissue region division result;

[0016] The regional feature analysis submodule analyzes the edge features, tissue morphology features and density distribution information of each region based on the tissue region division results, extracts the edge smoothness, density distribution and morphological change details of the partition, and generates regional feature analysis data;

[0017] The feature hierarchical mapping submodule maps the regional features to the hierarchical path of the large language model based on the regional feature analysis data, establishes a corresponding relationship between the features of the organizational level and the hierarchical path, and obtains the image feature hierarchical mapping result.

[0018] The present invention is improved in that the image feature path generation module includes:

[0019] The feature sequence construction submodule establishes a feature description path layer by layer according to the hierarchical order from common tissue to lesion area based on the image feature hierarchical mapping result, describes the order of tissue features at each level, and generates a tissue hierarchical feature list;

[0020] The feature trend identification submodule identifies the change trend of the features within the hierarchy based on the tissue hierarchy feature list, identifies the transformation of the tissue morphology, density and edge features by analyzing the structural transformation from the normal to the diseased area, and generates feature transformation trend information;

[0021] Based on the feature conversion trend information, the abnormal feature marking submodule marks the lesion feature set for the area with edge and density abnormalities in the mapping path, describes the distribution and position of the lesion feature set, and obtains the pathological diagnosis path information.

[0022] The present invention is improved in that the lesion feature similarity transfer module comprises:

[0023] The abnormal feature extraction submodule extracts the abnormal edge contour, abnormal texture and abnormal density features of the lesion area based on the pathological diagnosis path information, analyzes the numerical distribution and local differences of the density of the lesion area, and generates abnormal features of the lesion;

[0024] The similar region identification submodule performs feature comparison on the unlabeled region based on the abnormal features of the lesion, detects whether the edge contour morphology of the unlabeled region matches the lesion region, analyzes the density distribution of the unlabeled region and identifies the region with the smallest density difference with the lesion region, and obtains potential similar region information;

[0025] The potential lesion marking submodule verifies that the density difference range meets the lesion region characteristics based on the potential similar region information, marks the unmarked region that meets the requirements as a potential lesion region, records the structural feature information of the region, and obtains lesion structural feature transfer data.

[0026] The present invention is improved in that the lesion structure classification processing module comprises:

[0027] The lesion feature analysis submodule identifies texture similarity based on the lesion structure feature transfer data according to the local distribution of texture features, and extracts density difference data according to the data distribution of density features to obtain lesion feature similarity data;

[0028] The lesion classification submodule classifies regions with similar edge features into the same category based on the lesion feature similarity data, and classifies them by comparing the distribution similarity of texture features to generate a lesion classification result;

[0029] The lesion category updating submodule identifies the edge, texture and density features of the new lesion area based on the lesion category classification result, updates the area matching the features of the marked lesion area to the current category, and obtains a lesion structure update result.

[0030] The present invention is improved in that the regional feature probability analysis module comprises:

[0031] The density feature grading submodule, based on the lesion structure update result, grades the density and morphological features of the lesion area according to the density value and morphology to generate graded feature data;

[0032] The feature probability statistics submodule calculates the frequency distribution of hierarchical features based on the hierarchical feature data, calculates the feature occurrence probability value of each level, integrates the calculation results into probability distribution data, and generates lesion feature probability distribution information;

[0033] The target probability region marking submodule determines the region where the multi-frequency feature is located based on the lesion feature probability distribution information and the feature occurrence frequency and distribution range, marks the high-probability feature region as a potential risk region, and forms a lesion abnormality probability distribution map.

[0034] The present invention is improved in that the formula is used to calculate the probability of occurrence of features at each level:

[0035]

[0036] Get the probability value P(x) of the i-th level feature i );

[0037] Among them, f i is the frequency of occurrence of the feature in the i-th level, M is the total number of levels, and f j Represents the frequency of occurrence of features in the jth level.

[0038] The present invention is improved in that the lesion fuzzy grading module comprises:

[0039] The boundary fuzzy recognition submodule detects the boundary fuzzy features of the lesion area based on the lesion abnormality probability distribution map, analyzes the fuzziness degree and transition smoothness of the boundary, and obtains the fuzzy boundary features;

[0040] The fuzzy membership calculation submodule divides the density uneven area into membership levels according to the fuzzy degree based on the fuzzy boundary features, calculates the membership value of the area, and assigns it to the corresponding fuzzy membership level to obtain the fuzzy membership value of the lesion area;

[0041] The risk level classification submodule evaluates the risk level of density fluctuation and boundary fuzzy features based on the fuzzy membership value of the lesion area, divides the area into risk levels, and generates fuzzy grading results for the lesions.

[0042] The present invention has the following improvements: for dividing the density uneven area into membership levels according to the degree of fuzziness, the formula is adopted:

[0043]

[0044] Get the fuzzy membership value μ(x′);

[0045] Among them, x′ is the fuzziness parameter of the current area, x′ min is the minimum ambiguity value in the entire analysis area, x′ max is the maximum fuzziness value in the entire analysis area, y is the boundary transition smoothness parameter of the current area, and y min is the minimum boundary smoothness value in the entire analysis area, y max is the maximum value of boundary smoothness in the entire analysis area, and α and β are constant coefficients.

