Color feature transformation method of hysteroscopic images based on deep learning

Abstract

The present invention discloses a method for transforming color features of hysteroscopic images based on deep learning, which specifically relates to the field of medical image processing, and solves the problem of insufficient analysis accuracy caused by redundancy of color features in hysteroscopic images. High-quality image data is ensured by multi-level precise screening, and multi-dimensional color features are extracted by nonlinear mapping and structural analysis to enhance the expression and discrimination of features. The constructed multi-level index system combines spectral analysis and graph theory operations to accurately measure the macroscopic distribution differences and microscopic robustness of features. Through Pareto-like frontier search and topological data analysis technology, multiple rounds of iterations and graph theory clustering screening, a key candidate dimension subset with high discrimination and robustness in complex scenes is successfully extracted. The problems of redundancy, insufficient stability and poor adaptability of color features are effectively solved, and the accuracy and reliability of hysteroscopic image analysis are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and more specifically, to a method for transforming color features of hysteroscopic images based on deep learning. Background Art

[0002] As an important endoscopic device for the diagnosis and treatment of gynecological diseases, the high-quality images obtained by hysteroscopy play a key role in helping clinicians accurately identify the lesion area in the uterine cavity, assess the severity of the disease, and formulate effective treatment plans. However, in actual operations, hysteroscopic images are often interfered by multiple factors such as fluctuations in lighting conditions, blood reflections, changes in tissue surface humidity, and dynamic offsets caused by equipment operation. These factors cause the color features of the image to present a complex distribution with high dimensions, nonlinearity, and time-series instability, which increases the difficulty of image analysis and processing and limits the accuracy and real-time nature of clinical decision-making.

[0003] The existing technology faces the following core technical problems in the process of color feature transformation and analysis of hysteroscopic images: first, the color feature dimension is seriously redundant, and the traditional multi-color space feature extraction method is difficult to effectively screen out key features in high-dimensional data, resulting in waste of computing resources and reduced analysis accuracy; second, the feature stability is insufficient, and the image is prone to feature fluctuations under slight environmental changes (such as local illumination changes or slight displacement of tissue structure), affecting subsequent mode switching and decision support; third, the adaptability to multiple scenes is poor, and different lesion states and complex application scenarios require features to have both macroscopic differentiation capabilities and robustness under local disturbance conditions. Traditional methods are difficult to achieve this goal in multi-dimensional and multi-scale feature spaces. Therefore, there is an urgent need for a hysteroscopic image color feature transformation method based on deep learning, which can extract key color feature dimensions with high discrimination and robustness through multi-level image screening, nonlinear mapping and structural analysis, multi-level index evaluation and comprehensive feature screening, thereby significantly improving the accuracy and robustness of hysteroscopic image analysis and meeting the high requirements of clinical applications for real-time and accuracy. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a hysteroscopic image color feature transformation method based on deep learning, which ensures high-quality image data through multi-level precise screening, and extracts multidimensional color features using nonlinear mapping and structural analysis, significantly enhancing the expressiveness and distinguishing power of the features; the constructed multi-level index system combines spectral analysis and graph theory operations to accurately measure the performance of the features in terms of macroscopic distribution differences and microscopic robustness; further, by adopting Pareto-like frontier search and topological data analysis technology, through multiple rounds of iterations and graph theory clustering screening, a subset of key candidate dimensions with high distinguishing power and robustness in complex scenes is successfully extracted; the problems of color feature dimension redundancy, insufficient feature stability and poor adaptability to multiple scenes existing in the existing methods are effectively solved, and the accuracy and reliability of hysteroscopic image analysis based on deep learning in clinical applications are greatly improved to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] S1. By evaluating the clarity, illumination balance and structural consistency of the image, we can filter out samples that do not meet the quality standards and build a high-quality image dataset.

[0007] S2. Apply nonlinear mapping and structural analysis methods to the screened high-quality images to extract multidimensional color features and mark their distribution relationships to form a feature candidate set;

[0008] S3. Construct a multi-level index system for each dimension in the feature candidate set. One type of index measures the distribution differences in macroscopic scenarios through spectral analysis and graph theory operations, and identifies features that can distinguish multiple states in the overall framework; the other type of index examines the fluctuation characteristics in the spatiotemporal neighborhood through perturbation field construction and smoothing operations, and selects dimensions that maintain high-order robustness under slight changes;

[0009] S4. Map the feature dimensions evaluated by the index to a dual-index coordinate system, use Pareto-like frontier search and topological data analysis to perform multiple rounds of iterations and graph clustering on the dimension set, gradually eliminate features that do not meet the stability requirements of complex scenarios, and obtain a subset of key candidate dimensions that can support mode switching decisions.

[0010] In a preferred embodiment, step S1 includes the following contents:

[0011] After hysteroscopic image acquisition, a clarity assessment method based on frequency domain analysis is first applied. The image is converted to the frequency domain through Fourier transform, and the energy distribution of the high-frequency component is calculated to quantify the edge clarity of the image. Images with high-frequency energy lower than the corresponding threshold are judged as blurred images and eliminated. Then, an adaptive illumination correction algorithm is used to perform illumination equalization on the image using local brightness equalization technology. By analyzing the brightness gradient and contrast distribution of the image, images with uneven illumination or severe overexposure and underexposure areas are identified and eliminated. Subsequently, a structural feature detection model based on deep learning is used to identify and verify the lesion area in the image to ensure that only image samples containing clear lesion structures and consistent structural features are retained. Finally, time series consistency detection is implemented to screen out image sequences with significant changes or instability in the time dimension by analyzing the similarity of color and structural features between consecutive frames. After multiple levels of screening steps, a hysteroscopic image dataset, i.e., a high-quality image dataset, is constructed.

