Gynecological endoscopic image intelligent analysis and cervical lesion precise diagnosis system

Abstract

The invention relates to the technical field of medical image processing, in particular to a gynecological endoscope image intelligent analysis and cervical lesion precise diagnosis system. Comprising an image enhancement module, a differential geometric feature extraction module, a topology preserving manifold learning module, a curvature-blood vessel correlation analysis module, a biopsy point accurate positioning module, a real-time intelligent auxiliary diagnosis module, a clinical feedback module and a self-adaptive optimization module. Micro morphological change features are extracted through multi-scale curvature analysis; a topology sensitive encoder and a manifold alignment decoder are adopted, and a key topology structure is kept; establishing a correlation model of the tissue surface curvature change and the blood vessel morphology, and analyzing differential morphological characteristics; determining an optimal biopsy position based on the comprehensive characteristics; according to the method, the diagnosis result and basis are generated, visual display is provided, the detection rate of cervical lesions, especially early minimal lesions, is remarkably increased, the missed diagnosis rate is reduced, and multi-point biopsy is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, in particular to a gynecological endoscope image intelligent analysis and cervical lesion precise diagnosis system for intelligent analysis of cervical endoscope images and precise diagnosis of cervical lesions. BACKGROUND

[0002] Cervical cancer is one of the most common malignant tumors in women, and early diagnosis and treatment are crucial for improving patient survival rate. Currently, the diagnosis of cervical lesions mainly relies on endoscopy and biopsy pathological analysis. However, traditional endoscopy has the following problems: on the one hand, endoscope images are often affected by problems such as reflection, shadow and color distortion, reducing image quality; on the other hand, the identification of lesion areas and the selection of biopsy points are highly dependent on the experience of doctors, with strong subjectivity and poor consistency. In addition, traditional image analysis methods mainly rely on color and texture features, which are difficult to capture the subtle morphological changes in the early stage of cervical lesions, resulting in a high rate of missed diagnosis of early lesions.

[0003] In the prior art, although there are some deep learning-based image analysis systems applied to cervical lesion diagnosis, these systems usually use general convolutional neural network architecture, lacking special optimization for cervical tissue characteristics. In particular, these systems often lose key topological structure information in the feature extraction process, resulting in insufficient recognition accuracy for irregularly shaped lesions. At the same time, existing systems lack comprehensive analysis of the differential characteristics of tissue surfaces and the correlation of blood vessel morphology, and cannot fully capture the biological characteristics of cervical lesions, affecting the accuracy of diagnosis.

[0004] Therefore, there is an urgent need for a gynecological endoscope image intelligent analysis and cervical lesion precise diagnosis system that can overcome the above problems. SUMMARY

[0005] The purpose of the present application is to provide a gynecological endoscope image intelligent analysis and cervical lesion precise diagnosis system that can accurately analyze cervical endoscope images and accurately diagnose lesions through differential geometric feature extraction, topologically preserving manifold learning and curvature-vascular correlation analysis.

[0006] The present application provides a gynecological endoscope image intelligent analysis and cervical lesion precise diagnosis system, which comprises:

[0007] An image enhancement module for processing the acquired cervical endoscope images, improving image contrast, eliminating reflections and shadows, and performing color correction to obtain enhanced cervical images;

[0008] The differential geometric feature extraction module, connected to the image enhancement module, is used to receive the enhanced cervical image, model the surface of the enhanced cervical image as a Riemannian manifold, calculate multi-scale curvature features, extract features of abnormal regions, and output the curvature feature map and preliminary segmentation results of the cervical image.

[0009] The topology-preserving manifold learning module, connected to the differential geometry feature extraction module, receives the curvature feature map and preliminary segmentation results, processes the feature map through a topology-sensitive encoder and a manifold-aligned decoder, preserves key topological structures, and outputs refined segmentation results.

[0010] The curvature-vascular correlation analysis module, connected to the topology-preserving manifold learning module, is used to receive the fine segmentation results, establish a correlation model between tissue surface curvature changes and vascular morphology, analyze differential morphological features, and construct a comprehensive feature descriptor.

[0011] The biopsy point precise positioning module is connected to the curvature-vascular correlation analysis module and is used to receive the comprehensive feature descriptor, analyze the abnormal area features, and determine the optimal biopsy location.

[0012] The real-time intelligent auxiliary diagnosis module is connected to the biopsy point precise positioning module. It is used to receive the optimal biopsy location and the comprehensive feature descriptor, generate cervical lesion diagnosis results and diagnostic basis, and provide auxiliary diagnostic information.

[0013] Preferably, the differential geometric feature extraction module includes:

[0014] The manifold modeling unit is used to map the enhanced cervical image surface into a parameterized surface, establish a local differential coordinate system, and calculate the local metric tensor.

[0015] The curvature calculation unit, connected to the manifold modeling unit, is used to calculate the Gaussian curvature and average curvature of the parameterized surface at multiple scales and generate a curvature feature map.

[0016] An anomaly region preliminary location unit, connected to the curvature calculation unit, is used to determine the preliminary location of the anomaly region based on the curvature feature map through adaptive threshold segmentation and curvature gradient field analysis.

[0017] Preferably, the curvature calculation unit is implemented in the following manner:

[0018] For the parametric surface, a scale sequence is set. ,in ;

[0019] For each scale The parameterized surface is processed using a Gaussian filter with the corresponding kernel width;

[0020] a curvature feature set corresponding to the scale sequence is obtained by calculating the curvature map at each scale;

[0021] The curvature features at multiple scales are fused through an adaptive weight function to obtain a fused curvature feature map.

