Common intrauterine diseases: hysteroscopic diagnostic image recognition enhancement methods

Abstract

This invention relates to the field of intelligent hysteroscopic imaging technology, specifically a method for image recognition and enhancement in hysteroscopic diagnosis of common intrauterine diseases. The method includes: synchronously acquiring a continuous stream of raw images and physical motion trajectories via hysteroscopy; performing spatiotemporal continuity analysis to generate a dynamic deformation mapping atlas; and identifying and marking areas of abnormal tissue compliance. Multiple virtual scanning windows are generated based on the contours of these abnormal areas. The image information entropy and expected feature abundance values ​​for each window are calculated, and an acquisition urgency level is assigned. An execution queue is constructed based on the urgency level and equipment physical constraints, driving the device to perform sequential pause-focusing scans to acquire high-resolution locally enhanced images. The local images are then feature-aligned and stitched with the raw image stream to synthesize an enhanced fusion diagnostic image covering the entire uterine cavity. This method can objectively assess the mechanical properties of uterine tissue and intelligently and adaptively allocate scanning resources, thereby improving the information content of the images and diagnostic efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent hysteroscopic imaging technology, and in particular to a method for enhancing hysteroscopic diagnostic images of common intrauterine diseases. Background Technology

[0002] Currently, the hysteroscopic diagnostic techniques widely used in clinical practice mainly rely on the experience of the operating physician, making diagnoses by observing real-time video images on a monitor. Existing technologies typically only provide a continuous two-dimensional optical image stream, and the diagnostic process is highly dependent on the physician's visual judgment and subjective experience. The image acquisition mode passively follows the physician's manual operation, lacking objective means of assessing the biomechanical properties of uterine tissue.

[0003] Existing technologies have shortcomings. Traditional hysteroscopic images cannot quantitatively reflect the mechanical properties of tissues, such as elasticity and compliance. They struggle to objectively identify and define changes in tissue stiffness and abnormal mobility associated with lesions like submucosal fibroids, endometrial polyps, and adhesions, easily leading to missed diagnoses of early or minute lesions. Furthermore, image acquisition during the examination lacks intelligent guidance. Doctors need to repeatedly and time-consuming zoom in and manually focus to obtain local details. The scanning path and focus points throughout the uterine cavity depend entirely on personal habits, resulting in inconsistent examination efficiency and the potential for missing fine images of critical areas due to uneven attention distribution.

[0004] This invention aims to address the challenge of upgrading hysteroscopy from a mere morphological observation tool into a diagnostic system capable of dynamically sensing and intelligently scanning tissue biomechanical properties. Its core lies in overcoming the limitations of static image analysis and achieving adaptive priority allocation of examination resources to high-value diagnostic areas. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method for enhancing the recognition of hysteroscopic diagnostic images of common intrauterine diseases.

[0006] To achieve the above objectives, the present invention employs the following technical solution: a method for enhancing hysteroscopic diagnostic images of common intrauterine diseases, comprising:

[0007] By simultaneously acquiring continuous raw image streams and physical motion trajectories using hysteroscopic equipment, and performing spatiotemporal continuity analysis, dynamic deformation mapping maps are generated.

[0008] Based on the dynamic deformation mapping atlas, regions with abnormal tissue compliance are identified, and the boundary contours of these regions are marked in the three-dimensional uterine cavity space.

[0009] Based on the shape complexity and area size of the boundary contour, multiple virtual scanning windows with different field of view are dynamically generated;

[0010] Based on the type of the abnormal tissue compliance region and the spatial location of the virtual scanning window, calculate the image information entropy and expected value of feature abundance for each virtual scanning window;

[0011] The image information entropy and the expected value of feature abundance are combined with a preset diagnostic confidence threshold to assign a collection urgency level to each virtual scanning window;

[0012] Based on the urgency level of the data acquisition and the physical movement constraints of the hysteroscopic equipment, an ordered virtual scanning window execution queue is constructed.

[0013] The hysteroscopic device is driven to perform pause-focus scanning on the target uterine cavity region sequentially according to the virtual scanning window execution queue, thereby acquiring a set of high-resolution locally enhanced images;

[0014] The high-resolution locally enhanced image is feature-aligned and stitched with the continuous original image stream to synthesize an enhanced fusion diagnostic image covering the entire uterine cavity.

[0015] As a further aspect of the present invention, the step of simultaneously acquiring a continuous raw image stream and physical motion trajectory through a hysteroscopic device, and performing spatiotemporal continuity analysis to generate a dynamic deformation mapping map includes:

[0016] The hysteroscopic device acquires a continuous stream of raw images of the target uterine cavity while in continuous motion, and simultaneously records the physical motion trajectory of the hysteroscopic device.

[0017] Spatiotemporal continuity analysis is performed on the continuous raw image stream to extract the parallax changes and tissue texture drift features between image frames, and the physical motion trajectory is fused to generate a dynamic deformation mapping map of the endometrial surface.

[0018] The process involves performing spatiotemporal continuity analysis on the continuous raw image stream, extracting parallax changes and tissue texture drift features between image frames, fusing the physical motion trajectory, and generating a dynamic deformation mapping map of the endometrial surface. Specifically:

[0019] The continuous raw image stream is decomposed into continuous image frame pairs. Dense optical flow algorithm is applied to each image frame pair to calculate the motion vector of each pixel between consecutive frames, forming dense optical flow field data.

[0020] Motion pattern decomposition was performed on the dense optical flow field data to separate the background flow field caused by the global motion of the hysteroscopic equipment and the characteristic flow field caused by the local deformation of the endometrium.

[0021] Cluster analysis is performed on the characteristic flow field to identify flow field regions with consistent motion directions, and each region is defined as a local deformation unit.