[0012] In a preferred embodiment, step S2 includes the following contents:

[0013] First, each high-quality hysteroscopic image is converted to multiple color parameter systems simultaneously;

[0014] Secondly, nonlinear mapping technology is applied to embed high-dimensional image features of each color space after conversion. An autoencoder architecture is used to compress high-dimensional color features into low-dimensional latent space through the encoder, and then the original features are reconstructed through the decoder. The reconstruction error is minimized during the training process to learn the nonlinear distribution pattern of color features.

[0015] Subsequently, the local structural features in the image are extracted by combining the structural analysis method. Specifically, the feature extraction layer in the convolutional neural network is used to capture the subtle structural changes in the image. A comprehensive multi-dimensional feature vector is generated by fusing the structural features with the color features after nonlinear mapping.

[0016] Next, we use topological data analysis technology to mark the distribution relationship of the fused multi-dimensional feature vectors. We use the persistent homology method to analyze the topological structure of features at different scales and identify feature dimensions with significant distribution patterns.

[0017] Finally, based on the comprehensive results of multi-color space conversion, nonlinear mapping, structural feature extraction and topological data analysis, feature dimensions with high topological significance and distinguishing ability are screened out to form a feature candidate set.

[0018] In a preferred embodiment, step S3 includes the following contents:

[0019] After completing the multidimensional color feature extraction and distribution relationship labeling, a multi-level index system is constructed for each dimension in the feature candidate set to systematically evaluate the performance of each feature dimension in terms of macro distribution difference and micro robustness; including the construction and calculation of two types of indexes: macro distribution difference measurement index and micro robustness measurement index.

[0020] In a preferred embodiment, S301. The macroscopic distribution difference measurement index is constructed by combining spectral analysis with graph theory operations, and the specific steps are as follows:

[0021] a. Based on each dimension in the feature candidate set, a corresponding feature space graph is constructed; for each feature dimension, a node set is defined as the value of the corresponding dimension in all samples, and an edge set is determined based on the similarity relationship between feature values;

[0022] b. Apply the spectral clustering method to the constructed feature space graph and extract the previous The smallest non-zero eigenvalues, where is the preset number of clusters;

[0023] c. Definition: Macro-distribution difference refers to the comprehensive measure of the spectral clustering eigenvalues ​​corresponding to the feature dimension.

[0024] In a preferred embodiment, S302. The microscopic robustness metric index is constructed by using a disturbance field construction and smoothing operation method, and the specific steps are as follows:

[0025] a. In the feature candidate set, for the feature dimension, define its disturbance field to simulate the impact of small environmental changes on the feature value; specifically, by introducing local random perturbations to the feature value, the generated perturbed feature value obeys the preset small-amplitude noise distribution;

[0026] b. Apply a smoothing filter to the disturbed eigenvalue to eliminate the influence of noise and obtain a smoothed eigenvalue;

[0027] c. Obtain the fluctuation characteristic quantization value by calculating the degree of change of the characteristic value before and after the disturbance, and then use the average fluctuation characteristic quantization value of all sample points as the micro-robustness measure;

[0028] d. Define the micro robustness measure index as the inverse of the quantified value of the average volatility characteristic.

[0029] In a preferred embodiment, S303. The macro-distribution difference index and the micro-robustness measurement index are normalized, and a comprehensive index is obtained by weighted summation calculation; all feature dimensions in the feature candidate set are sorted from large to small according to the comprehensive index, and the feature dimensions ranked at the top of the rated proportion are selected as candidate dimensions with high discrimination and robustness.

[0030] In a preferred embodiment, step S4 includes the following contents:

[0031] S401. First, the macro-distribution difference index and the micro-robustness measurement index of each candidate dimension are mapped into a dual-index coordinate system to form a two-dimensional feature space.

[0032] In a preferred embodiment, S402. Using a Pareto-like frontier search method, a candidate dimension having both a high macroscopic distribution difference index and a high microscopic robustness metric index value in a dual index space is identified, and the process is as follows:

[0033] a. Sort all candidate dimensions. If a candidate dimension is not inferior to another candidate dimension in two indicators, then this candidate dimension dominates the other candidate dimension. By iteratively eliminating the dominated candidate dimensions, the Pareto-like frontier is gradually approached.

[0034] b. Select candidate dimensions on the Pareto frontier to form a preliminary high-priority feature set;

[0035] S403. Perform topological data analysis on the high-priority feature set obtained through preliminary screening:

[0036] a. Use the persistent homology method to analyze the topological connectivity of candidate dimensions of high-priority feature sets at different scales; the specific steps include:

[0037] Construct the Rips complex: Based on the distance matrix of the candidate dimensions, define the corresponding neighborhood radius and gradually increase it, and record the connectivity changes between the candidate dimensions;

[0038] Calculate persistent stripes: record the birth and death scales of each topological feature to form a persistent stripe set;

[0039] b. Evaluate the persistent strips of each candidate dimension and define the topological significance score of the candidate dimension as the weighted sum of the lengths of the persistent strips;

[0040] c. Based on the candidate dimension topological significance scores, select candidate dimensions whose candidate dimension topological significance scores are greater than the corresponding threshold.