[0022] Preferably, the topologically preserved manifold learning module comprises:

[0023] a topologically sensitive encoder configured to perform multi-layer feature extraction on the curvature feature map and the enhanced cervical image, and to calculate persistent homology features after each layer of extraction to supervise the topological structure of the feature space;

[0024] a manifold alignment decoder connected to the topologically sensitive encoder and configured to construct a feature manifold and ensure geometric structure consistency in the decoding process through a manifold alignment mechanism;

[0025] an adaptive skip connection unit connected between the topologically sensitive encoder and the manifold alignment decoder and configured to dynamically adjust the weight of the skip connection based on feature distribution similarity and topological feature similarity.

[0026] Preferably, the topologically sensitive encoder calculates the persistent homology features in the following manner:

[0027] a filter complex is constructed on each layer of convolutional feature map to capture the topological structure of the feature space;

[0028] 0-dimensional and 1-dimensional Betti numbers are calculated to represent the number of connected components and loops;

[0029] a topological feature vector is constructed and compared with the topological feature vector of the original image to calculate a topological similarity index;

[0030] a topological loss term is constructed based on the topological similarity index to guide the network to preserve key topological structures.

[0031] Preferably, the curvature-blood vessel correlation analysis module comprises:

[0032] a blood vessel-curvature collaborative detection unit configured to extract a blood vessel feature map and a curvature feature map and to enhance the two types of features through a collaborative attention mechanism;

[0033] a differential morphological analysis unit connected to the blood vessel-curvature collaborative detection unit and configured to calculate Minkowski functionals and fractal dimensions through multi-scale morphological processing to quantify the morphological complexity of abnormal regions;

[0034] A feature tensor construction unit, connected with the differential morphological analysis unit, is configured to integrate all extracted features, construct a feature correlation graph, and analyze the interaction strength between different features.

[0035] As preferred, the vessel-curvature synergy detection unit is implemented in the following manner:

[0036] The enhanced cervical image is processed using a dedicated vessel enhancement filter to extract a vessel feature map;

[0037] The curvature feature map is input as a curvature channel;

[0038] A conversion function is designed to respectively map the curvature feature to the vessel feature space and map the vessel feature to the curvature feature space;

[0039] Attention weights are calculated to respectively generate a curvature-enhanced vessel feature and a vessel-enhanced curvature feature;

[0040] The enhanced features are fused through an adaptive weight factor to generate a synergy feature map.

[0041] As preferred, the biopsy point accurate positioning module comprises:

[0042] A feature comprehensive analysis unit is configured to analyze the comprehensive feature descriptor, evaluate the severity of the abnormal area, and evaluate the feature significance;

[0043] A biomarker correlation unit, connected with the feature comprehensive analysis unit, is configured to correlate the features of the abnormal area with known cervical lesion biomarkers;

[0044] A biopsy point optimization unit, connected with the biomarker correlation unit, is configured to determine the optimal biopsy location and the number of biopsies based on the features of the abnormal area and the biomarker correlation results, and in combination with clinical biopsy rules.

[0045] As preferred, the real-time intelligent auxiliary diagnosis module comprises:

[0046] A lesion classification unit is configured to classify the cervical state as normal, cervical intraepithelial neoplasia, and cervical cancer based on the comprehensive feature descriptor;

[0047] A diagnosis basis generation unit, connected with the lesion classification unit, is configured to extract key features and evidence supporting the diagnosis result, and generate diagnosis basis information;

[0048] A visual display unit, connected with the diagnosis basis generation unit, is configured to visually display the cervical lesion diagnosis result, diagnosis basis, and optimal biopsy location on the enhanced cervical image.

[0049] As preferred, the system further comprises:

[0050] A clinical feedback module connected with the real-time intelligent auxiliary diagnosis module, configured to receive feedback information of the diagnosis result from a doctor;

[0051] An adaptive optimization module connected with the clinical feedback module, configured to adjust parameters of the differential geometry feature extraction module, the topologically preserving manifold learning module, the curvature-vascular correlation analysis module, the biopsy point accurate positioning module and the real-time intelligent auxiliary diagnosis module according to the feedback information, and improve the system diagnosis accuracy.

[0052] The beneficial effects of the present application include:

[0053] 1. The accuracy of cervical lesion detection is improved, especially the detection rate of early micro-lesions is significantly improved, which is helpful for early diagnosis and treatment of cervical cancer.

[0054] 2. Through differential geometry modeling, the small morphological changes of the cervical tissue surface can be captured, and more rich feature information can be provided for lesion recognition.

[0055] 3. The topologically preserving manifold learning mechanism ensures that the key structural information is not lost in the deep learning process, and improves the recognition accuracy of irregular shape lesions.

[0056] 4. Through curvature-vascular correlation analysis, the internal relationship between tissue morphological changes and vascular abnormalities is revealed, and the expression ability of lesion features is enhanced from the biological mechanism.

[0057] 5. The accurate positioning of biopsy points is realized, unnecessary multiple biopsy is reduced, and the biopsy efficiency and the representativeness of tissue samples are improved.

[0058] 6. Intuitive visual display and diagnosis basis explanation are provided, the explainability of the system is enhanced, and the doctor is assisted to make more accurate diagnosis decision. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 It is the overall architecture schematic diagram of the system of the present application;

[0060] Figure 2 It is the structure schematic diagram of the differential geometry feature extraction module in the present application;

[0061] Figure 3 It is the structure schematic diagram of the topologically preserving manifold learning module in the present application;

[0062] Figure 4 It is the structure schematic diagram of the curvature-vascular correlation analysis module in the present application;

[0063] Figure 5This is a schematic diagram of the biopsy point precise positioning module in this invention;

[0064] Figure 6 This is a schematic diagram of the real-time intelligent auxiliary diagnosis module in this invention;

[0065] Figure 7 This is a schematic diagram of the processing flow of the system of the present invention. Detailed Implementation

[0066] Please refer to Figure 1 - Figure 7 The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0067] like Figure 1 As shown, the intelligent analysis system for gynecological endoscopic images and precise diagnosis of cervical lesions provided by the present invention includes: an image enhancement module 1, a differential geometric feature extraction module 2, a topology-preserving manifold learning module 3, a curvature-vascular correlation analysis module 4, a biopsy point precise positioning module 5, a real-time intelligent auxiliary diagnosis module 6, and a clinical feedback module 7.