[0022] Synchronously read the physical motion trajectory data corresponding to the timestamp of each image frame, the physical motion trajectory data including the three-dimensional spatial coordinates and attitude angle of the end of the hysteroscopic device;

[0023] Using inverse kinematics algorithm, the physical motion trajectory data is converted into a theoretical observation vector of the endometrial surface by the lens of the hysteroscopic device;

[0024] The motion vector of the local deformation unit is compared with the theoretical observation vector, and the deviation between the actual motion and the theoretical observation of each local deformation unit is calculated.

[0025] Based on the magnitude and direction of the deviation, each local deformation unit is assigned a deformation intensity coefficient and a deformation mode label;

[0026] The deformation intensity coefficients and deformation mode labels of all local deformation units are mapped back to the initial spatial position on the surface of the endometrium, constructing a dynamic deformation mapping map containing spatiotemporal dimensions. The dynamic deformation mapping map records the dynamic response characteristics of each point on the surface of the endometrium during the movement of the device.

[0027] As a further aspect of the present invention, the step of identifying tissue compliance abnormalities based on the dynamic deformation mapping map and marking the boundary contours of the tissue compliance abnormalities in the three-dimensional uterine cavity space specifically includes:

[0028] In the dynamic deformation mapping map, a normal variation range threshold for the deformation intensity coefficient is set, and continuous regions where the deformation intensity coefficient continuously exceeds the normal variation range threshold are initially screened as candidate abnormal regions.

[0029] Analyze the distribution consistency of deformation pattern labels within each candidate anomaly region. If a candidate anomaly region contains more than a preset proportion of distinct deformation pattern labels, then the candidate anomaly region is divided into multiple sub-regions.

[0030] For each candidate anomaly region or its sub-region, calculate the statistical distribution of the deformation intensity coefficients of all points within it, including the mean, variance, and gradient rate of change.

[0031] Based on the statistical distribution, a multidimensional feature vector is constructed, which is used to characterize the deformation inhomogeneity and dynamic stability of the candidate anomaly region.

[0032] The multidimensional feature vector is input into a pre-trained tissue compliance classification model, which outputs the probability that each region belongs to a rigid region, an adhesion region, an edematous region, or a normal region.

[0033] Areas classified as rigid, adhesion, or edema areas are formally defined as areas of abnormal tissue compliance.

[0034] Obtain the initial set of two-dimensional image coordinates corresponding to the abnormal tissue compliance region in the continuous raw image stream;

[0035] Using the depth sensor data from the hysteroscopic device, the initial two-dimensional image coordinate set is back-projected onto a three-dimensional uterine cavity space model constructed from physical motion trajectories;

[0036] In the three-dimensional uterine cavity space model, the concave hull algorithm is used to process the back-projected three-dimensional point set to generate the minimum closed surface that can enclose all three-dimensional points. The minimum closed surface is the boundary contour of the tissue compliance abnormal region in the three-dimensional uterine cavity space.

[0037] As a further aspect of the present invention, the dynamic generation of multiple virtual scanning windows with different field of view ranges based on the shape complexity and area size of the boundary contour is specifically as follows:

[0038] Geometric features are extracted from the boundary contour of each region with abnormal tissue compliance. The circularity, aspect ratio, and edge curvature of the boundary contour are calculated, and the shape complexity score of the boundary contour is obtained by combining them.

[0039] Calculate the area of ​​the projected surface of the boundary contour in the three-dimensional uterine cavity space model, and use it as the area size of the abnormal tissue compliance region;

[0040] Establish a table showing the correspondence between shape complexity score, area size, and basic field of view size of the virtual scanning window. The higher the shape complexity or the larger the area, the smaller the corresponding basic field of view.

[0041] Based on the corresponding table, a basic field of view size for a virtual scanning window is determined for each region of tissue compliance abnormality.

[0042] The geometric center of the boundary contour of each region with abnormal tissue compliance is used as the center point of the viewport;

[0043] Based on the basic field of view size, in the three-dimensional uterine cavity space model, a square pyramid-shaped observation space is constructed with the center point of the window as the vertex. The size of the rectangular base of the observation space is determined by the basic field of view size.

[0044] Adjust the observation direction of the observation space so that its axis is aligned with the optimal observation posture that the hysteroscopic device can reach at the center point of the window.

[0045] The observation space volume is defined as a virtual scanning window covering the region of abnormal tissue compliance. The virtual scanning window includes parameters such as spatial location, observation direction, and field of view.

[0046] As a further aspect of the present invention, the step of calculating the image information entropy and expected feature abundance value corresponding to each virtual scanning window based on the type of the tissue compliance abnormality region and the spatial position of the virtual scanning window specifically involves:

[0047] Based on the type of abnormal tissue compliance region, the image texture feature library and morphological feature library of typical diseases are called from the pre-stored knowledge base;

[0048] For each virtual scanning window, historical image segments located within the window area are extracted from the continuous raw image stream based on its spatial location parameters;

[0049] Multi-scale filtering and feature point detection are performed on the historical image segments. The types of textures and the number of morphological features that have been revealed are counted. The results are compared with the typical image texture feature library and morphological feature library to calculate the feature manifestation ratio.

[0050] Based on the feature display ratio and the basic field of view size of the virtual scanning window, the number of unknown or fuzzy features captured within the window range is estimated as a potential feature increment.

[0051] Calculate the image information entropy by combining the grayscale distribution of the historical image fragments themselves;

[0052] The image information entropy and the latent feature increment are weighted and fused to obtain the expected value of the feature abundance of the virtual scanning window, wherein the weight of the latent feature increment is higher than the weight of the historical image information entropy.