[0046] Compared with the prior art, the advantages and positive effects of the present invention are:

[0047] In the present invention, through multi-dimensional means such as feature extraction and hierarchical mapping, breast imaging data is subdivided into multiple feature levels and the feature relationship of the lesion area is constructed, so as to realize multi-level refinement in the lesion diagnosis path. By identifying the changing trend and abnormal points of the features of each level in the lesion, it is helpful to lock the potential lesion area more accurately and accelerate the lesion feature recognition process, so as to ensure the effective transmission of information and the comprehensive record of the lesion structure. The feature similarity transfer process can effectively discover and improve the hidden lesion information by comparing the unmarked area while expanding the existing marked lesion features, thereby improving the comprehensiveness of lesion detection and the accuracy of features. At the same time, the lesion category structure is dynamically updated, and the classification and adjustment of lesion features are gradually optimized, so as to maintain the real-time and adaptability of image data analysis, generate an abnormal probability map based on feature frequency and distribution range, and more clearly present the feature distribution of high-risk areas, so as to provide doctors with an intuitive reference for abnormal distribution of lesions and reduce the risk of misdiagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a module diagram of the platform of the present invention;

[0049] Figure 2 It is a framework diagram of the present invention;

[0050] Figure 3 It is a schematic diagram of the organizational level feature segmentation module of the present invention;

[0051] Figure 4 A schematic diagram of an image feature path generation module of the present invention;

[0052] Figure 5 A schematic diagram of a transfer module with similar lesion characteristics according to the present invention;

[0053] Figure 6 Schematic diagram of the lesion structure classification processing module of the present invention;

[0054] Figure 7 It is a schematic diagram of the regional feature probability analysis module of the present invention;

[0055] Figure 8 It is a schematic diagram of the lesion fuzzy grading module of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0057] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating positions or positional relationships, are based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0058] See also Figure 1 The present invention provides a technical solution: a breast cancer imaging data intelligent analysis and hierarchical diagnosis platform based on a large language model, the platform comprising:

[0059] The tissue-level feature segmentation module is based on breast imaging data. It uses the feature text extracted from the image description by the large language model to segment the image into multiple lesion areas, map them to the hierarchical path of the large language model, and obtain the image feature hierarchical mapping results.

[0060] The image feature path generation module identifies the feature conversion trend within each level based on the image feature hierarchical mapping results, marks the abnormal feature set where the edge or density abnormality area appears as the lesion, and obtains the pathological diagnosis path information;

[0061] The lesion feature similarity transfer module selects the key features of the lesion area in the abnormal feature concentration of the lesion based on the pathological diagnosis path information, compares the key features with the unlabeled area one by one, records the structural features of the newly identified lesion area, and generates lesion structural feature transfer data;

[0062] The lesion structure classification processing module classifies the features of each lesion area based on the lesion structure feature transfer data, dynamically updates the lesion category structure, and obtains the lesion structure update result;

[0063] The regional feature probability analysis module updates the lesion structure results, calculates the feature probability distribution of the lesion, determines the high probability area by referring to the feature occurrence frequency and distribution range, and forms a lesion abnormality probability distribution map;

[0064] The lesion fuzzy grading module identifies the boundary fuzzy features of each lesion area based on the lesion abnormality probability distribution map, analyzes the risk level of each area, and obtains the lesion fuzzy grading results;

[0065] The image feature hierarchical mapping results include the edge smoothness description of the common tissue area, the morphological heterogeneity description of the suspected lesion area, and the density increase characteristics of the confirmed lesion area. The pathological diagnosis path information includes the transition feature description from common tissue to suspected lesions, the morphological transformation description from suspected lesions to lesions, and the density feature path of the lesion. The lesion structure feature transfer data includes the edge roughness parameters, texture irregularity parameters, and density contrast analysis of the newly discovered lesion area. The lesion structure update results include the classification of lesions with clear edge definition, irregular morphology, and rough texture. The probability distribution map of lesion abnormality includes the division of density concentration areas with probability values ​​higher than the predetermined threshold, heterogeneous areas with abnormal morphology, and lesion areas with prominent edge roughness. The lesion fuzzy grading results include the membership evaluation of unclear edge areas, the membership evaluation of mixed texture areas, and the fuzzy membership evaluation of irregular morphology areas.

[0066] See also Figure 2 and Figure 3 ,The organizational hierarchical feature segmentation module includes:

[0067] The tissue region division submodule divides the tissue regions in the image based on breast image data, extracts tissue feature text in the image through a large language model, divides the image into common tissue, suspected lesion area and confirmed lesion area, and obtains the tissue region division result;

[0068] Based on breast imaging data, tissue region division is performed. First, multiple potential tissue regions in the image are obtained through multi-scale image segmentation methods. For example, the density distribution in the image is clustered, and blocks of regions with similar density are determined and marked as potential tissue regions. Then, the specific contours of these regions are divided using edge detection methods. On this basis, the large language model is applied to parse the text feature information of each region one by one. For example, the model is used to extract the pixel density, morphological description, boundary features, etc. of each region in a textual manner, and these feature information are converted into structured text data and further quantified into numerical features. Specifically, after the model parses the description of "high-density circular tissue", it is converted into a quantitative data format of "density value = high, shape = circular", and feature encoding is used to generate independent feature vectors for each region for subsequent processing. Finally, the text feature data of all extracted regions are integrated to obtain the quantitative feature set of the image, which is used for further analysis to form the tissue region division result.