[0068] Image enhancement module 1 processes the acquired cervical endoscopic images to improve image contrast, eliminate reflections and shadows, and perform color correction to obtain an enhanced cervical image. In one embodiment of the invention, image enhancement module 1 employs a multi-step processing strategy. First, it improves image contrast through adaptive histogram equalization. Then, it eliminates reflections and shadows using a combination of Gaussian filtering and median filtering. Finally, it performs color correction based on a color correction algorithm to ensure the tissue color matches the actual color. Preferably, the contrast enhancement step uses a local region adaptive processing method, setting different enhancement parameter coefficients for different regions. For example, for regions with low brightness, the enhancement coefficient can be set to 1.2-1.5, while for regions with high brightness, the enhancement coefficient can be set to 0.8-1.0, thereby achieving a more balanced image effect.

[0069] like Figure 2 As shown, the differential geometric feature extraction module 2 is connected to the image enhancement module 1. It is used to receive the enhanced cervical image, model the surface of the enhanced cervical image as a Riemannian manifold, calculate multi-scale curvature features, extract features of abnormal regions, and output the curvature feature map and preliminary segmentation results of the cervical image.

[0070] The differential geometric feature extraction module 2 includes a manifold modeling unit 21, a curvature calculation unit 22, and an anomaly region preliminary localization unit 23.

[0071] The manifold modeling unit 21 is used to map the enhanced cervical image surface to a parametric surface, establish a local differential coordinate system, and calculate the local metric tensor. Specifically, a mapping function from a two-dimensional image to a three-dimensional manifold is first established. ,for each pixel point ,calculate its corresponding point on the manifold In a preferred embodiment of the present application, the mapping function can use the image intensity value as the third dimensional coordinate to construct the parameterized surface . Then, a local tangent plane is established at each point on the manifold , the local coordinate basis vectors and are calculated, and the local metric tensor is constructed. To eliminate noise and small artifacts while preserving the true tissue morphology, an adaptive Gaussian filter can be applied to smooth the surface, with the filter kernel width automatically adjusted according to the local curvature. For example, in areas with high curvature, can be set to a small value (e.g. 1.0-1.5) to preserve details; while in areas with low curvature, can be set to a large value (e.g. 2.0-3.0) to more effectively suppress noise.

[0072] The curvature calculation unit 22 is connected to the manifold modeling unit 21 and is used to calculate the Gaussian curvature and mean curvature of the parameterized surface at multiple scales to generate a curvature feature map. The curvature calculation unit 22 is implemented in the following manner:

[0073] 1) For the parameterized surface, a scale sequence is set, where ;

[0074] 2) For each scale , a Gaussian filter with a corresponding kernel width is used to process the parameterized surface;

[0075] 3) The curvature map at each scale is calculated to obtain a curvature feature set corresponding to the scale sequence;

[0076] 4) The curvature features at multiple scales are fused by an adaptive weight function to obtain a fused curvature feature map.

[0077] In an embodiment of the present application, the scale sequence can be set as , corresponding to different sizes of tissue features. For each scale , a Gaussian filter with a kernel width of is first used to process the surface , then the principal curvatures and are calculated, and then the Gaussian curvature and the mean curvature are obtained, generating the Gaussian curvature map and the mean curvature map .

[0078] To integrate curvature features at different scales, an adaptive weighting function is designed. Weights are assigned based on the discriminative power of curvature features at different scales. In practical applications, the weight function can be optimized using a validation set, initially set to... The formula for calculating the fused curvature feature map is:

[0079] ,

[0080] in, To fuse curvature feature maps, Let be the weight of the i-th scale. This is the Gaussian curvature plot at the i-th scale. The total number of scales, This represents the summation operation from i=1 to i=n, ​​used to accumulate the weighted curvature feature maps at each scale. Finally, the fused features are normalized to obtain the standardized curvature feature map. .

[0081] The preliminary anomaly region localization unit 23 is connected to the curvature calculation unit 22, and is used to determine the preliminary location of the anomaly region based on the curvature feature map through adaptive threshold segmentation and curvature gradient field analysis. In the specific implementation process, firstly based on... Based on the histogram characteristics, design a dynamic threshold function. :

[0082] ,

[0083] in, Let be the dynamic threshold at point (x, y). This represents the mean curvature of a local region (typically a 15×15 pixel window) centered at the point (x,y). The standard deviation of curvature for this local region. This is an adjustment factor that can be set according to the sensitivity requirements for lesion detection, typically ranging from 2.0 to 3.0. For each point (x, y), if If it is, then it is marked as a potential outlier.

[0084] Subsequently, the gradient field of the curvature feature map is calculated. Identify gradient magnitude Greater than the threshold The region is a potential boundary. In a preferred embodiment of the invention, Can be set to The range of 15% to 20% of the maximum value has been validated by a large number of clinical images and can effectively balance the sensitivity and specificity of boundary detection.

[0085] Finally, the level set method is used to refine the preliminary segmentation results. Using the preliminary segmentation results as the initial contour, the velocity field F is defined as follows:

[0086] ,

[0087] in, For the velocity field of the level set evolution, Let be the stopping function based on image gradient, defined as For image, For image gradient, Describes the Euclidean norm. This is the balloon force term, controlling the expansion or contraction of the balloon's shape; it is typically set to 0.5-1.5. The curvature term, used to smooth the contour, is calculated as follows: ,in For level set functions, This represents the divergence operation.