[0053] For newly discovered virtual scan windows that are not covered by historical image fragments, the average feature abundance of the corresponding tissue compliance anomaly region type is used as the initial estimate of its feature abundance expected value.

[0054] As a further aspect of the present invention, the step of combining the image information entropy and the expected value of feature abundance with a preset diagnostic confidence threshold to assign a collection urgency level to each virtual scanning window specifically involves:

[0055] Multiple diagnostic confidence threshold intervals are set, and each threshold interval corresponds to the minimum amount of feature information required for a diagnosis.

[0056] The current image information entropy of each virtual scanning window is compared with the minimum feature information amount. If the image information entropy is lower than the minimum feature information amount, it is determined that the current information of the virtual scanning window is insufficient.

[0057] For virtual scanning windows with insufficient current information, the difference between their expected feature abundance and the current image information entropy is further calculated. This difference reflects the potential for additional information to be obtained through enhanced scanning.

[0058] Based on the magnitude of the difference, the virtual scanning window is divided into a high information potential window, a medium information potential window, and a low information potential window;

[0059] The type weights of the abnormal tissue compliance regions are combined, and these type weights are pre-set according to the clinical risk level of the disease.

[0060] The urgency score is obtained by multiplying the information potential level by the type weight and then normalizing the result.

[0061] Based on the distribution of urgency scores, the virtual scanning window is divided into three acquisition urgency levels: high, medium, and low. The high acquisition urgency level indicates the window that needs to be scanned with priority and high quality.

[0062] As a further aspect of the present invention, the step of constructing an ordered virtual scanning window execution queue based on the acquisition urgency level and the physical motion constraints of the hysteroscopic equipment specifically involves:

[0063] Obtain the physical motion constraints of the hysteroscopic device, including the maximum turning angle, minimum travel distance, device length limit, and safe obstacle avoidance space;

[0064] Starting from the current position of the hysteroscopic device, all virtual scanning windows are considered as nodes to be accessed.

[0065] A heuristic graph search algorithm is used for path planning, with the center point of each virtual scanning window as a node in the graph. The movement cost between nodes is determined by the actual movement distance of the device and the complexity of attitude adjustment.

[0066] The collection urgency level is used as the priority weight for node access, and nodes with higher collection urgency levels receive higher access priority in path cost calculation.

[0067] The heuristic graph search algorithm, under the premise of satisfying physical motion constraints, finds an approximately optimal path that starts from the starting point, visits all nodes in sequence, and finally covers all nodes with high acquisition urgency level.

[0068] Based on the access order of nodes in the near-optimal path, a virtual scanning window execution queue is generated. The queue contains not only the execution order of the windows, but also a description of the transition trajectory of the hysteroscopic device moving from one window to the next.

[0069] As a further aspect of the present invention, the driving hysteroscopy device sequentially performs pause-focus scanning on the target uterine cavity region according to the virtual scanning window execution queue to acquire a set of high-resolution locally enhanced images, specifically:

[0070] The hysteroscopy device controller receives the virtual scanning window execution queue and parses the spatial position and observation direction parameters of the first virtual scanning window in the queue;

[0071] Control the hysteroscopic equipment to move from its current position along the transition trajectory specified in the queue to the observation position of the first virtual scanning window;

[0072] Adjust the orientation of the hysteroscopy equipment so that its lens optical axis is aligned with the observation direction of the virtual scanning window;

[0073] In the pause-focus scanning mode, the hysteroscopic device remains stationary, and the lens focal length and illumination intensity are adjusted to the optimal values ​​preset for the virtual scanning window. These optimal values ​​are pre-configured according to the urgency level of the acquisition and the tissue type.

[0074] While the device is stationary, multiple local images within the window area are acquired at a frame rate and resolution higher than that of the continuous raw image stream.

[0075] Real-time sharpness evaluation and motion artifact detection are performed on multiple acquired local images, and the image with the highest sharpness and no artifacts is selected as the high-resolution local enhancement image of the virtual scanning window.

[0076] Store the high-resolution locally enhanced image and its corresponding virtual scanning window identifier;

[0077] The hysteroscopic device is controlled to move to the next virtual scanning window in the queue order, and the pause-focusing scanning process, from adjusting the posture of the hysteroscopic device to aligning the observation direction to selecting and storing high-resolution local enhancement images, is repeated until all virtual scanning windows in the queue have been scanned, and finally a set of high-resolution local enhancement images corresponding one-to-one with the virtual scanning windows are obtained.

[0078] As a further aspect of the present invention, the step of performing feature alignment and stitching between the high-resolution locally enhanced image and the continuous original image stream to synthesize an enhanced fusion diagnostic image covering the entire uterine cavity specifically involves:

[0079] Extract scale-invariant feature transform feature points and their descriptors for each high-resolution locally enhanced image;

[0080] In the continuous raw image stream, locate the reference image frame that is closest to each high-resolution locally enhanced image in time and space;

[0081] Extract feature points and descriptors of the same type from the reference image frame;

[0082] A correspondence between feature points in a high-resolution locally enhanced image and feature points in a reference image frame is established using a feature descriptor matching algorithm.

[0083] Based on the correspondence pairs, the homography matrix transformation model is used to align each high-resolution local enhancement image to the global image coordinate system formed by the continuous original image stream;

[0084] A multi-band fusion algorithm is used to seamlessly stitch the aligned high-resolution local enhanced image with the corresponding region in the global image coordinate system, achieving a smooth transition of texture and color at the stitching boundary.