[0069] The regional feature analysis submodule analyzes the edge features, tissue morphology features and density distribution information of each region based on the tissue region division results, extracts the edge smoothness, density distribution and morphological change details of the partition, and generates regional feature analysis data;

[0070] First, edge detection is applied to obtain the boundary morphological features of the region. For example, the Sobel operator is used to extract the boundary gradient change to obtain the integrity and smoothness data of the boundary, and the morphological features of the edge are quantified, such as the roundness or irregularity of the shape, which are recorded as edge feature values ​​after extraction. Next, the tissue morphological features are processed, and the morphological structure analysis method is used to obtain the internal features of the region. For example, the regional area and shape complexity in morphological image processing are analyzed and quantified into feature vectors. In the process of extracting density features, the density distribution data of the region is obtained based on the grayscale histogram, and the distribution ratio of high and low density points is classified into high and low levels, and the density extreme value and average density value are extracted. These data are integrated to form a density feature set. Finally, the edge features, morphological structure features and density features are summarized to form regional feature analysis data.

[0071] The feature hierarchical mapping submodule maps the regional features to the hierarchical path of the large language model based on the regional feature analysis data, establishes the correspondence between the features of the organizational level and the hierarchical path, and obtains the image feature hierarchical mapping result;

[0072] First, the morphology, density and edge features of each region are matched separately, and the corresponding relationship between these features and the nodes of the hierarchical path of the large language model is established step by step. For example, for a region with high density and smooth edges, its density feature is first mapped to the high-density node of the hierarchical path, then the region is mapped to the "edge complete" node according to the edge smoothness, and finally the morphology (such as circle or ellipse) is compared and matched with the morphological classification node of the large language model. In the matching process, a similarity comparison method, such as Cosine similarity, is used to quantify the similarity value between the regional features and the hierarchical path nodes, and the nodes with higher matching degree are used as the hierarchical position of the feature. For example, assuming that the density distribution value of a region is 80 (high density), the edge smoothness is 90 (high smoothness), and the morphology is described as "ellipse". After mapping, the density feature of the region is associated with the high-density path node, the edge feature is associated with the "edge smoothness" node, and the morphological feature is mapped to the "ellipse" node. Finally, after all matching is completed, a feature path containing density, edge and morphology is generated, and these mapping relationships are integrated to obtain the hierarchical mapping result of image features.

[0073] See also Figure 2 and Figure 4 , the image feature path generation module includes:

[0074] The feature sequence construction submodule is based on the image feature hierarchical mapping results, establishes a feature description path layer by layer according to the hierarchical order from common tissue to the lesion area, describes the order of tissue features at each level, and generates a tissue hierarchical feature list;

[0075] By arranging and parsing the feature data of each level from low to high level, a description path is gradually formed. For example, in the normal tissue area, low density (such as density value between 10-30) and high edge smoothness (smoothness index is above 90) are the main features; in the lesion area, the density feature gradually increases to medium density (such as density value between 40-60), and the edge gradually changes irregularly (smoothness index drops to between 60-80), accompanied by a change in morphology from regular to irregular (such as a circle becomes an irregular polygon). First, the low density and smooth edge features of the normal tissue level are recorded as the basic level, and then the features such as increased density, rough edges and morphological changes in the lesion area are added to ensure that the feature order maintains an increasing relationship from the basic level to the advanced level. The feature description of each level is recorded in the corresponding feature description path, and the complete list of tissue level features is summarized in turn to describe the feature evolution path at different levels.

[0076] The feature trend recognition submodule identifies the changing trend of features within the hierarchy based on the tissue hierarchy feature list, identifies the transformation of tissue morphology, density and edge features by analyzing the structural transformation from normal to diseased areas, and generates feature transformation trend information;

[0077] By comparing the features of each level layer by layer, the transformation process of the features can be identified. For example, key features such as morphology, density and edge are extracted from the feature list of normal tissue to the lesion area. First, the morphological features of the normal tissue are described as regular circles with low density (density values ​​are between 10-30) and high edge smoothness (edge ​​smoothness is 90-100); then, the morphological features of the lesion area are gradually transformed from regular to elliptical or irregular shapes, the density increases to medium (density values ​​are between 40-60), and the edge smoothness decreases to below 60. For this continuous change process, the increase, decrease and change of the features are recorded step by step. For example, in the edge feature, the change process from smooth to rough is recorded, and the density records the gradual change from low to high. By summarizing the change trajectory of each layer of features, the feature conversion path description from normal to lesion area is established step by step, and the feature conversion trend information is obtained based on the layer-by-layer transformation of features.