[0041] In a preferred embodiment, S404. clustering the candidate dimensions screened by the Pareto frontier and topological data analysis using graph theory methods to identify the synergistic relationship between features, and gradually eliminating the candidate dimensions that do not meet the stability requirements of complex scenarios through multiple rounds of iterations;

[0042] a. Treat the selected candidate dimensions as nodes in the graph and define edge weights based on their similarity in the dual-index space;

[0043] b. Apply the spectral clustering algorithm to the constructed candidate dimension graph, calculate the Laplacian matrix of the graph, and extract the first rated number of eigenvectors for clustering;

[0044] c. For each cluster, evaluate the consistency and synergy of the candidate dimensions within it; use a topological structure-based connectivity score, which is defined as the sum of the candidate dimension topological significance scores of all candidate dimensions within the cluster;

[0045] d. According to the intra-cluster consistency score, set the removal threshold, remove clusters whose intra-cluster consistency score is lower than the removal threshold, and thus remove candidate dimensions with insufficient synergy in complex scenarios; repeat the spectral clustering and intra-cluster evaluation process until all clusters meet the consistency score requirements;

[0046] S405. After multiple rounds of iterative screening, a set of key candidate dimension subsets are finally determined, which have excellent performance in both macro-distribution differences and micro-robustness, and have stable synergistic relationships in complex scenarios.

[0047] Technical effects and advantages of the hysteroscopic image color feature transformation method based on deep learning of the present invention:

[0048] First, the present invention adopts multi-level precise screening technology to effectively eliminate low-quality images, ensure that subsequent feature extraction is based on high-quality data, and significantly reduce the interference of noise and redundant information. Secondly, the nonlinear mapping and structural analysis methods are used to comprehensively extract and mark multidimensional color features, thereby enhancing the expressiveness and distinguishing power of the features. By constructing a multi-level index system, combining spectral analysis and graph theory operations, the distribution differences of features in macroscopic scenes are accurately measured, and at the same time, the high robustness of features under slight changes is ensured through perturbation field construction and smoothing operations. Further, using Pareto-like frontier search and topological data analysis technology, multiple rounds of iterations and graph theory clustering screening are performed on the feature dimensions, and a key candidate dimension subset with high distinguishing power and robustness in complex scenes is successfully extracted. This key candidate dimension subset not only effectively solves the problems of color feature dimension redundancy, insufficient feature stability, and poor adaptability to multiple scenes in the existing methods, but also greatly improves the accuracy and reliability of hysteroscopic image analysis based on deep learning in actual medical applications. Overall, the present invention achieves efficient and intelligent color feature extraction and screening through systematic method optimization, providing a more accurate and stable image analysis tool for clinical diagnosis, with significant technological progress and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flowchart of the method for transforming color features of hysteroscopic images based on deep learning of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] Embodiment 1: Figure 1 The present invention provides a method for transforming color features of hysteroscopic images based on deep learning, including:

[0052] S1. By evaluating the clarity, illumination balance, and structural consistency of the image, we can filter out samples that do not meet the quality standards and build a high-quality image dataset.

[0053] S2. Apply nonlinear mapping and structural analysis methods to the screened high-quality images to extract multidimensional color features and mark their distribution relationships to form a feature candidate set.

[0054] S3. A multi-level index system is constructed for each dimension in the feature candidate set. One type of index measures the distribution differences in macroscopic scenarios through spectral analysis and graph theory operations, and identifies features that can distinguish multiple states in the overall framework. The other type of index examines the fluctuation characteristics in the spatiotemporal neighborhood through disturbance field construction and smoothing operations, and screens dimensions that maintain high-order robustness under slight changes.

[0055] S4. Map the feature dimensions evaluated by the index to a dual-index coordinate system, use Pareto-like frontier search and topological data analysis to perform multiple rounds of iterations and graph clustering on the dimension set, gradually eliminate features that do not meet the stability requirements of complex scenarios, and obtain a subset of key candidate dimensions that can support mode switching decisions.

[0056] Step S1 includes the following contents:

[0057] After the hysteroscopic image is acquired, the clarity assessment method based on frequency domain analysis is first applied. The image is converted to the frequency domain through Fourier transform, and the energy distribution of the high-frequency component is calculated to quantify the edge clarity of the image. The image with high-frequency energy lower than the corresponding threshold is judged as a blurred image and eliminated. Then, the adaptive illumination correction algorithm is used to perform illumination equalization on the image using the local brightness equalization technology. By analyzing the brightness gradient and contrast distribution of the image, the image with uneven illumination or severe overexposure and underexposure areas is identified and eliminated. Subsequently, the structural feature detection model based on deep learning is used to identify and verify the lesion area in the image to ensure that only image samples containing clear lesion structures and consistent structural features are retained. Finally, the time series consistency detection is implemented to screen out image sequences with significant changes or instability in the time dimension by analyzing the similarity of color and structural features between consecutive frames. After the above multi-level screening steps, a high-quality and structurally consistent hysteroscopic image dataset, namely the high-quality image dataset, is finally constructed, which provides a solid foundation for the subsequent multi-dimensional color feature extraction and analysis.

[0058] The specific processing process is as follows:

[0059] The clarity assessment uses a clarity assessment method based on frequency domain analysis to convert the spatial domain image into the frequency domain through discrete Fourier transform. The specific steps include:

[0060] Apply two-dimensional DFT to each image to get its frequency domain representation .

[0061] Calculate the high frequency energy ratio of the frequency domain image , defined as the high-frequency region (e.g. ) Internal Energy With total energy Ratio of: ;in, is the preset high frequency cutoff threshold.

[0062] If the high frequency energy ratio is lower than the set threshold , the image is judged as a blurred image and is removed.