[0085] For the uterine cavity region not covered by the high-resolution local enhancement image, the image content of the corresponding region in the continuous original image stream is preserved;

[0086] All image blocks after stitching and fusion are integrated onto a unified image canvas to generate an enhanced fusion diagnostic image that is spatially continuous and has high-resolution details in key areas. The enhanced fusion diagnostic image preserves the complete uterine cavity topology.

[0087] As a further aspect of the present invention, the method further includes a step of performing quantitative analysis of intrauterine diseases based on the enhanced fusion diagnostic image, specifically:

[0088] Load the enhanced fusion diagnostic image covering the entire uterine cavity, and load the associated three-dimensional uterine cavity spatial model and the boundary contour data of the tissue compliance abnormality area;

[0089] On the enhanced fusion diagnostic image, image sub-regions of each tissue compliance abnormality region are segmented based on boundary contour data;

[0090] For each image sub-region, intensive feature extraction is performed. The extracted features include, but are not limited to, color histogram statistics, multi-directional Gabor filter response, local binary pattern histogram, and deep convolutional neural network activation features.

[0091] The extracted dense features are input into a multi-task analysis network, which performs two tasks in parallel: disease type classification and disease severity regression.

[0092] The multi-task analysis network outputs a disease type probability distribution vector for each region of tissue compliance abnormality, as well as a continuous numerical score representing the severity.

[0093] Based on the disease type probability distribution vector, the type with the highest probability is taken as the preliminary diagnostic type of the tissue compliance abnormality region;

[0094] By combining the continuous numerical score of the severity with the area size and shape complexity of the tissue compliance abnormality region, the contribution weight of the tissue compliance abnormality region in the overall intrauterine health assessment is calculated using a pre-calibrated quantitative formula.

[0095] By summarizing the diagnostic types, severity continuous numerical scores, and contribution weights of all areas with abnormal tissue compliance, a structured quantitative analysis report of intrauterine diseases is generated. This report describes the distribution, type, and severity of intrauterine diseases in data form.

[0096] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0097] By simultaneously acquiring the physical motion trajectory of the hysteroscopic device and performing spatiotemporal continuity analysis with a continuous image stream, a dynamic deformation mapping atlas is generated. This technology quantifies and visualizes the minute deformations generated during the interaction between the instrument and tissue, making it possible to explicitly express local differences in tissue compliance and elastic modulus that are traditionally invisible to the naked eye. Physicians can identify abnormal areas based on objective biomechanical maps, rather than relying solely on shape and color, providing a new diagnostic dimension for identifying features such as tiny protrusions of intramural fibroids into the uterine cavity and rigid edges of dense adhesions.

[0098] Based on image information entropy and expected feature abundance, the urgency level of data acquisition is calculated for each virtual scanning window, and an ordered execution queue is constructed accordingly. This mechanism simulates the decision-making logic of human experts who "carefully examine key areas" during examinations, but achieves fully automated quantification and optimization. It drives the device to automatically prioritize high-resolution focused scanning of complex, information-rich, or potentially malignant areas, while quickly passing through normal or homogeneous areas. Thus, within the constraints of limited examination time and the physical movement capabilities of the equipment, it maximizes the density of effective diagnostic information acquired in a single examination, avoiding the time waste or obscuration of critical details caused by indiscriminate uniform scanning. Attached Figure Description

[0099] Figure 1 This is a flowchart of the hysteroscopic image recognition enhancement method for common intrauterine diseases described in this invention;

[0100] Figure 2 A flowchart for generating a dynamic deformation mapping map;

[0101] Figure 3 A flowchart for dynamically generating virtual scanning windows;

[0102] Figure 4 A bar chart comparing the feature points extracted and matched from high-resolution images of each virtual scanning window;

[0103] Figure 5 This is a scatter plot showing the correlation between feature abundance and urgency level distribution. Detailed Implementation

[0104] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0105] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0106] See Figure 1 The hysteroscopic device synchronously acquires a continuous stream of raw images and their corresponding physical motion trajectories. Spatiotemporal continuity analysis is performed on the raw image stream to generate a mapping atlas reflecting the dynamic deformation characteristics of the endometrial surface. Based on this atlas, regions with abnormal tissue compliance are identified, and the boundary contours of these regions are precisely marked in a three-dimensional uterine cavity spatial model. Based on the geometric properties of the boundary contours, multiple virtual scanning windows with different field of view ranges are dynamically planned and generated. Combining the type of abnormal region and the spatial location of the window, the image information entropy and feature abundance of each window are estimated. Based on these estimated information and preset diagnostic confidence standards, an appropriate acquisition urgency level is assigned to each virtual scanning window. Taking into account both the acquisition urgency level and the physical motion limitations of the hysteroscopic device itself, an ordered virtual scanning window execution queue is planned and constructed. The hysteroscopic device is driven to move sequentially to designated positions according to this queue and perform pause-focused scanning to acquire a set of high-resolution locally enhanced images. These locally enhanced images are aligned and seamlessly stitched with the original continuous image stream at the feature level to synthesize a fused diagnostic image that covers the entire uterine cavity and enhances details in key areas.

[0107] In one embodiment of the present invention, see [reference] Figure 2The hysteroscopic device performs a continuous linear scan from the fundus to the internal cervical os within the uterine cavity. This process simulates the examination of suspected endometrial polyps. While the hysteroscopic device moves at a constant speed, the image sensor captures a continuous stream of raw images at a rate of 30 frames per second. Simultaneously, the motion sensor outputs the coordinate sequence and Euler angle attitude data of the hysteroscopic device's distal end in three-dimensional space in real time, forming a physical motion trajectory. The continuous raw image stream is decomposed chronologically, with each pair of adjacent images forming an image frame pair. For each image frame pair, the Farneback dense optical flow algorithm is applied to calculate the two-dimensional motion vector of each pixel between consecutive frames, forming a dense optical flow field data matrix of the same size as the image resolution. Motion mode decomposition is performed on the dense optical flow field data matrix. Robust principal component analysis is used to separate a low-rank background flow field matrix caused by the overall translation and rotation of the hysteroscopic device, and a sparse feature flow field matrix caused by local contraction of the endometrium and protrusion of proliferative tissue. Density-based spatial clustering analysis is performed on the sparse characteristic flow field matrix to identify pixel clusters whose motion vector directions and amplitudes are spatially continuous and consistent. Each pixel cluster is defined as a local deformation unit.