[0078] The abnormal feature marking submodule marks the lesion feature set for the areas with edge and density abnormalities in the mapping path based on the feature conversion trend information, describes the distribution and location of the lesion feature set, and obtains the pathological diagnosis path information;

[0079] First, based on the high-frequency abnormal features in the feature conversion trend information, screen and determine the abnormal areas. For example, by comparing the conversion trend path, identify areas with significantly increased density (such as density values ​​higher than 60) and significantly increased edge roughness (smoothness lower than 40) as abnormal features. Mark these feature positions in the lesion feature set. For example, if the density of a certain area increases beyond the normal increase range in the trend path, or the edge roughness is significantly higher than the corresponding value in the trend path, the area is marked as an abnormal lesion area. Finally, mark all abnormal features in the image feature distribution map, record their positions and distribution, and generate pathological diagnosis path information based on the distribution.

[0080] See also Figure 2 and Figure 5 , the lesion feature similarity transfer module includes:

[0081] The abnormal feature extraction submodule extracts the abnormal edge contour, abnormal texture and abnormal density features of the lesion area based on the pathological diagnosis path information, analyzes the numerical distribution and local differences of the density of the lesion area, and generates abnormal features of the lesion;

[0082] By gradually analyzing the edge contour morphology in the image data, the abnormal contour data of the edge is refined into angle and curvature values. Specifically, the changes in edge morphology are graded in units of 10 degrees, and the angle and curvature values ​​of each turning point are recorded to obtain the texture features in the region, such as by calculating the consistency index of the texture direction, and standardizing it to a numerical representation from 0 (no consistency) to 1 (complete consistency). For example, if the texture consistency of a lesion area is 0.8 and the fineness is higher than 75%, it is judged that its texture is significantly abnormal. When further analyzing the density features, the local differences in density in the region are statistically analyzed by calculating the difference between the average density value of the region and the surrounding area. For example, if the density value of the lesion area reaches 80 and the mean value of the surrounding area is 50, the density difference reaches 30, which significantly exceeds the common range of 5-10, indicating that there is a density abnormality. Finally, the abnormal indicators such as edge angle changes, texture consistency and density differences are summarized, and the abnormal feature data set of the area is generated to obtain the abnormal features of the lesion.

[0083] The similar region recognition submodule compares the features of the unlabeled regions based on the abnormal features of the lesions, detects whether the edge contour of the unlabeled regions matches the lesion region, analyzes the density distribution of the unlabeled regions and identifies the region with the smallest density difference with the lesion region, and obtains the potential similar region information;

[0084] First, extract the edge contour morphological data of each unlabeled area, obtain the edge curvature and continuity through edge detection, and compare them with the abnormal contour features of the lesion area. For example, if the average edge curvature of the lesion area is 0.4 and the continuity is high, then prioritize the identification of unlabeled areas close to these features. Then, analyze the density distribution data of the unlabeled area, perform a difference analysis on its density value and the lesion area, and screen areas with a density difference of less than 10. If the density value of an unlabeled area is between 60-70 and the edge curvature is between 0.35 and 0.45, which is close to the characteristics of the lesion area, it is considered to have a high similarity. Use these criteria to screen out qualified unlabeled areas and integrate them into potential similar area information.

[0085] The potential lesion marking submodule verifies that the density difference range meets the characteristics of the lesion area based on the potential similar area information, marks the unmarked area that meets the requirements as the potential lesion area, records the structural feature information of the area, and obtains the lesion structural feature transfer data;

[0086] By controlling the difference between the density value of each potential similar area and the density of the lesion area within the set threshold range. For example, if the density value of the lesion area is 75 and the density difference range is set to ±5, the area with a density value between 70-80 is screened as the area that meets the density feature. Further, according to the similarity of the edge features, MATLAB is used to obtain the mutual correlation coefficient of the edge contour of the unlabeled area and the lesion area, and the similarity between the two is calculated by the xcorr2 function. In MATLAB, the xcorr2 function can directly process two two-dimensional matrices, and extract the mutual correlation coefficient by calculating the maximum value in the mutual correlation matrix to judge the similarity. When the result of the mutual correlation coefficient is 1, it means complete similarity, and when the result is -1, it means complete dissimilarity. If the mutual correlation coefficient of a region is greater than 0.8, it is considered that the edge feature of the region is highly similar to the lesion area. For example, if the density of an unlabeled area is 77 and the mutual correlation coefficient of the edge is 0.85, it meets the requirements of the lesion feature and is marked as a potential lesion area. Finally, the structural feature information of these areas is recorded, and all qualified areas are marked to form the lesion structural feature transfer data.