[0088] Iteratively update the level set function ,

[0089] Stop iterating when the value is 0.001 or the maximum number of iterations is reached (usually set to 100).

[0090] like Figure 3 As shown, the topology-preserving manifold learning module 3 is connected to the differential geometry feature extraction module 2. It is used to receive curvature feature maps and preliminary segmentation results, process the feature maps through a topology-sensitive encoder and a manifold alignment decoder, preserve key topological structures, and output fine segmentation results.

[0091] The topology-preserving manifold learning module 3 includes a topology-sensitive encoder 31, a manifold alignment decoder 32, and an adaptive jump connection unit 33.

[0092] A topology-sensitive encoder 31 is used to perform multi-layer feature extraction on the curvature feature map and the enhanced cervical image, and calculates persistent homology features after each layer extraction to supervise the topological structure of the feature space. In one embodiment of the present invention, a three-channel input tensor is first constructed. ,in To enhance the grayscale version of the image, This is the preliminary segmentation result. This is to standardize the curvature feature map. Multi-channel information is fused using a 1×1 convolution to generate the initial feature map. .

[0093] The encoder adopts a 5-layer structure, each layer containing two 3x3 convolutional layers and one 2x2 max-pooling layer. After each convolution, the persistent homology features are calculated by constructing a filter complex on the feature map to capture the topological structure of the feature space, and then calculating the 0-dimensional and 1-dimensional Betti numbers and , representing the number of connected components and loops, respectively, and storing the topological feature vector of each layer .

[0094] The topologically sensitive encoder 31 calculates the persistent homology features by the following ways:

[0095] Constructing a filter complex on each layer of convolutional feature map to capture the topological structure of the feature space;

[0096] Calculating the 0-dimensional and 1-dimensional Betti numbers representing the number of connected components and loops;

[0097] Constructing the topological feature vector and comparing it with the topological feature vector of the original image to calculate the topological similarity index;

[0098] Based on the topological similarity index, the topological loss term is constructed to guide the network to maintain the key topological structure.

[0099] In actual implementation, the construction of filter complex adopts a threshold-based method, applying a series of thresholds to the feature map F to get a sequence of nested binary maps. For each binary map, the number of connected components and the number of loops are calculated to form the persistent homology barcode. The calculation of the topological similarity index is based on the Wasserstein distance of the barcodes:

[0100] ,

[0101] where is the topological similarity index, with a value range of [0, 1], and the larger the value, the more similar the topological structure, is the natural exponential function, is the scaling factor, controlling the sensitivity of similarity change with distance, usually set to 0.1-0.5, is the Wasserstein distance between the original image barcode and the encoder feature barcode , measuring the difference between the two barcodes. The topological loss term is defined as:

[0102] ,

[0103] where, The range of the topological loss term is [0, 1], and the smaller the value is, the better the topological structure is maintained. The loss term encourages the network to maintain the key topological structure in the feature extraction process.

[0104] The manifold alignment decoder 32 is connected with the topologically sensitive encoder 31, and is used to construct a feature manifold and ensure the geometric structure consistency in the decoding process through a manifold alignment mechanism. In the specific implementation process, first, the feature map of the bottleneck layer of the encoder is taken as the high-dimensional feature manifold In order to maintain the geometric structure of the manifold in the feature space, a manifold modeling method based on the geodesic distance is adopted. For each feature vector in the feature map, a local neighborhood is constructed through a K-neighbor algorithm, and the neighborhood size is usually set to 8 to 15. The Euclidean distance between the feature points in the local neighborhood is calculated, and the geodesic distance on the manifold is estimated through the Isomap algorithm to construct a geodesic distance matrix . Based on the geodesic distance matrix, a local connection graph of the feature manifold is established , in which the nodes represent the feature vectors, and the weight of the edge reflects the geodesic distance, thereby accurately depicting the manifold geometric relationship between the features.

[0105] In order to realize the differentiable mapping from the feature space of the encoder to the feature space of the decoder, a lightweight manifold mapping network is designed, which is composed of three fully connected layers, each followed by a LeakyReLU activation function. The mapping network receives the encoder features as input and outputs the corresponding decoder feature space representation . The training target of the mapping network is to minimize the change of the geodesic distance between the feature points before and after mapping, so as to ensure that the intrinsic geometric structure of the manifold is maintained in the mapping process.

[0106] The decoder also adopts a 5-layer structure, each layer containing a transpose convolutional layer for upsampling, a skip connection with the corresponding layer feature of the encoder, and two convolutional layers. The last layer uses convolution and Sigmoid activation to output the segmentation probability map. In order to ensure the geometric structure consistency of the feature space, a manifold alignment loss is designed, which includes two parts: a geodesic distance preservation term and a local reconstruction term. The geodesic distance preservation term ensures that the geodesic distance relationship of the encoder and decoder features on their respective manifolds is consistent, and the calculation formula is:

[0107] ,

[0108] wherein, ​the first summation symbol denotes the accumulation over 5 levels, is the weight factor of the i-th level, denotes the accumulation over all feature point pairs in the neighborhood set of the i-th level, and is a feature point pair within the neighborhood, denotes the geodesic distance in the encoder feature space, denotes the geodesic distance in the decoder feature space, denotes the manifold mapping network, denotes the squared Euclidean norm.