[0108] In some embodiments, the complex rotational motion of the hysteroscopic device during uterine horn exploration is processed. The synchronously read physical motion trajectory data includes the three-dimensional spatial coordinates and attitude angles of the hysteroscopic device's end effector at the corresponding moment in each frame of the image. The attitude angles are represented in quaternion form. Using an inverse kinematics algorithm, the spatial pose of the hysteroscopic device's end effector is converted into a direction vector of the lens optical axis in the uterine coordinate system. This direction vector is the theoretical observation vector. For each identified local deformation unit, the average vector of all pixel motion vectors within it is calculated, and this average vector is used as the actual motion vector of the local deformation unit. The actual motion vector of the local deformation unit is subtracted from the theoretical observation vector at the same timestamp. The magnitude of the resulting difference vector is the magnitude of the deviation, and the angle between the direction of the difference vector and the theoretical observation vector is the deviation direction. Based on the magnitude and direction of the deviation, each local deformation unit is assigned a deformation intensity coefficient and a deformation mode label. The deformation mode labels include three categories: "co-directional reinforcement," "reverse resistance," and "lateral slippage."

[0109] It is understandable that when processing the rapid passage of hysteroscopic equipment through the smooth mid-segment of the uterine cavity, the characteristic flow field matrix elements in the dense optical flow field data matrix are sparse, resulting in a small number of local deformation units and generally low deformation intensity coefficients. Conversely, when processing the slow passage of the hysteroscopic equipment through suspected adhesion areas, the characteristic flow field matrix elements are significant, the deformation intensity coefficients of the clustered local deformation units are high, and the deformation mode labels often exhibit "reverse resistance." All local deformation unit deformation intensity coefficients and deformation mode labels are mapped onto a two-dimensional parametric planar map unfolded from the 3D model of the uterine cavity, based on their pixel coordinates in the first frame image. For cases where the same spatial location is observed multiple times at different time stamps, the time series of its deformation intensity coefficients is recorded. The final constructed dynamic deformation mapping map is a multidimensional data set containing spatial two-dimensional coordinates, a time dimension, deformation intensity coefficients, and deformation mode labels.

[0110] In some embodiments, the calculation of the strain strength coefficient combines the magnitude and direction of the deviation, using the following formula:

[0111]

[0112] in: Represents the overall deformation strength coefficient. This represents the magnitude of the deviation. This represents the angle between the deviation direction and the theoretical observation vector. and These are pre-set weighting coefficients used to balance the influence of the magnitude and direction differences of the deviation on the overall deformation intensity coefficient. (Magnitude of deviation) The deviation direction angle is directly derived from the calculation of the magnitude of the difference between the motion vector and the theoretical observation vector. The weighting coefficients are obtained using the vector dot product formula. and The adjustment is made dynamically based on the movement speed of the hysteroscopic equipment; when the equipment moves at a high speed, the [adjustment] is increased. The weighting is increased when the equipment is moving slowly or stationary. The weight of the deformation pattern label. The assignment rule for deformation pattern labels is: when When the angle is less than 15 degrees, the label is "homogeneous reinforcement"; when... When the angle is greater than 165 degrees, the label is "reverse resistance"; otherwise, the label is "lateral slip".

[0113] Optionally, the dense optical flow algorithm can use a deep learning-based optical flow estimation network to replace the traditional algorithm. Motion pattern decomposition can also use independent component analysis. If the calculation of theoretical observation vectors encounters complex instrument kinematic models, a pre-calibrated lookup table can be used for mapping. Cluster analysis of local deformation units is performed directly in the three-dimensional uterine cavity model space, rather than in the two-dimensional image plane. Optionally, the data structure of the dynamic deformation mapping atlas is organized using a spatiotemporal voxel grid, with each voxel storing its spatial location, the time window of occurrence, the average deformation intensity coefficient, and the label of the most important deformation pattern. For time-series data, the temporal statistical characteristics of the deformation intensity coefficient at each voxel location, including the mean and variance, are additionally calculated and recorded in the dynamic deformation mapping atlas. It can be understood that the generation of the dynamic deformation mapping atlas depends on the precise spatiotemporal synchronization of the continuous original image stream and the physical motion trajectory. The time synchronization error needs to be controlled within one frame image period, and the spatial synchronization depends on the precise calibration of the motion sensor of the hysteroscopic equipment. The essence of dynamic deformation mapping is the quantitative expression of the dynamic mechanical response of endometrial tissue under external mechanical stimulation on an image sequence. By separating the movement of the device from the deformation of the tissue itself, it explicitly depicts the biomechanical characteristics of the tissue implied in the movement of image pixels.