[0063] The illumination balance detection uses an adaptive illumination correction algorithm to perform illumination balance processing on the image and evaluate the illumination balance. The specific steps include:

[0064] Adaptive histogram equalization is applied to adjust the brightness distribution in local areas and enhance the local contrast of the image.

[0065] Calculating the Light Balance Score , based on the brightness gradient and contrast distribution characteristics of the image. Using the gradient amplitude The lighting balance is quantified by the following formula: ;in, For images At the point The gradient amplitude at is the total number of pixels in the image.

[0066] If the light balance score is lower than the predetermined threshold , the image lighting condition is determined to be unstable and is removed.

[0067] Structural feature detection uses a deep learning-driven structural feature detection model to ensure that the image contains clear and consistent lesion structural features. The specific steps include:

[0068] Use pre-trained convolutional neural networks, such as U-Net or ResNet architecture, to perform semantic segmentation or feature extraction of lesion areas.

[0069] The lesion area probability map outputted from each image is thresholded to extract the binary mask of the lesion structure.

[0070] The consistency score of the lesion structure was calculated based on the area distribution and morphological characteristics of the lesion area to ensure that the lesion area maintained structural coherence between different images.

[0071] If the lesion structure consistency score is lower than the set threshold , the image structure features are determined to be unclear and are removed.

[0072] Time series consistency detection evaluates the temporal consistency of color and structural features of continuously acquired image sequences and screens out images with significant jumps.

[0073] The specific steps include:

[0074] Define similarity metrics between consecutive frames, such as based on structural similarity index , calculate the similarity score between adjacent frames ;in, and They are two consecutive frames of images.

[0075] Setting similarity threshold If the similarity score is lower than the threshold, it is judged that there is a significant color or structure jump and is removed.

[0076] The sliding window technique is applied to comprehensively evaluate the similarity of multiple consecutive frames to ensure the temporal consistency of the entire image sequence.

[0077] After the above multi-level and precise screening steps, samples with insufficient image clarity, uneven illumination, blurred structural features and poor time consistency were eliminated, and finally a high-quality and structurally consistent hysteroscopic image dataset was constructed. This high-quality dataset will provide a stable and reliable input basis for subsequent multi-dimensional color feature extraction and analysis, ensuring that subsequent steps can be performed on consistent and high-quality image data, effectively improving the accuracy and robustness of overall image processing and analysis.

[0078] Step S2 includes the following contents:

[0079] After building a high-quality image dataset, the next key step is to extract rich and representative multidimensional features from these images and mark their distribution relationships to form a feature candidate set. The specific processing process is as follows:

[0080] First, each high-quality hysteroscopic image is simultaneously converted to multiple color parameter systems, such as RGB, HSV, and Lab color spaces, to fully capture the color information of the image. This multi-color space conversion ensures the full expression of different color features and provides a multi-dimensional data foundation for subsequent feature extraction.

[0081] Each high-quality hysteroscopic image is simultaneously converted to multiple color parameter systems, including but not limited to RGB, HSV and Lab color spaces. This step is achieved through a specific color space conversion matrix:

[0082] ;

[0083] ;

[0084] in, and These are functions for converting RGB color space to HSV and Lab color space respectively. This multi-color space conversion ensures comprehensive capture of different color features.

[0085] Secondly, nonlinear mapping technology is applied to perform high-dimensional embedding of the image features of each converted color space. The autoencoder architecture is used to compress the high-dimensional color features into a low-dimensional latent space through the encoder, and then the original features are reconstructed through the decoder. The reconstruction error is minimized during the training process to learn the nonlinear distribution pattern of color features. This process not only effectively reduces the dimensionality, but also retains the key structural information of the color features.

[0086] Apply the autoencoder to the converted image features of each color space for nonlinear mapping to capture the complex distribution pattern of color features. The specific steps include:

[0087] Encoder design: Build an encoder network with multiple hidden layers to transform high-dimensional color features Mapping to a low-dimensional latent space , through the following nonlinear activation function: ;in, is the ReLU activation function, and are the weights and biases of the encoder.

[0088] Decoder design: Construct the corresponding decoder network to reconstruct the features of the latent space back to the original high-dimensional space: ;

[0089] By minimizing the reconstruction error To train the autoencoder, we ensure that the nonlinear mapping can effectively preserve the key structural information of the color features.

[0090] Subsequently, the local structural features in the image are extracted by combining the structural analysis method. Specifically, the feature extraction layer in the convolutional neural network is used to capture the subtle structural changes of the edges, textures and lesion areas in the image. By fusing the structural features with the color features after nonlinear mapping, a comprehensive multi-dimensional feature vector is generated to enhance the expressiveness of the features.

[0091] Specific layers in the convolutional neural network are used to extract local structural features of the image, such as edges, textures, and subtle structural changes in lesion areas. The specific steps include:

[0092] Feature extraction network: Use a pre-trained deep convolutional network (e.g., VGG16 or ResNet50) to extract feature maps through its intermediate convolutional layers ,in and are the height and width of the feature map, respectively. is the number of channels.

[0093] Feature aggregation: Apply global average pooling to the extracted feature maps to aggregate local structural features into a global description vector: ;in, is the structural feature vector; Indicates the pixel index in the height direction, ranging from 1 to ; Indicates the pixel index in the width direction, ranging from 1 to .

[0094] The low-dimensional latent space after nonlinear mapping is fused with the extracted structural feature vector to generate a comprehensive multi-dimensional feature vector ;in, represents the feature concatenation operation, ; This fusion method ensures the coordinated expression of color features and structural features, and enhances the ability to distinguish features.