[0087] See also Figure 2 and Figure 6 , the lesion structure classification processing module includes:

[0088] The lesion feature analysis submodule transfers the data based on the lesion structure features, identifies the texture similarity according to the local distribution of the texture features, and extracts the density difference data according to the data distribution of the density features to obtain the lesion feature similarity data;

[0089] The gray-level co-occurrence matrix is ​​used to integrate contrast, energy, and uniformity parameters to analyze the texture features of the lesion area layer by layer. For example, the lesion area is divided into multiple local areas, and starting from each small area, the contrast (characterizing the intensity of texture change), energy (representing the uniformity of texture), and uniformity (reflecting the consistency of texture) of the gray-level co-occurrence matrix are obtained. In a certain local area, if the contrast is high (for example, above 0.5), it shows that the texture in the area changes dramatically; at the same time, if the energy and uniformity are low, it indicates that the texture is rough and inconsistent. Through such parameters, it is gradually extended from the local to the whole to the entire lesion area, and the differences in texture features between local areas are identified. For example, when the contrast in the central area of ​​the lesion is high and the energy and uniformity are low, it indicates that there may be obvious texture abnormalities in the central area, and then it is judged that the area may be a high-risk lesion area, and the lesion feature similarity data is obtained.

[0090] The lesion classification submodule classifies the areas with similar edge features into the same category based on the lesion feature similarity data, and classifies them by comparing the distribution similarity of texture features to generate lesion classification results;

[0091] By comparing the curvature and edge angle change features, areas with consistent edge morphology are identified and classified into the same category. For example, the curvature change of the edge of the lesion area is obtained, and the smooth edge with a curvature between 0.4 and 0.5 and a small angle change (such as within 5 degrees) is observed, then the area has a smoother edge feature. Further screening of the edge features of unclassified areas, if the curvature value and angle change of the new area are similar to those of the known lesion area (such as the curvature is also in the range of 0.4 to 0.5, and the angle change is within 5 degrees), then the edge morphology of these areas is determined to be similar, and they are classified into the same category for further analysis of the consistency of the lesion edge, to confirm the structural characteristics of the lesion edge, and to generate lesion category classification results.

[0092] The lesion category update submodule identifies the edge, texture, and density features of the new lesion area based on the lesion category classification results, updates the area matching the features of the marked lesion area to the current category, and obtains the lesion structure update result;

[0093] By mapping the feature data of the newly identified lesion area into the feature space, the Euclidean distance between the newly identified lesion area and the classified lesion category features is calculated using MATLAB software to determine the category affiliation. The edge curvature, texture consistency and density features of the new area are quantified into feature vectors and compared with the average feature vectors of the existing lesion categories. For example, the feature vector of the classified lesion category is [0.4, 0.75, 50] (representing the mean of curvature, texture consistency and density features respectively), while the feature vector of the new area is [0.42, 0.7, 52]. In the similarity judgment, the distance between the new area obtained by the Euclidean distance method and the category is compared with the preset similarity threshold range. If the distance is within the threshold range (for example, the threshold is set to 0.1), the new area is considered to be close enough to the category feature and thus classified into the category. For example, if the calculated distance is 0.08, which is less than the threshold 0.1, the area is considered to meet the feature standards of the category, classified into the category, and its label is updated so that the classification, integration and analysis of the lesion structure can be further performed to obtain the lesion structure update result.

[0094] See also Figure 2 and Figure 7 , the regional characteristic probability analysis module includes:

[0095] The density feature grading submodule is based on the lesion structure update result, and divides the density and morphological characteristics of the lesion area into grades according to the density value and morphology to generate graded feature data;

[0096] First, the density value data in the lesion area was obtained using ImageJ software, and the density was divided into three levels: low, medium, and high. The density value of each area was classified into the corresponding level, for example, low density was defined as 0-30, medium density was 31-70, and high density was above 71. Then, according to the morphological characteristics, the structure in the image was layered and refined according to the complexity of the regional morphology, and the morphology was divided into three levels: simple, regular, and irregular. The morphological characteristics were evaluated region by region using image analysis tools, such as edge detection to analyze the regularity and complexity of the morphology. If the shape contour of a certain area is relatively smooth and regular, it is classified as a regular level; if the contour is irregular and has multiple protrusions, it is marked as an irregular level. The density and morphological feature data of all areas are summarized by level to generate hierarchical feature data, laying the data foundation for the next step of probability statistics and distribution analysis.

[0097] The feature probability statistics submodule is based on the hierarchical feature data, statistically analyzes the frequency distribution of hierarchical features, calculates the probability of occurrence of features at each level, integrates the calculation results into probability distribution data, and generates lesion feature probability distribution information;

[0098] To calculate the probability of occurrence of features at each level, the formula is used:

[0099]

[0100] Get the probability value P(x) of the i-th level feature i );

[0101] Among them, f i is the frequency of occurrence of features in the i-th level, obtained by analyzing the image data statistics, M is the total number of levels, obtained by setting the hierarchical level of the image data, and f j Represents the frequency of occurrence of features in the jth level.

[0102] If the image classification contains two levels, their occurrence frequencies are counted separately. Assuming that the characteristic frequency of the first level is f1=15 and the characteristic frequency of the second level is f2=35, the total frequency is:

[0103]

[0104] Substitute the frequency of features at each level into the formula, calculate the probability of features at each level, and convert the result into a percentage:

[0105] For the first level:

[0106]

[0107] For the second level:

[0108]

[0109] The results show that the first level feature has a 30% probability of occurrence, while the second level feature has a 70% probability of occurrence.