[0114] In one embodiment of the present invention, when processing an example scenario involving endometrial adhesions, a dynamic deformation mapping atlas shows a continuous region on the posterior wall of the uterine cavity, whose deformation intensity coefficient fluctuates continuously within the range of 0.62 to 0.85, while the preset normal variation range threshold is 0.15 to 0.40. This region is initially screened as a candidate abnormal region because its deformation intensity coefficient continuously exceeds the threshold. Analyzing the distribution of deformation pattern labels within this candidate abnormal region, it is found that 70% of the local deformation units are labeled "reverse resistance", 25% are labeled "lateral slippage", and 5% are labeled "homogeneous enhancement". Since the proportion of dissimilar deformation pattern labels exceeds the preset 60% splitting threshold, the candidate abnormal region is split into two sub-regions, one sub-region dominated by the "reverse resistance" label and the other sub-region dominated by the "lateral slippage" label. For the sub-region dominated by "reverse resistance," the statistical distribution of the deformation intensity coefficients of all points within it was calculated, yielding a mean of 0.78, a variance of 0.12, and a gradient rate of change of 0.45 calculated using discrete gradients. In contrast, the statistical distribution of the normal uterine cavity region showed a mean of 0.25, a variance of 0.05, and a gradient rate of change of 0.08. Based on these statistical values, a multidimensional feature vector was constructed. For example, for this sub-region, the multidimensional feature vector is represented as follows: The vector elements represent the mean values ​​of the deformation intensity coefficients, respectively. ,variance and gradient rate of change Gradient rate of change The specific calculation uses the following formula:

[0115]

[0116] in: Represents the rate of change of the gradient. This represents the total number of spatially adjacent point pairs within a candidate anomaly region or sub-region. Representing the The difference in deformation intensity coefficient between adjacent points in the horizontal direction, Representing the For the difference in deformation intensity coefficient between adjacent points in the vertical direction, this formula quantifies the drastic change in the deformation intensity coefficient in space by averaging the absolute differences of all pairs of adjacent points. This multidimensional feature vector is input into a pre-trained tissue compliance classification model. The model outputs a probability of 0.05 for this sub-region belonging to a rigid region, 0.91 for an adhesion region, 0.03 for an edema region, and 0.01 for a normal region. Since the probability of an adhesion region is the highest, this sub-region is officially identified as an adhesion region within the abnormal tissue compliance regions.

[0117] In some embodiments, in a scenario of diffuse endometrial edema, the dynamic deformation mapping atlas shows that the deformation intensity coefficient of a large area at the fundus of the uterine cavity is in the range of 0.35 to 0.50, slightly exceeding the upper limit of the normal threshold, and the deformation pattern label distribution is highly consistent, all being "homogeneous enhancement". Therefore, this area is not segmented. The calculated statistical distribution has a mean of 0.42, a variance of 0.08, and a gradient change rate of 0.15. The constructed multidimensional feature vector is... The input to the tissue compliance classification model shows a probability of 0.88 for edema regions, thus identifying it as an edema region. The initial set of two-dimensional image coordinates corresponding to such abnormal tissue compliance regions in the continuous raw image stream is obtained; for example, the set of pixel coordinates in the image corresponding to adhesion regions. Using the depth map data collected simultaneously by the depth sensor of the hysteroscopy device, the coordinates of each two-dimensional pixel are... and its corresponding depth value Back-projecting onto the three-dimensional uterine cavity space model constructed from the physical motion trajectory yields a three-dimensional point set. In the three-dimensional uterine cavity spatial model, the concave hull algorithm is used to process the three-dimensional point set. The algorithm marks the boundary contour by finding a minimal closed surface that can enclose all three-dimensional points. For the adhesion region, the generated boundary contour may be an irregular polyhedral surface, while for the edema region, the generated boundary contour may be a relatively smooth surface.

[0118] In one embodiment of the present invention, see [reference] Figure 3The study addressed a region of abnormal tissue compliance identified as endometrial adhesion. Its three-dimensional boundary contour on the surface of a uterine cavity model appeared as an irregular elongated shape. The circularity value of this boundary contour was calculated to be 0.35, the aspect ratio to be 4.2, and the edge curvature to be 1.8. A weighted summation method was used to calculate the shape complexity score. The formula is:

[0119]

[0120] in: Represents the shape complexity score. The roundness representing the boundary contour. The aspect ratio representing the boundary profile. The curvature of the edge representing the boundary contour. , , These are preset weighting coefficients, which assign different weights to roundness, aspect ratio, and edge curvature to comprehensively evaluate the geometric complexity of the boundary contour. For this adhered region, the shape complexity score is calculated by substituting the numerical values. The score is 5.6, and the area of ​​the projected surface of this boundary contour in the 3D uterine cavity space model is calculated to be 22.5 square millimeters. Based on a pre-established correspondence table, the shape complexity score is... When the score is between 5.0 and 6.0 and the area is greater than 20 square millimeters, the corresponding basic field of view of the virtual scanning window is 4 mm × 3 mm. For comparison, a similarly sized window with a different shape complexity score... A circular edema area of ​​only 2.0 corresponds to a basic field of view size of 8 mm × 8 mm. Using the geometric center of the adhesion region's boundary contour as the center point of the viewing window, a pyramidal observation space is constructed in the three-dimensional uterine cavity model with this center point as its vertex. The rectangular base of this observation space has dimensions of 4 mm × 3 mm. The observation direction of the observation space is adjusted so that its axis is aligned with the optimal observation posture of the hysteroscopic device calculated based on the kinematic model at the center point of the viewing window; that is, the lens optical axis is as perpendicular as possible to the tangent plane of the endometrial surface at this point. Finally, this observation space is defined as a virtual scanning window.