[0095] Next, we use topological data analysis technology to mark the distribution relationship of the fused multidimensional feature vectors. We use the persistent homology method to analyze the topological structure of features at different scales and identify feature dimensions with significant distribution patterns. This process can reveal the morphology and connection relationship of color features in the feature space, ensuring that the feature candidate set contains key features with high discrimination and structural consistency.

[0096] Apply topological data analysis technology to mark the distribution relationship of the fused multi-dimensional feature vectors, specifically using the persistent homology method to identify the topological structure of features at different scales. The specific steps include:

[0097] Construct Rips complex: Construct Rips complex based on multidimensional feature vectors, define the neighborhood radius and gradually increase it, and record the connectivity changes of features at different scales.

[0098] Compute persistent bands: Calculate the persistent bands of each feature dimension at different scales through persistent homology to quantify the topological significance of the feature: ;in, and Respectively The birth and death scale of a topological feature.

[0099] Characteristic marking: According to the length of the persistent strip , a topological significance score is assigned to each feature dimension, and feature dimensions with longer persistent bands are retained, indicating their important distribution patterns in the feature space.

[0100] Finally, based on the comprehensive results of multi-color space conversion, nonlinear mapping, structural feature extraction and topological data analysis, feature dimensions with high topological significance and distinguishing ability are screened to form a feature candidate set. This feature candidate set not only covers the diversity and complexity of color information, but also enhances the distinguishing ability of features through structural labeling, providing a solid foundation for subsequent multi-level index construction and feature screening.

[0101] Through the above-mentioned specific and rigorous processing logic, it is ensured that in the process of multi-dimensional color feature extraction and distribution relationship labeling, the key color information and its structural characteristics in the hysteroscopic image can be effectively captured and expressed, laying a solid data foundation for accurate analysis and decision-making in subsequent steps.

[0102] Step S3 includes the following contents:

[0103] After completing the multi-dimensional color feature extraction and distribution relationship labeling, the next step is to construct a multi-level index system for each dimension in the feature candidate set to systematically evaluate the performance of each feature dimension in terms of macroscopic distribution difference and microscopic robustness. This process includes the construction and calculation of two types of indexes, including the macroscopic distribution difference measurement index and the microscopic robustness measurement index. The corresponding specific technical logic is as follows:

[0104] S301. Construction of macro-distribution difference measurement index:

[0105] In order to measure the distribution differences of each feature dimension in the overall data framework, a method combining spectral analysis and graph theory operations is used. The specific steps are as follows:

[0106] a. Feature space graph construction: Based on each dimension in the feature candidate set, a corresponding feature space graph is constructed. , define the node set is the value of the corresponding dimension in all samples, and the edge set Determined based on the similarity or proximity between feature values. The similarity measure can use a similarity function based on local density, such as Gaussian kernel similarity.

[0107] b. Application of spectral clustering and graph Laplacian operator: constructing feature space graph Apply the spectral clustering method, by calculating the graph Laplacian matrix ,in is the degree matrix, is the similarity matrix. Calculate the eigenvalue decomposition of the graph Laplacian matrix and extract The smallest non-zero eigenvalue ,in is the preset number of clusters.

[0108] These eigenvalues ​​reflect the feature dimensions Connectivity and distribution diversity in graph structures.

[0109] c. Distribution Difference Index Calculation: Define the Macro Distribution Difference Index is the feature dimension The corresponding comprehensive measure of the spectral clustering eigenvalue is calculated as follows: ; Among them, the smaller eigenvalue This shows that there is an obvious clustering structure in the graph, reflecting that the feature dimension has a high distribution difference ability in the macro scenario.

[0110] S302. Construction of micro robustness measurement index:

[0111] In order to evaluate the robustness of each feature dimension in the local space-time neighborhood, the method of perturbation field construction and smoothing operation is adopted. The specific steps are as follows:

[0112] a. Perturbation field construction: In the feature candidate set, for the feature dimension , defining its disturbance field To simulate the impact of small environmental changes on the eigenvalues. Specifically, by introducing local random perturbations To eigenvalue , generating the perturbed eigenvalues ,in Obey a preset small-amplitude noise distribution, such as uniform distribution or Gaussian distribution.

[0113] b. Smoothing operation application: Apply a smoothing filter (such as bilateral filtering or Gaussian smoothing) to the disturbed eigenvalue to eliminate the influence of noise and obtain the smoothed eigenvalue The purpose of the smoothing operation is to simulate local fluctuations caused by factors such as illumination changes and tissue movement in actual applications.

[0114] c. Fluctuation characteristic quantification: The fluctuation characteristic quantification value is obtained by calculating the degree of change of the characteristic value before and after the disturbance, and then the average fluctuation characteristic quantification value of all sample points is calculated. As a micro-robustness measure.

[0115] d. Quantification of volatility characteristics: Defining micro-robustness measurement index It is the reciprocal of the average fluctuation feature quantization value, reflecting the robustness of the feature dimension under slight changes. ;in, A very small constant to prevent division by zero.

[0116] S303. Normalize the macro-distribution difference index and the micro-robustness measurement index, and perform weighted sum calculation to obtain a comprehensive index.