[0109] The local reconstruction term is based on the principle of local linear embedding, which ensures that the local neighborhood structure of the feature points is preserved after mapping, and the calculation formula is:

[0110] ,

[0111] wherein, is the local reconstruction loss, the summation symbol denotes the accumulation over 5 levels, is the weight factor of the i-th level, is the i-th level feature map of the decoder, denotes the neighborhood set of the i-th level feature, is the local reconstruction weight, which is obtained by minimizing the reconstruction error of the features within the neighborhood and satisfies the constraint condition , is the manifold mapping network, is the neighborhood feature of the corresponding layer of the encoder, denotes the squared Euclidean norm. The total loss of manifold alignment is:

[0112]

[0113] wherein, is the total loss of manifold alignment, and is the balance coefficient, which controls the relative weight of the two loss terms, and in the preferred embodiment can be set to , . By optimizing this loss function through end-to-end training, the manifold mapping network can learn a feature transformation that preserves the consistency of the geometric structure, and the entire manifold alignment mechanism is completely differentiable, supports gradient backpropagation, and realizes end-to-end training optimization.

[0114] ​The adaptive skip connection unit 33 connects the topology-sensitive encoder 31 and the manifold-aligned decoder 32, and is used to dynamically adjust the weight of the skip connection based on the feature distribution similarity and the topological feature similarity. In the traditional U-Net network, the skip connection is static and cannot adapt to different lesion types. The present application designs an adaptive weight function. For the encoder feature map and the decoder feature map , the feature distribution similarity and the topological feature similarity are calculated, and the total similarity is obtained by comprehensive consideration:

[0115] ,

[0116] wherein, is the total similarity of the i-th layer, and the value range is [0, 1], is a balance factor, which controls the relative importance of the feature distribution similarity and the topological feature similarity, and is usually set to 0.5-0.7, is the feature distribution similarity, which is calculated based on the cosine similarity of the feature vector, is the topological feature similarity, which is calculated based on the Betti number similarity. The adaptive weight function is defined as:

[0117] ,

[0118] wherein, is the weight of the i-th layer skip connection, and the value range is (0, 1), is an S-shaped activation function, which is defined as , is a sensitivity factor, which controls the sensitivity of the weight to the similarity, and is usually set to 2.0-5.0. The feature fusion formula of the skip connection is:

[0119] ,

[0120] wherein, is the i-th layer fused feature map, is the weight of the i-th layer skip connection, is the i-th layer encoder feature map, is the i-th layer decoder feature map. The updated decoder feature map is generated by integrating the fused feature through 1x1 convolution.

[0121] As shown in Figure 4 , the curvature-vascular correlation analysis module 4 is connected with the topology-preserving manifold learning module 3, and is used to receive the fine segmentation result, establish the correlation model between the curvature change of the tissue surface and the vascular morphology, analyze the differential morphological features, and construct the comprehensive feature descriptor.

[0122] The curvature-vascular correlation analysis module 4 includes a vascular-curvature co-detection unit 41, a differential morphology analysis unit 42, and a feature tensor construction unit 43.

[0123] The blood vessel-curvature co-detection unit 41 is used to extract blood vessel feature maps and curvature feature maps, and the two types of features are mutually enhanced through a collaborative attention mechanism. The blood vessel-curvature co-detection unit 41 is implemented in the following way:

[0124] The enhanced cervical image was processed using a dedicated vascular enhancement filter to extract vascular feature maps.

[0125] Use the curvature feature map as the curvature channel input;

[0126] Design transformation functions to map curvature features to the vascular feature space and vascular features to the curvature feature space, respectively;

[0127] Calculate attention weights to generate vascular features with enhanced curvature and vascular curvature features with enhanced curvature, respectively;

[0128] A collaborative feature map is generated by fusing enhanced features through adaptive weighting factors.

[0129] In a preferred embodiment of the present invention, the dedicated vascular enhancement filter is an improvement upon the Frangi filter and is optimized for the characteristics of cervical blood vessels:

[0130] ,

[0131] in, For scale The vascular response value is set below, ranging from [0,1]. A larger value indicates a higher probability that the point belongs to a blood vessel. and The eigenvalues ​​of the Hessian matrix , The eigenvalue ratio reflects the shape characteristics of the structure. The magnitude of the second derivative structure reflects the contrast of the structure. It is a natural exponential function. and These are parameters that control sensitivity to blood vessel shape and contrast. For cervical blood vessels, the preferred settings are... , The filter operates at multiple scales, capturing blood vessels of different diameters, and ultimately generating a vascular feature map. .

[0132] For the collaborative attention mechanism, the following transformation function is designed:

[0133] ,

[0134] ,

[0135] where, is the transformation result of curvature feature mapping to vessel feature space, is the transformation result of vessel feature mapping to curvature feature space, and is the learnable transformation matrix, with size and , where and are the dimensions of curvature feature and vessel feature respectively, and are the curvature feature and vessel feature after 3x3 convolution processing, and the calculation formula of attention weight is:

[0136] ,

[0137] ,

[0138] where, is the attention weight map of vessel feature, is the attention weight map of curvature feature, is the S-type activation function, which maps the value to the interval (0, 1). The enhanced feature is calculated as:

[0139] ,

[0140] ,

[0141] where, is the curvature-enhanced vessel feature, is the vessel-enhanced curvature feature, represents element-wise multiplication (Hadamard product). The final collaborative feature is fused by adaptive weight factors:

[0142] ,

[0143] where, is the fused collaborative feature, and are adaptive weight factors, which are dynamically adjusted according to image characteristics, satisfying In practical applications, the weight factor can be designed based on the information entropy of the feature map:

[0144] ,

[0145] ,

[0146] where, denotes the information entropy of denotes the information entropy of The information entropy formula is , where is the normalized histogram of eigenvalues. The final collaborative feature map is generated by integrating the fused features through 1x1 convolution .

[0147] The differential morphological analysis unit 42 is connected with the vessel-curvature collaborative detection unit 41, and is used for calculating Minkowski functionals and fractal dimensions by multi-scale morphological processing to quantify the morphological complexity of the abnormal region. In the specific implementation process, first, a set of morphological operators is defined erosion, dilation, opening operation, and closing operation and a set of structure elements containing structure elements of different sizes and shapes are defined. Various morphological operators and structure element combinations are applied to the segmentation result to generate a morphological feature sequence .