[0121] In some embodiments, another area of ​​tissue compliance abnormality identified as an endometrial polyp is treated, based on its shape complexity score. The baseline field of view (FLAV) is determined to be 6 mm × 6 mm after consulting a corresponding relationship table, based on a 3.1 FLAV and an area of ​​8 square millimeters. A virtual scanning window is then generated accordingly. For each generated virtual scanning window, based on the type of the corresponding tissue compliance abnormality region, typical image texture and morphological feature libraries for that type of disease are retrieved from a pre-stored knowledge base. According to the spatial position parameters of the virtual scanning window, historical image segments that previously covered this field of view are extracted from the continuous raw image stream. Multi-scale Gabor filtering and SIFT feature point detection are performed on these historical image segments, and the number of typical features that have been revealed is counted. The feature manifestation ratio is calculated. For example, in a historical image segment corresponding to a virtual scanning window of an adhesion region, a "stripe-like texture" feature is detected, but a "sharp boundary" feature is not detected; its texture feature manifestation ratio is 50%, and its morphological feature manifestation ratio is 30%. Based on the feature manifestation ratio and the baseline field of view of the virtual scanning window, the number of unknown or ambiguous features that may be captured within the virtual scanning window is estimated as a potential feature increment. The estimation formula is: ,in This is an empirical coefficient for feature density per unit area. Simultaneously, the image information entropy of the historical image fragments themselves is calculated. Image information entropy With potential feature increment By weight and Weighted fusion is performed to obtain the expected value of feature abundance of the virtual scanning window. The weight Higher than weight For newly discovered virtual scan windows that are not covered by historical image fragments, the global average feature abundance of the corresponding tissue compliance anomaly region type is used as the initial estimate of the feature abundance expected value.

[0122] It is understandable that the generation strategy of virtual scanning windows is based on adaptive adjustments to the morphological features of areas with abnormal tissue compliance. Complex or large areas often require a smaller field of view to focus on details, while small, regularly shaped areas can be viewed from a larger field of view. The correspondence between the basic field of view size and the shape complexity score S and area size is pre-calibrated through analysis of a large amount of clinical data. The calculation of the expected value of feature abundance aims to quantify the potential information gain from enhanced scanning of each virtual scanning window. Image information entropy H reflects the information content of the existing image, and the potential feature increment... Based on the known feature display ratio and field of view, the potential for supplementary information is estimated. A high potential feature increment means that there are still a large number of unclear features within the window. A urgency level for acquisition is assigned to each virtual scanning window, and multiple diagnostic confidence threshold intervals are set. For example, interval one requires an image information entropy greater than 7.0, interval two requires an image information entropy greater than 5.0, and interval three requires an image information entropy greater than 3.0. The current image information entropy H of each virtual scanning window is compared with the interval threshold. If the image information entropy H is lower than the minimum feature information content of the interval, the virtual scanning window is determined to have insufficient information. For virtual scanning windows with insufficient information, their expected feature abundance is calculated. The difference between the current image information entropy H and the current image information entropy H According to the difference The size of the information potential level is divided, such as the difference. For high information potential windows, the difference The information potential window is between 1.5 and 3.0, with the difference being... For low-information-potential windows, the clinical risk weights of areas with abnormal tissue compliance are combined. For example, the clinical risk weight of adhesion areas is set to 1.5, that of polyp areas to 1.2, and that of edema areas to 1.0. The information potential level is quantified into a numerical value and multiplied by the type weight. Then, normalization is performed to obtain an urgency score. Based on the distribution of urgency scores, the virtual scanning window is divided into three acquisition urgency levels: high, medium, and low.

[0123] Optionally, the shape complexity score S can be calculated using Fourier descriptors as features. The correspondence table can be dynamically generated using a machine learning model. Latent feature increments. The estimation can incorporate prior knowledge of vascular density based on tissue type. The threshold for classifying information potential levels can be dynamically adjusted according to the overall progress of the current examination. The setting of type weights can incorporate more clinical parameters, such as patient age and medical history.

[0124] In one embodiment of the present invention, the hysteroscopy device controller receives and parses a virtual scanning window execution queue. The queue contains three sequentially arranged virtual scanning windows, identified as VW_01, VW_02, and VW_03. The controller first parses the spatial position and observation direction parameters of the first virtual scanning window, VW_01. The spatial position parameters are three-dimensional coordinates (X1, Y1, Z1), and the observation direction parameters are attitude quaternions (q1, q2, q3, q4). The hysteroscopy device is controlled to move from its current position along a transition trajectory specified in the queue to the observation position of the virtual scanning window VW_01. The transition trajectory is generated by interpolation of a series of path points. The attitude of the hysteroscopy device is adjusted by controlling the servo motor to align the lens optical axis with the observation direction of the virtual scanning window VW_01, with the attitude alignment error controlled within 0.5 degrees. Start the pause-focus scanning mode. In pause-focus scanning mode, the hysteroscopy device remains stationary. Adjust the lens focal length and illumination intensity to the optimal values ​​preset in the virtual scanning window VW_01. Refer to Table 1 for the preset parameter configurations corresponding to different acquisition urgency levels.

[0125] Table 1: Preset Scanning Parameters for Different Acquisition Urgency Levels

[0126]

[0127] The virtual scanning window VW_01 has a high acquisition urgency level, therefore a multi-focus depth-of-field fusion strategy is adopted. The illumination intensity is set to 10000 lux, and multiple local images within the window area are acquired at a frame rate of 15 frames per second and a resolution of 3840x2160. Real-time sharpness evaluation is performed on the acquired local images, using a sharpness evaluation function. Defined as:

[0128]

[0129] in: Represents the clarity score. Represents the total number of pixels in the image. Representative image in the The gradient vector at each pixel Represents the square of the gradient magnitude. Represents the image The noise level estimate, These are weighting coefficients used to balance the gradient and noise terms. This function quantifies image sharpness by comprehensively evaluating the image gradient and noise level. Simultaneously, motion artifact detection is performed, achieved by comparing the feature matching errors of corresponding blocks between consecutively acquired frames, and a sharpness score is selected. The image with the highest resolution and no motion artifacts is used as the high-resolution local enhancement image of the virtual scanning window VW_01. The high-resolution local enhancement image and its corresponding virtual scanning window identifier VW_01 are stored. The hysteroscopic device is controlled to move to the next virtual scanning window VW_02 in queue order, and the pause-focusing scanning process, from adjusting the hysteroscopic device attitude to aligning the observation direction to selecting and storing the high-resolution local enhancement image, is repeated until all virtual scanning windows in the queue have been scanned, ultimately resulting in a set of high-resolution local enhancement images corresponding one-to-one with each virtual scanning window.