[0117] The size of the comprehensive index reflects the comprehensive performance of each feature dimension in terms of macroscopic distribution differences and microscopic robustness. Specifically, a larger comprehensive index value indicates that the feature dimension has a higher distribution differentiation ability in macroscopic scenarios and can effectively distinguish multiple pathological states. At the same time, it exhibits strong robustness at the microscopic level and can maintain the consistency and reliability of its features under subtle environmental changes and disturbance conditions. Therefore, the size of the comprehensive index is used as a quantitative standard to evaluate and rank the pros and cons of feature dimensions. A higher comprehensive index value indicates that the feature dimension is more suitable as a key feature for subsequent mode switching decisions, thereby ensuring high-accuracy and stable diagnostic support in complex clinical application scenarios.

[0118] According to the comprehensive index, all feature dimensions in the feature candidate set are sorted from large to small according to the numerical value, and the feature dimensions with the highest rated proportion are selected as candidate dimensions with high discrimination and robustness to enter the next step of comprehensive evaluation and screening process.

[0119] Through the construction and evaluation of the above multi-level index system, the performance of each feature dimension in terms of macroscopic distribution differences and microscopic robustness can be systematically quantified, thereby effectively identifying key color feature dimensions that have high discrimination capabilities in the overall framework and remain robust under local micro-change conditions. This process not only improves the scientificity and accuracy of feature screening, but also lays a solid quantitative foundation for subsequent comprehensive decision-making and mode switching.

[0120] Step S4 includes the following contents:

[0121] After completing the construction of the multi-level index system and the evaluation of the characteristic dimensions, the next step is to conduct a comprehensive analysis of these evaluated candidate dimensions to screen out a subset of key candidate dimensions that can stably support mode switching decisions in complex scenarios. The specific processing process is as follows:

[0122] S401. Quasi-Pareto frontier search: First, the macroscopic distribution difference index and microscopic robustness measurement index of each candidate dimension are mapped into a dual-index coordinate system to form a two-dimensional feature space. This mapping process ensures that the performance of each candidate dimension on two important evaluation indicators is considered, thus providing a comprehensive quantitative basis for subsequent comprehensive screening.

[0123] ;in, Indicates The position of the candidate dimensions in the dual-index coordinate system, and They are the index values ​​of macro-distribution difference and micro-robustness respectively.

[0124] S402. Quasi-Pareto frontier search: Use the quasi-Pareto frontier search method to identify candidate dimensions that have both high macro-distribution difference index and high micro-robustness metric index values ​​in the dual-index space. This process defines a non-dominated sorting mechanism to ensure that the selected candidate dimensions perform well in both indicators, avoiding the characteristics that perform well in only one indicator from entering the next screening stage.

[0125] a. Non-dominated sorting: Sort all candidate dimensions. If a candidate dimension is not inferior to another candidate dimension in two indicators, then this candidate dimension dominates the other candidate dimension. By iteratively eliminating the dominated candidate dimensions, the Pareto-like frontier is gradually approached.

[0126] b. Frontier screening: Select candidate dimensions on the Pareto-like frontier to form a preliminary high-priority feature set. The candidate dimensions in this set show significant comprehensive advantages in the dual-index space.

[0127] S403. Application of topological data analysis: Perform topological data analysis on the high-priority feature set obtained through preliminary screening to further identify and retain candidate dimensions with robust topological structures in complex scenarios.

[0128] a. Persistent homology calculation: Use the persistent homology method to analyze the topological connectivity of candidate dimensions of high-priority feature sets at different scales. The specific steps include:

[0129] Construct the Rips complex: Based on the distance matrix of the candidate dimensions, define the corresponding neighborhood radius and gradually increase it, and record the connectivity changes between the candidate dimensions.

[0130] Calculate persistent strips: Record the birth and death scales of each topological feature (such as connected branches and rings) to form a set of persistent strips.

[0131] b. Topological significance evaluation: The persistent strip of each candidate dimension is evaluated, and the candidate dimension topological significance score is defined as the weighted sum of the persistent strip lengths.

[0132] c. Topological screening: Based on the topological significance scores of candidate dimensions, candidate dimensions with topological significance scores greater than the corresponding threshold are screened to ensure that these candidate dimensions maintain a robust topological structure at different scales.

[0133] S404. Graph clustering and iterative elimination: Graph theory methods are used to cluster candidate dimensions that have been screened by Pareto-like frontier and topological data analysis to identify synergistic relationships between features, and candidate dimensions that do not meet the stability requirements of complex scenarios are further eliminated through multiple rounds of iterations.

[0134] a. Candidate dimension graph construction: The screened candidate dimensions are regarded as nodes in the graph, and their edge weights are defined based on their similarity in the dual-index space (such as Euclidean distance or Gaussian kernel similarity).

[0135] b. Spectral clustering application: Apply the spectral clustering algorithm to the constructed candidate dimension graph, calculate the Laplacian matrix of the graph, and extract a predetermined number of eigenvectors for clustering.

[0136] c. Intra-cluster consistency evaluation: For each cluster, the consistency and synergy of the candidate dimensions within it are evaluated. A topologically based connectivity score is used, which is defined as the sum of the candidate dimension topological significance scores of all candidate dimensions within the cluster.

[0137] d. According to the intra-cluster consistency score, set the removal threshold, and remove clusters whose intra-cluster consistency score is lower than the removal threshold, so as to remove candidate dimensions with insufficient synergy in complex scenarios. Repeat the spectral clustering and intra-cluster evaluation process until all clusters meet the consistency score requirements.

[0138] S405. Determination of key candidate dimension subsets: After multiple rounds of iterative screening using Pareto-like frontier search, topological data analysis, and graph clustering, a set of key candidate dimension subsets with excellent performance in both macroscopic distribution differences and microscopic robustness and stable synergy in complex scenarios are finally determined. This subset can effectively support subsequent mode switching decisions and ensure high accuracy and robustness in a changing clinical application environment.