[0148] For each morphologically processed image , Minkowski functionals are calculated, including area , perimeter , and Euler number . The area represents the total area of the lesion region, the perimeter represents the boundary length of the lesion region, and the Euler number is equal to the number of connected components minus the number of holes, reflecting the topological complexity of the region. The Minkowski feature vector is constructed.

[0149] In addition, the fractal dimension of the lesion region is estimated using the box counting method :

[0150] ,

[0151] where is the fractal dimension, represents the limit when tends to 0, is the side length of the box, is the number of boxes with a side length of required to cover the lesion region, denotes the natural logarithm. In actual calculation, estimation is performed at multiple scales, and a more stable fractal dimension value is obtained through linear regression. The fractal dimension reflects the irregularity and complexity of the lesion tissue, and is usually between 1.2-1.4 for normal tissue, and can reach 1.7-1.9 for cancerous tissue.

[0152] ​The feature tensor construction unit 43 is connected to the differential morphology analysis unit 42 to integrate all extracted features, construct a feature correlation graph, and analyze the interaction strength between different features. In the specific implementation, all extracted features are first collected: segmentation results... Synergistic features Minkowski characteristics and fractal dimension Construct feature vectors To reduce dimensionality, principal component analysis was applied, retaining 95% of the variance information.

[0153] Then, construct the feature association graph. Nodes represent features, and edges represent the correlations between features. Mutual information between features... Used to quantify nonlinear correlation:

[0154] ,

[0155] in, Features and Mutual information between them Indicates the feature and Double summation of all possible values, It is a feature and The joint probability distribution, and It is a marginal probability distribution. This represents the natural logarithm. By setting a mutual information threshold (typically 0.3-0.5), highly correlated feature groups are identified, forming feature clusters.

[0156] Finally, a unique identifier is assigned to each anomalous region, and the center coordinates, direction, and size of the region are recorded. A spatial relationship tree is constructed to describe the relative positions between anomalous regions. The output is a comprehensive feature descriptor D, which includes a segmentation mask, feature vectors, spatial information, and a relationship graph.

[0157] like Figure 5 As shown, the biopsy point precise positioning module 5 is connected to the curvature-vascular correlation analysis module 4, which is used to receive comprehensive feature descriptors, analyze abnormal area features, and determine the optimal biopsy location.

[0158] The biopsy site precise positioning module 5 includes a feature comprehensive analysis unit 51, a biomarker association unit 52, and a biopsy site optimization unit 53.

[0159] The feature comprehensive analysis unit 51 is configured to analyze the comprehensive feature descriptor, evaluate the severity of the abnormal region and the feature saliency. In an embodiment of the present application, by comprehensively considering the curvature feature, the blood vessel feature and the morphological feature, a risk score function of the abnormal region is constructed:

[0160] ,

[0161] wherein, is the risk score of the region R, with a value range of [0, 1], and the higher the value, the greater the risk of the lesion, 、 and respectively represent the curvature abnormality degree, the blood vessel abnormality degree and the morphological abnormality degree of the region R, which are all normalized to the interval [0, 1], 、 and are weight coefficients, which control the contribution of each feature to the total risk score, and satisfy . For cervical lesions, the preferred weight settings are = 0.3, = 0.4, = 0.3, which shows good risk assessment ability in clinical verification.

[0162] The biomarker association unit 52 is connected with the feature comprehensive analysis unit 51, and is configured to perform association analysis on the features of the abnormal region and the known biomarkers of cervical lesions. In cervical lesions, the main biomarkers include abnormal blood vessel patterns, atypical glands and abnormal epithelial changes. This unit establishes a feature-biomarker mapping relationship, for example, a high curvature region is related to abnormal epithelial changes, a blood vessel network complexity is related to abnormal blood vessel patterns, and a specific morphological feature is related to atypical glands. Through this mapping, the extracted features are converted into biomarker indicators with clinical significance.

[0163] The biopsy point optimization unit 53 is connected with the biomarker association unit 52, and is configured to determine the best biopsy location and the number of biopsies based on the features and biomarker association results of the abnormal region, and in combination with clinical biopsy rules. This unit adopts a multi-objective optimization strategy, and the objective function includes:

[0164] maximizing the representativeness of the biopsy point: selecting a position that can best reflect the characteristics of the lesion;

[0165] minimizing the number of biopsies: reducing unnecessary biopsies while ensuring diagnostic accuracy;

[0166] following the clinical biopsy rules: such as the Reid Colposcopic Index rule, giving priority to regions with obvious blood vessel abnormalities and significant color changes.

[0167] The mathematical expression of the optimization problem is:

[0168] ,

[0169] ,

[0170] where, represents finding the maximum value in all possible subsets P of the region R, represents a representative measure of the biopsy point set P, defined as , is the number of biopsies, and is a balancing factor that controls the relative importance of representativeness and biopsy number, represents a clinical rule constraint, which is a Boolean value, true when the biopsy point set P satisfies the clinical rule, and false otherwise. In practical applications, can be set to 0.7-0.8, 0.2-0.3 to achieve a good balance between representativeness and biopsy number.

[0171] As shown in Figure 6 , the real-time intelligent auxiliary diagnosis module 6 is connected with the biopsy point accurate positioning module 5, used to receive the optimal biopsy position and comprehensive feature descriptor, generate the cervical lesion diagnosis result and diagnosis basis, and provide auxiliary diagnosis information.

[0172] The real-time intelligent auxiliary diagnosis module 6 includes a lesion classification unit 61, a diagnosis basis generation unit 62, and a visualization display unit 63.