[0110] The target probability region marking submodule determines the region where the multi-frequency features are located based on the probability distribution information of the lesion features according to the frequency of occurrence and distribution range of the features, marks the high-probability feature region as a potential risk region, and forms a probability distribution map of the lesion abnormality;

[0111] First, summarize the frequency of features in each region and calculate the probability value of each level of features, and bring the obtained probability results into it. For example, if the probability value of the first level feature is 30% and the probability of the second level feature is 70%, then mark the occurrence probabilities of these levels on the distribution map in the lesion area. Subsequently, use the spatial distribution of high-frequency features as a reference to visualize these probability data on the image, and use color or heat map to mark high-probability areas to highlight multi-frequency feature areas. Through this visualization, the high probability distribution of multi-frequency features can be shown in the image, and finally generate a probability distribution map of lesion abnormalities for further diagnostic analysis and risk area positioning.

[0112] See also Figure 2 and Figure 8 , the lesion fuzzy grading module includes:

[0113] The boundary fuzzy recognition submodule detects the boundary fuzzy features of the lesion area based on the probability distribution map of lesion abnormalities, analyzes the fuzziness degree and transition smoothness of the boundary, and obtains the fuzzy boundary features;

[0114] First, the image data is preprocessed, and the noise interference in the data is reduced by the noise reduction method to ensure the accuracy of the image details. Then, the grayscale gradient detection technology is used to obtain the grayscale change rate at the edge of the lesion, and the blur degree of the regional boundary is further analyzed. If the grayscale gradient change rate of the boundary area is small, it indicates that the boundary has a high blur feature. In order to analyze the transition smoothness of the boundary, after the boundary detection, the boundary gradient data is interpolated to calculate the smoothness of the boundary change. The areas where the detected gradient change rate is lower than the set threshold are marked as fuzzy areas, and the boundary features of these fuzzy areas are extracted. Finally, through data screening and classification, a set of fuzzy boundary features containing all fuzzy area information is formed.

[0115] The fuzzy membership calculation submodule divides the density uneven area into membership levels according to the degree of fuzziness based on the fuzzy boundary features, calculates the membership value of the area, and assigns it to the corresponding fuzzy membership level to obtain the fuzzy membership value of the lesion area;

[0116] For dividing the density uneven areas into membership levels according to the degree of fuzziness, the formula is used:

[0117]

[0118] Get the fuzzy membership value μ(x′);

[0119] Among them, μ(x′) represents the relative position of the current area within the fuzzy membership range, and evaluates the fuzziness of the area. x′ is the fuzziness parameter of the current area, which characterizes the fuzzy characteristics of the area. The acquisition method is to detect the grayscale gradient of the area in the image. Usually, the rate of change of the grayscale gradient is measured as the fuzziness through image processing technology. x′ min It is the minimum value of blur in all analysis areas, indicating the lower limit of blur in image data. By measuring and counting blur characteristics of all areas, the minimum value is selected as the reference lower limit of blur. max is the maximum value of blur in all analyzed areas, indicating the upper limit of blur in image data. It is obtained by measuring blur characteristics of all areas and taking the highest value. y is the boundary transition smoothness parameter of the current area, describing the smoothness characteristics of the area boundary. The smoothness value is measured by analyzing the gray gradient change rate of the area boundary. The higher the smoothness value, the more natural the boundary transition. min It is the minimum value of boundary smoothness in all analysis areas, indicating the lower limit of smoothness in image data. It is measured by counting the boundary smoothness of all areas. max It is the maximum value of boundary smoothness in the entire analysis area, indicating the upper limit of smoothness in the image data. It is obtained by counting the smoothness data of all areas. α and β are constant coefficients used to adjust the relative contribution of blur and smoothness when calculating membership values. They are usually set to 1 to ensure that blur and smoothness have equal weights in the calculation.

[0120] For example, the current region fuzziness x′=8 and the boundary transition smoothness y=5; the minimum fuzziness x′ min =2, maximum value x′ max =10; minimum smoothness value y min =1, maximum value y max =8:

[0121] Substituting into the formula, the fuzzy membership value is calculated as follows:

[0122]