[0139] The key candidate dimension subset is of vital importance in the present invention. By systematically eliminating the redundancy of color feature dimensions, the subset significantly improves the distinguishing power and robustness of the features, ensuring that the color features of hysteroscopic images can be accurately and reliably identified and analyzed in variable clinical application scenarios. The construction process of this subset is based on multi-level index evaluation and graph theory clustering methods, which comprehensively considers the performance of features in terms of macroscopic distribution differences and microscopic robustness, ensuring that the selected features have a high degree of distinguishing ability in the overall data framework and maintain stability under local micro-change conditions. This not only effectively solves the core problems of poor adaptability to multiple scenarios and insufficient feature stability in the prior art, but also significantly improves the accuracy and robustness of the hysteroscopic image color feature transformation method based on deep learning in mode switching decisions. Therefore, the construction of the key candidate dimension subset provides solid technical support for the accuracy and reliability of the present invention in clinical applications, and promotes the development of hysteroscopic image analysis technology to a higher level.

[0140] The present invention maps the color feature dimensions evaluated by multi-level indexes to a dual-index coordinate system, and adopts a Pareto-like frontier search and topological data analysis technology to perform multiple rounds of iterations and graph theory clustering processing on the feature dimension set. In the dual-index space, the Pareto-like frontier search effectively identifies the feature dimensions that are excellent in both macroscopic distribution differences and microscopic robustness, ensuring that the selected features have a high degree of distinguishing ability and stability in the overall data framework. Subsequently, the topological structure between the feature dimensions is further analyzed by topological data analysis, and the feature dimensions that do not meet the stability requirements in complex application scenarios are identified and eliminated by the graph theory clustering method. This process is optimized through multiple rounds of iterations to gradually screen out a subset of key candidate dimensions that can maintain consistency and robustness in a variable clinical environment. The key candidate dimension subsets finally obtained not only eliminate the redundancy of color feature dimensions, improve the distinguishing power and robustness of features, but also have the ability to support mode switching decisions in complex scenarios, effectively solving the core problems of poor adaptability to multiple scenarios and insufficient feature stability in the prior art, thereby significantly improving the accuracy and reliability of the hysteroscopic image color feature transformation method based on deep learning in clinical applications.

[0141] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0142] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0143] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0144] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for transforming color features of hysteroscopic images based on deep learning, characterized in that: Includes steps: S1. By evaluating the clarity, illumination balance and structural consistency of the image, we can filter out samples that do not meet the quality standards and build a high-quality image dataset. S2. Apply nonlinear mapping and structural analysis methods to the screened high-quality images to extract multidimensional color features and mark their distribution relationships to form a feature candidate set; S3. Construct a multi-level index system for each dimension in the feature candidate set. One type of index measures the distribution differences in macroscopic scenarios through spectral analysis and graph theory operations, and identifies features that can distinguish multiple states in the overall framework; the other type of index examines the fluctuation characteristics in the spatiotemporal neighborhood through perturbation field construction and smoothing operations, and selects dimensions that maintain high-order robustness under slight changes; S4. Map the feature dimensions evaluated by the index to the dual-index coordinate system, use the Pareto-like frontier search and topological data analysis to perform multiple rounds of iterations and graph clustering on the dimension set, gradually eliminate the features that do not meet the stability requirements of complex scenarios, and obtain the key candidate dimension subsets that can support mode switching decisions; Step S2 includes the following contents: First, each high-quality hysteroscopic image is converted to multiple color parameter systems simultaneously; Secondly, nonlinear mapping technology is applied to embed high-dimensional image features of each color space after conversion. An autoencoder architecture is used to compress high-dimensional color features into low-dimensional latent space through the encoder, and then the original features are reconstructed through the decoder. The reconstruction error is minimized during the training process to learn the nonlinear distribution pattern of color features. Subsequently, the local structural features in the image are extracted by combining the structural analysis method. Specifically, the feature extraction layer in the convolutional neural network is used to capture the subtle structural changes in the image. A comprehensive multi-dimensional feature vector is generated by fusing the structural features with the color features after nonlinear mapping. Next, we use topological data analysis technology to mark the distribution relationship of the fused multi-dimensional feature vectors. We use the persistent homology method to analyze the topological structure of features at different scales and identify feature dimensions with significant distribution patterns. Finally, based on the comprehensive results of multi-color space conversion, nonlinear mapping, structural feature extraction and topological data analysis, feature dimensions with high topological significance and distinguishing ability are screened out to form a feature candidate set.

2. The method for transforming color features of hysteroscopic images based on deep learning according to claim 1, characterized in that: Step S1 includes the following contents: After hysteroscopic image acquisition, a clarity assessment method based on frequency domain analysis is first applied. The image is converted to the frequency domain through Fourier transform, and the energy distribution of the high-frequency component is calculated to quantify the edge clarity of the image. Images with high-frequency energy lower than the corresponding threshold are judged as blurred images and eliminated. Then, an adaptive illumination correction algorithm is used to perform illumination equalization on the image using local brightness equalization technology. By analyzing the brightness gradient and contrast distribution of the image, images with uneven illumination or severe overexposure and underexposure areas are identified and eliminated. Subsequently, a structural feature detection model based on deep learning is used to identify and verify the lesion area in the image to ensure that only image samples containing clear lesion structures and consistent structural features are retained. Finally, time series consistency detection is implemented to screen out image sequences with significant changes or instability in the time dimension by analyzing the similarity of color and structural features between consecutive frames. After multiple levels of screening steps, a hysteroscopic image dataset, i.e., a high-quality image dataset, is constructed.