[0173] The lesion classification unit 61 is used to classify the cervical state as normal, cervical intraepithelial neoplasia, and cervical cancer based on the comprehensive feature descriptor. This unit uses an ensemble learning method to combine the results of multiple classifiers to improve the robustness of classification. In specific implementation, random forest, support vector machine, and gradient boosting tree classifiers are used, and each classifier is trained for different feature subsets. The final classification result is determined based on weighted voting:

[0174] ,

[0175] where, is the classification result of the region R, represents finding the class c that maximizes the following expression, represents the summation operation from i=1 to i=k, where k is the number of classifiers, is the weight of the i-th classifier, satisfying , Pi,c represents the probability prediction of the i-th classifier that the region R belongs to class c. The weights can be determined according to the performance of each classifier on the validation set, for example, the weights of the classifiers with better performance can be set to 0.4-0.5, and the weights of the classifiers with general performance can be set to 0.2-0.3.

[0176] The diagnosis basis generation unit 62 is connected with the lesion classification unit 61, and is used to extract key features and evidence supporting the diagnosis result, and generate diagnosis basis information. This unit identifies the features that contribute most to the classification result through feature importance analysis, and associates these features with clinical diagnosis criteria. For example, for cervical intraepithelial neoplasia, key features can include specific blood vessel patterns, abnormal gland openings, and irregular edges of transformation zones. The diagnosis basis is ranked by importance to form a structured diagnosis report.

[0177] The visualization display unit 63 is connected with the diagnosis basis generation unit 62, and is used to visually display the cervical lesion diagnosis result, diagnosis basis, and optimal biopsy location on the enhanced cervical image. This unit adopts a multi-layer visualization strategy, including:

[0178] The base layer displays the enhanced original cervical image;

[0179] The segmentation layer displays the abnormal area using a semi-transparent color, and different colors represent different risk levels;

[0180] The feature layer marks the positions of key features such as abnormal blood vessels and atypical glands;

[0181] The biopsy layer marks the recommended biopsy point position and indicates the priority.

[0182] The visualization interface also provides interactive functions, allowing doctors to adjust the display content and view detailed feature analysis results to assist in making more accurate diagnosis decisions.

[0183] As shown in Figure 7 The system of the present application also includes a clinical feedback module 7 and an adaptive optimization module 8.

[0184] The clinical feedback module 7 is connected with the real-time intelligent auxiliary diagnosis module 6, and is used to receive the feedback information of the doctor on the diagnosis result. This module designs a structured feedback interface, allowing doctors to evaluate the diagnosis accuracy of the system, the rationality of the biopsy point recommendation, and provide specific improvement suggestions.

[0185] The adaptive optimization module 8 is connected with the clinical feedback module 7, and is used for adjusting parameters of the differential geometry feature extraction module 2, the topological preserving manifold learning module 3, the curvature-vascular correlation analysis module 4, the biopsy point accurate positioning module 5 and the real-time intelligent auxiliary diagnosis module 6 according to the feedback information, so as to improve the system diagnosis accuracy. The module adopts an incremental learning strategy, continuously optimizes the system parameters, and enables the system to adapt to different clinical environments and equipment conditions. The optimization process considers the consistency and reliability of the feedback, and gives higher weight to the highly consistent feedback.

[0186] Figure 7 The overall data flow and processing flow of the system of the present application are shown. First, the image enhancement module 1 processes the acquired cervical endoscope images, and outputs enhanced cervical images. Then, the differential geometry feature extraction module 2 receives the enhanced images, extracts curvature features, and outputs curvature feature maps and preliminary segmentation results. Next, the topological preserving manifold learning module 3 performs deep learning based on these features, preserves key topological structures, and outputs fine segmentation results. Subsequently, the curvature-vascular correlation analysis module 4 establishes the correlation between tissue morphology and vascular features, and constructs comprehensive feature descriptors. The biopsy point accurate positioning module 5 determines the optimal biopsy location according to the feature analysis, and finally the real-time intelligent auxiliary diagnosis module 6 generates diagnosis results and basis, and provides visual display. The whole process forms a closed-loop system, and the performance is continuously optimized through clinical feedback.

[0187] In the preferred embodiment of the present application, the collected gynecological endoscope image data generally has a resolution of 512x512 pixels, an RGB color format, and includes different types such as normal cervical tissue, inflammatory areas, low-grade lesions and high-grade lesions. The processing time of the system for the image is controlled within 2 seconds, ensuring that real-time analysis can be supported. The accuracy of curvature feature extraction reaches the pixel level, and can identify minor morphological changes. The training of the topological preserving network uses 5000 labeled images, and the verification accuracy reaches more than 92%. The overall diagnostic accuracy of the system is improved by 23.5% compared with traditional methods, and the sensitivity in early lesion detection is improved by 31.2%.

[0188] In summary, the gynecological endoscope image intelligent analysis and cervical lesion accurate diagnosis system provided by the present application realizes accurate analysis of cervical endoscope images and accurate diagnosis of lesions through innovative technologies such as differential geometry feature extraction, topological preserving manifold learning and curvature-vascular correlation analysis, significantly improves the detection rate of cervical lesions, especially early and small lesions, and provides a powerful intelligent auxiliary diagnosis tool for clinicians.