[0123] The results show that the fuzzy membership value of the current area is 0.67. Based on the calculated fuzzy membership value μ(x), the density uneven area is graded and divided into different levels of fuzziness. The membership values ​​of all areas are sorted by size. The preset fuzzy membership level threshold is set to divide the membership values ​​into multiple levels, such as low fuzziness, medium fuzziness and high fuzziness. The specific division process is as follows: Sort the membership values ​​of all areas in ascending order. Assuming that the total membership value range is from 0 to 1, the division thresholds are 0.3 (low fuzzy threshold) and 0.7 (high fuzzy threshold). This divides the membership value range into three levels: Low fuzziness: areas with membership values ​​less than 0.3. Medium fuzziness: areas with membership values ​​between 0.3 and 0.7. High fuzziness: areas with membership values ​​greater than 0.7. For example, if the fuzzy membership value of a region is 0.15, which is lower than 0.3, it will be divided into the low fuzzy membership level, indicating that the fuzziness of the region is low and the boundary is clear. If the membership value of another area is 0.55, which is between 0.3 and 0.7, it is classified into the medium fuzzy membership level, which is used to identify areas with moderate fuzziness. For the area with a membership value of 0.85, since it is higher than 0.7, it is assigned to the high fuzzy membership level, indicating that the area is significantly blurred and the boundary may not be clear, which requires further attention. All areas form a hierarchical structure with different fuzzy membership levels based on the allocation of their membership values. The low fuzzy level represents areas with clearer boundaries, the medium fuzzy level represents areas with moderately blurred boundaries, and the high fuzzy level reflects areas with significant fuzziness. This hierarchical structure provides a clear basis for subsequent lesion risk assessment and diagnosis, allowing areas with different fuzzy characteristics to be intuitively classified for further analysis and processing.

[0124] The risk level classification submodule evaluates the risk level of density fluctuation and boundary fuzzy features based on the fuzzy membership value of the lesion area, divides the area into risk levels, and generates fuzzy grading results of the lesion;

[0125] Based on the fuzzy membership value of the lesion area and the different fuzzy membership levels, the risk level is divided by evaluating the density fluctuation and boundary fuzzy features. First, the areas are grouped according to the fuzzy membership level, the density fluctuation range of each area in each group is obtained, and these fluctuation values ​​are counted. Density fluctuation is detected by the change amplitude of the pixel gray value in the image data, and its fluctuation amplitude is recorded by level, and the boundary fuzzy membership value of the area in each group is compared. For low fuzzy membership levels (such as areas with membership values ​​below 0.3), if the regional density fluctuation amplitude is small, the area is divided into a low risk level. Then, for the medium fuzzy membership level (areas with membership values ​​between 0.3 and 0.7), the medium risk level is further confirmed by the area with higher density fluctuation, and the area with medium density fluctuation can be regarded as a potential lesion area. Finally, the high fuzzy membership level (areas with membership values ​​greater than 0.7) is detected, and the areas with large density fluctuation and high fuzzy membership values ​​are divided into high risk levels and marked as areas with higher lesion risks. Finally, a stratified risk grading result of the lesion area is formed, and a fuzzy grading structure of lesions divided into low, medium and high risk levels is obtained, which provides a grading basis for further diagnosis and analysis of the lesion area in the imaging data.

[0126] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A breast cancer imaging data intelligent analysis and hierarchical diagnosis platform based on a large language model, characterized by: The platform includes: The tissue-level feature segmentation module is based on breast imaging data. It uses the feature text extracted from the image description by the large language model to segment the image into multiple lesion areas, map them to the hierarchical path of the large language model, and obtain the image feature hierarchical mapping results. The image feature path generation module identifies the feature conversion trend in each level according to the image feature hierarchical mapping result, marks the abnormal feature set where the edge or density abnormality area appears as the lesion, and obtains the pathological diagnosis path information; The lesion feature similarity transfer module selects key features of the lesion area in the lesion abnormality feature set based on the pathological diagnosis path information, compares the key features with the unlabeled area item by item, records the newly identified lesion area structural features, and generates lesion structural feature transfer data; The lesion structure classification processing module classifies the features of each lesion area based on the lesion structure feature transfer data, dynamically updates the lesion category structure, and obtains a lesion structure update result; The regional feature probability analysis module updates the lesion structure, calculates the feature probability distribution of the lesion, determines the high probability area by referring to the feature occurrence frequency and distribution range, and forms a lesion abnormality probability distribution map; The lesion fuzzy grading module identifies the boundary fuzzy features of each lesion area based on the lesion abnormality probability distribution map, analyzes the risk level of each area, and obtains the lesion fuzzy grading result.

2. According to the intelligent analysis and hierarchical diagnosis platform for breast cancer image data based on a large language model in claim 1, the image feature hierarchical mapping result includes an edge smoothness description of a common tissue area, a morphological heterogeneity description of a suspected lesion area, and a density increase feature of a determined lesion area; the pathological diagnosis path information includes a transition feature description from a common tissue to a suspected lesion, a morphological transformation description from a suspected lesion to a lesion, and a density feature path of the lesion; the lesion structure feature transfer data includes an edge roughness parameter, a texture irregularity parameter, and a density contrast analysis of a newly discovered lesion area; the lesion structure update result includes a classification of lesions with clearly defined edges, a classification of lesions with irregular morphology, and a classification of lesions with rough textures; the lesion abnormality probability distribution map includes a division of a density concentration area with a probability value higher than a predetermined threshold, a heterogeneous area with abnormal morphology, and a lesion area with prominent edge roughness; the lesion fuzzy grading result includes a membership evaluation of an unclear edge area, a membership evaluation of a mixed texture area, and a fuzzy membership evaluation of an irregular morphology area.

3. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 1, characterized in that: The organizational level feature segmentation module includes: The tissue region division submodule divides the tissue regions in the image based on breast image data, extracts tissue feature text in the image through a large language model, divides the image into common tissue, suspected lesion area and confirmed lesion area, and obtains the tissue region division result; The regional feature analysis submodule analyzes the edge features, tissue morphology features and density distribution information of each region based on the tissue region division result, extracts the edge smoothness, density distribution and morphological change details of the partition, and generates regional feature analysis data; The feature hierarchical mapping submodule maps the regional features to the hierarchical path of the large language model based on the regional feature analysis data, establishes a corresponding relationship between the features of the organizational level and the hierarchical path, and obtains the image feature hierarchical mapping result.

4. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 1, characterized in that: The image feature path generation module includes: The feature sequence construction submodule establishes a feature description path layer by layer according to the hierarchical order from common tissue to lesion area based on the image feature hierarchical mapping result, describes the order of tissue features at each level, and generates a tissue hierarchical feature list; The feature trend identification submodule identifies the change trend of the features within the hierarchy based on the tissue hierarchy feature list, identifies the transformation of the tissue morphology, density and edge features by analyzing the structural transformation from the normal to the diseased area, and generates feature transformation trend information; Based on the feature conversion trend information, the abnormal feature marking submodule marks the lesion feature set for the area with edge and density abnormalities in the mapping path, describes the distribution and location of the lesion feature set, and obtains pathological diagnosis path information.

5. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 1, characterized in that: The lesion feature similarity transfer module comprises: The abnormal feature extraction submodule extracts the abnormal edge contour, abnormal texture and abnormal density features of the lesion area based on the pathological diagnosis path information, analyzes the numerical distribution and local differences of the density of the lesion area, and generates abnormal features of the lesion; The similar region identification submodule performs feature comparison on the unlabeled region based on the abnormal features of the lesion, detects whether the edge contour morphology of the unlabeled region matches the lesion region, analyzes the density distribution of the unlabeled region and identifies the region with the smallest density difference with the lesion region, and obtains potential similar region information; The potential lesion marking submodule verifies that the density difference range meets the lesion region characteristics based on the potential similar region information, marks the unmarked region that meets the requirements as a potential lesion region, records the structural feature information of the region, and obtains lesion structural feature transfer data.

6. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 1, characterized in that: The lesion structure classification processing module comprises: The lesion feature analysis submodule identifies texture similarity based on the lesion structure feature transfer data according to the local distribution of texture features, and extracts density difference data according to the data distribution of density features to obtain lesion feature similarity data; The lesion classification submodule classifies regions with similar edge features into the same category based on the lesion feature similarity data, and classifies them by comparing the distribution similarity of texture features to generate a lesion classification result; The lesion category updating submodule identifies the edge, texture and density features of the new lesion area based on the lesion category classification result, updates the area matching the features of the marked lesion area to the current category, and obtains a lesion structure update result.

7. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 1, characterized in that: The regional feature probability analysis module includes: The density feature grading submodule, based on the lesion structure update result, grades the density and morphological features of the lesion area according to the density value and morphology to generate graded feature data; The feature probability statistics submodule calculates the frequency distribution of hierarchical features based on the hierarchical feature data, calculates the feature occurrence probability value of each level, integrates the calculation results into probability distribution data, and generates lesion feature probability distribution information; The target probability region marking submodule determines the region where the multi-frequency feature is located based on the lesion feature probability distribution information and the feature occurrence frequency and distribution range, marks the high-probability feature region as a potential risk region, and forms a lesion abnormality probability distribution map.

8. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 7, characterized in that: To calculate the probability of occurrence of features at each level, the formula is used: Get the probability value P(x) of the i-th level feature i ); Among them, f i is the frequency of occurrence of the feature in the i-th level, M is the total number of levels, and f j Represents the frequency of occurrence of features in the jth level.

9. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 1, characterized in that: The lesion fuzzy grading module comprises: The boundary fuzzy recognition submodule detects the boundary fuzzy features of the lesion area based on the lesion abnormality probability distribution map, analyzes the fuzziness degree and transition smoothness of the boundary, and obtains the fuzzy boundary features; The fuzzy membership calculation submodule divides the density uneven area into membership levels according to the fuzzy degree based on the fuzzy boundary features, calculates the membership value of the area, and assigns it to the corresponding fuzzy membership level to obtain the fuzzy membership value of the lesion area; The risk level classification submodule evaluates the risk level of density fluctuation and boundary fuzzy features based on the fuzzy membership value of the lesion area, divides the area into risk levels, and generates fuzzy grading results for the lesions.

10. The breast cancer image data intelligent analysis and hierarchical diagnosis platform based on a large language model according to claim 9, characterized in that: For dividing the density uneven areas into membership levels according to the degree of fuzziness, the formula is used: Get the fuzzy membership value μ(x′); Among them, x′ is the fuzziness parameter of the current area, x′ min is the minimum ambiguity value in the entire analysis area, x′ max is the maximum fuzziness value in the entire analysis area, y is the boundary transition smoothness parameter of the current area, and y min is the minimum boundary smoothness value in the entire analysis area, y max is the maximum value of boundary smoothness in the entire analysis area, and α and β are constant coefficients.

Images