3. The method for transforming color features of hysteroscopic images based on deep learning according to claim 2, characterized in that: Step S3 includes the following contents: After completing the multidimensional color feature extraction and distribution relationship labeling, a multi-level index system is constructed for each dimension in the feature candidate set to systematically evaluate the performance of each feature dimension in terms of macroscopic distribution differences and microscopic robustness; It includes the construction and calculation of two types of indexes: macro-distribution difference measurement index and micro-robustness measurement index.

4. The method for transforming color features of hysteroscopic images based on deep learning according to claim 3, characterized in that: S301. The macroscopic distribution difference measurement index is constructed by combining spectral analysis with graph theory operations. The specific steps are as follows: a. Based on each dimension in the feature candidate set, a corresponding feature space graph is constructed; for each feature dimension, a node set is defined as the value of the corresponding dimension in all samples, and an edge set is determined based on the similarity relationship between feature values; b. Apply the spectral clustering method to the constructed feature space graph and extract the previous The smallest non-zero eigenvalues, where is the preset number of clusters; c. Definition: Macro-distribution difference refers to the comprehensive measure of the spectral clustering eigenvalues ​​corresponding to the feature dimension.

5. The method for transforming color features of hysteroscopic images based on deep learning according to claim 4, characterized in that: S302. The construction of micro-robustness measurement index adopts the method of perturbation field construction and smoothing operation. The specific steps are as follows: a. In the feature candidate set, for the feature dimension, define its disturbance field to simulate the impact of small environmental changes on the feature value; specifically, by introducing local random perturbations to the feature value, the generated perturbed feature value obeys the preset small-amplitude noise distribution; b. Apply a smoothing filter to the disturbed eigenvalue to eliminate the influence of noise and obtain a smoothed eigenvalue; c. Obtain the fluctuation characteristic quantization value by calculating the degree of change of the characteristic value before and after the disturbance, and then use the average fluctuation characteristic quantization value of all sample points as the micro-robustness measure; d. Define the micro robustness measure index as the inverse of the quantified value of the average volatility characteristic.

6. The method for transforming color features of hysteroscopic images based on deep learning according to claim 5, characterized in that: S303. Normalize the macro-distribution difference index and the micro-robustness measurement index, and then perform weighted sum calculation to obtain a comprehensive index; sort all feature dimensions in the feature candidate set from large to small according to the comprehensive index, and select the feature dimensions with the highest rated proportion as candidate dimensions with high discrimination and robustness.

7. The method for transforming color features of hysteroscopic images based on deep learning according to claim 6, characterized in that: Step S4 includes the following contents: S401. First, the macro-distribution difference index and the micro-robustness measurement index of each candidate dimension are mapped into a dual-index coordinate system to form a two-dimensional feature space.

8. The method for transforming color features of hysteroscopic images based on deep learning according to claim 7, characterized in that: S402. Use the Pareto-like frontier search method to identify candidate dimensions that have both high macroscopic distribution difference index and high microscopic robustness metric index values ​​in the dual index space. The process is as follows: a. Sort all candidate dimensions. If a candidate dimension is not inferior to another candidate dimension in two indicators, then this candidate dimension dominates the other candidate dimension. By iteratively eliminating the dominated candidate dimensions, the Pareto-like frontier is gradually approached. b. Select candidate dimensions on the Pareto frontier to form a preliminary high-priority feature set; S403. Perform topological data analysis on the high-priority feature set obtained through preliminary screening: a. Use the persistent homology method to analyze the topological connectivity of candidate dimensions of high-priority feature sets at different scales; the specific steps include: Construct the Rips complex: Based on the distance matrix of the candidate dimensions, define the corresponding neighborhood radius and gradually increase it, and record the connectivity changes between the candidate dimensions; Calculate persistent stripes: record the birth and death scales of each topological feature to form a persistent stripe set; b. Evaluate the persistent strips of each candidate dimension and define the topological significance score of the candidate dimension as the weighted sum of the lengths of the persistent strips; c. Based on the candidate dimension topological significance scores, select candidate dimensions whose candidate dimension topological significance scores are greater than the corresponding threshold.

9. The method for transforming color features of hysteroscopic images based on deep learning according to claim 8, characterized in that: S404. Use graph theory methods to cluster the candidate dimensions screened by quasi-Pareto frontier and topological data analysis to identify the synergistic relationship between features, and gradually eliminate candidate dimensions that do not meet the stability requirements of complex scenarios through multiple rounds of iterations; a. Treat the selected candidate dimensions as nodes in the graph and define edge weights based on their similarity in the dual-index space; b. Apply the spectral clustering algorithm to the constructed candidate dimension graph, calculate the Laplacian matrix of the graph, and extract the first rated number of eigenvectors for clustering; c. For each cluster, evaluate the consistency and synergy of the candidate dimensions within it; use a topological structure-based connectivity score, which is defined as the sum of the candidate dimension topological significance scores of all candidate dimensions within the cluster; d. According to the intra-cluster consistency score, set the removal threshold, remove clusters whose intra-cluster consistency score is lower than the removal threshold, and thus remove candidate dimensions with insufficient synergy in complex scenarios; repeat the spectral clustering and intra-cluster evaluation process until all clusters meet the consistency score requirements; S405. After multiple rounds of iterative screening, a set of key candidate dimension subsets are finally determined, which have excellent performance in both macro-distribution differences and micro-robustness, and have stable synergistic relationships in complex scenarios.

Images