[0189] The above embodiments only express the specific implementation of the present application, which is described in more detail and in more detail, but cannot be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A gynecological endoscopic image intelligent analysis and precise diagnosis system for cervical lesions, characterized in that, The method comprises the following steps: An image enhancement module is used to process the obtained endocervical image, improve the image contrast, eliminate reflections and shadows, and perform color correction to obtain an enhanced cervical image; A differential geometric feature extraction module is connected with the image enhancement module and is used to receive the enhanced cervical image, model the surface of the enhanced cervical image as a Riemannian manifold, calculate multi-scale curvature features, extract features of abnormal regions, and output a curvature feature map and a preliminary segmentation result of the cervical image; A topological manifold learning module is connected with the differential geometric feature extraction module and is used to receive the curvature feature map and the preliminary segmentation result, process the feature map through a topologically sensitive encoder and a manifold alignment decoder, maintain key topological structures, and output a fine segmentation result; A curvature-vascular correlation analysis module is connected with the topological manifold learning module and is used to receive the fine segmentation result, establish a correlation model between the curvature change of the tissue surface and the vascular morphology, analyze differential morphological features, and construct a comprehensive feature descriptor; A biopsy point accurate positioning module is connected with the curvature-vascular correlation analysis module and is used to receive the comprehensive feature descriptor, analyze the features of abnormal regions, and determine the optimal biopsy position; A real-time intelligent auxiliary diagnosis module is connected with the biopsy point accurate positioning module and is used to receive the optimal biopsy position and the comprehensive feature descriptor, generate a cervical lesion diagnosis result and diagnosis basis, and provide auxiliary diagnosis information.

2. The system of claim 1, wherein, The differential geometric feature extraction module comprises: A manifold modeling unit is used to map the surface of the enhanced cervical image to a parameterized surface, establish a local differential coordinate system, and calculate a local metric tensor; A curvature calculation unit is connected with the manifold modeling unit and is used to calculate the Gaussian curvature and the mean curvature of the parameterized surface at multiple scales to generate a curvature feature map; An abnormal region preliminary positioning unit is connected with the curvature calculation unit and is used to determine the preliminary position of the abnormal region based on the curvature feature map through adaptive threshold segmentation and curvature gradient field analysis.

3. The system of claim 2, wherein, The curvature calculation unit is implemented in the following manner: For the parameterized surface, a sequence of scales is set , wherein ; for each scale processing the parameterized surface with a Gaussian filter corresponding to the kernel width; Curvature maps are calculated at each scale to obtain a curvature feature set corresponding to a scale sequence; The curvature features at multiple scales are fused through an adaptive weight function to obtain a fused curvature feature map.

4. The system of claim 1, wherein, The topological manifold learning module comprises: A topologically sensitive encoder is used to extract multi-layer features from the curvature feature map and the enhanced cervical image, and calculate persistent homology features after each layer of extraction to supervise the topological structure of the feature space; A manifold alignment decoder is connected with the topologically sensitive encoder and is used to construct a feature manifold and ensure the geometric structure consistency in the decoding process through a manifold alignment mechanism; An adaptive skip connection unit is connected with the topologically sensitive encoder and the manifold alignment decoder and is used to dynamically adjust the weight of the skip connection based on the feature distribution similarity and the topological feature similarity.

5. The system of claim 4, wherein, The topologically sensitive encoder calculates the persistent homology features in the following manner: Filter complexes are constructed on each layer of convolution feature maps to capture the topological structure of the feature space; Compute 0-dimensional and 1-dimensional Betti numbers, representing the number of connected components and loops; Construct a topological feature vector and compare it with the topological feature vector of the original image to calculate a topological similarity index; Based on the topological similarity index, construct a topological loss term to guide the network to maintain the key topological structure.

6. The system of claim 1, wherein, The curvature-vascular correlation analysis module includes: A vascular-curvature collaborative detection unit is configured to extract a vascular feature map and a curvature feature map, and to enhance the two types of features through a collaborative attention mechanism; A differential morphological analysis unit is connected to the vascular-curvature collaborative detection unit and is configured to calculate Minkowski functionals and fractal dimensions through multi-scale morphological processing to quantify the morphological complexity of the abnormal region; A feature tensor construction unit is connected to the differential morphological analysis unit and is configured to integrate all extracted features to construct a feature correlation graph and analyze the interaction strength between different features.

7. The system of claim 6, wherein, The vascular-curvature collaborative detection unit is implemented in the following manner: A dedicated vascular enhancement filter is used to process the enhanced cervical image to extract a vascular feature map; The curvature feature map is input as a curvature channel; A conversion function is designed to map the curvature features to the vascular feature space and to map the vascular features to the curvature feature space, respectively; Attention weights are calculated to generate curvature-enhanced vascular features and vascular-enhanced curvature features, respectively; The enhanced features are fused through an adaptive weight factor to generate a collaborative feature map.

8. The system of claim 1, wherein, The biopsy point precise positioning module includes: A feature comprehensive analysis unit is configured to analyze the comprehensive feature descriptor to evaluate the severity and feature saliency of the abnormal region; A biomarker association unit is connected to the feature comprehensive analysis unit and is configured to perform association analysis on the features of the abnormal region and known cervical lesion biomarkers; A biopsy point optimization unit is connected to the biomarker association unit and is configured to determine the optimal biopsy location and the number of biopsies based on the features of the abnormal region and the biomarker association results, combined with clinical biopsy rules.

9. The system of claim 1, wherein, The real-time intelligent auxiliary diagnosis module includes: A lesion classification unit is configured to classify the cervical state as normal, cervical intraepithelial neoplasia, and cervical cancer based on the comprehensive feature descriptor; A diagnosis basis generation unit is connected to the lesion classification unit and is configured to extract key features and evidence supporting the diagnosis result to generate diagnosis basis information; A visual display unit is connected to the diagnosis basis generation unit and is configured to visually display the cervical lesion diagnosis result, diagnosis basis, and optimal biopsy location on the enhanced cervical image.

10. The system of claim 1, wherein, The system further includes: A clinical feedback module is connected to the real-time intelligent auxiliary diagnosis module and is configured to receive feedback information from doctors on the diagnosis result; An adaptive optimization module is connected to the clinical feedback module and is configured to adjust the parameters of the differential geometric feature extraction module, the topological preserving manifold learning module, the curvature-vascular correlation analysis module, the biopsy point precise positioning module, and the real-time intelligent auxiliary diagnosis module according to the feedback information to improve the diagnosis accuracy of the system.

Images