[0130] In some embodiments, a virtual scanning window with a urgency level of 0 is processed, corresponding to a polyp tissue type, with a preset illumination intensity of 8000 lux, single-shot autofocus, a frame rate of 10 fps, and a resolution of 1920x1080. The weighting coefficients in the sharpness evaluation function... The system adaptively adjusts based on lighting intensity; higher lighting intensity may result in more significant noise estimation, with weighting coefficients adjusted accordingly. The corresponding increase is made. In motion artifact detection, image frames with feature matching errors exceeding a set threshold are excluded. Scale-invariant feature transform (SMT) points and their descriptors are extracted for each obtained high-resolution local enhancement image. For example, 1500 feature points and their 128-dimensional descriptor are extracted from the high-resolution local enhancement image identified as VW_01. In the continuous raw image stream, the reference image frame that is closest to each high-resolution local enhancement image in time and space is located based on the timestamp and spatial coordinates. For example, the raw image frame whose timestamp is adjacent to the high-resolution local enhancement image corresponding to VW_01 is used as the reference image frame. Scale-invariant SMT points and descriptors of the same type are extracted from the reference image frame, resulting in 1200 feature points. Through a feature descriptor matching algorithm, specifically a matcher based on fast nearest neighbor search, a correspondence between feature points in the high-resolution local enhancement image and feature points in the reference image frame is established, resulting in 256 matching pairs.

[0131] See Figure 4The study presents a comparison of the number of feature points in the enhanced images, the number of feature points in the original reference frames, and the number of feature matching points for three virtual scanning windows (VW_01, VW_02, and VW_03). The data shows that the number of feature points in the enhanced images is consistently higher than that in the original reference frames. For example, the enhanced image of VW_01 extracts 1500 feature points, while the original reference frame extracts only 1200. This indicates that high-resolution local enhancement images effectively improve the feature information density of the image through strategies such as multi-focus depth-of-field fusion. In the feature matching stage, the number of matching points for VW_01, VW_02, and VW_03 are 256, 189, and 127, respectively, showing a decreasing trend with the window sequence. This phenomenon is related to the acquisition urgency level of the virtual scanning windows and the feature abundance of regions with abnormal tissue compliance: high-urgency windows (such as VW_01) use higher scanning parameter configurations, resulting in a relatively higher feature matching degree between their enhanced images and the original reference frames. However, subsequent windows experience a decrease in the number of matching points due to the reduced expected value of feature abundance. The comparison results verified the effectiveness of the virtual scanning window execution queue and provided a quantitative basis for the feature alignment and stitching of subsequent enhanced fusion diagnostic images.

[0132] In one embodiment of the present invention, an enhanced fusion diagnostic image covering the entire uterine cavity is loaded. This image clearly shows an adhesion region on the posterior wall of the uterine cavity and a polyp region at the bottom of the uterine cavity. Simultaneously, a three-dimensional uterine cavity spatial model associated with this enhanced fusion diagnostic image and the boundary contour data of these two tissue compliance abnormalities are loaded. On the enhanced fusion diagnostic image, based on the boundary contour data, an image segmentation algorithm is used to segment the image sub-regions of the adhesion region and the polyp region. The image sub-region of the adhesion region contains approximately 15,000 pixels, and the image sub-region of the polyp region contains approximately 5,000 pixels. For each image sub-region, dense feature extraction is performed. The color histogram statistical features extracted from the image sub-region of the adhesion region are a 64-dimensional vector, the multi-directional Gabor filter response features are a 48-dimensional vector, the local binary pattern histogram features are a 256-dimensional vector, and the deep convolutional neural network activation features are a 1024-dimensional vector. All features are concatenated to form a 1392-dimensional dense feature vector. A similar 1392-dimensional dense feature vector is extracted from the image sub-region of the polyp region. The extracted dense feature vectors are input into a multi-task analysis network, which consists of a shared feature extractor and two parallel branches. The shared feature extractor is a fully connected layer, and the two parallel branches are a fully connected layer and a softmax layer for disease type classification, and a fully connected layer and a linear output layer for disease severity regression. The multi-task analysis network processes the input dense feature vector of the adhesion region, outputting a disease type probability distribution vector of [adhesion: 0.92, polyp: 0.05, edema: 0.02, normal: 0.01], with a continuous numerical score of severity of 0.85. It processes the dense feature vector of the polyp region, outputting a disease type probability distribution vector of [polyp: 0.88, adhesion: 0.08, edema: 0.03, normal: 0.01], with a continuous numerical score of severity of 0.60. Based on the disease type probability distribution vectors, the type with the highest probability is taken as the preliminary diagnostic type for the tissue compliance abnormality region. Therefore, the preliminary diagnostic type for the adhesion region is adhesion, and the preliminary diagnostic type for the polyp region is polyp.

[0133] In some embodiments, when processing a large edema region image sub-region, the probability distribution vector output by the multi-task analysis network is [edema: 0.90, normal: 0.08, adhesion: 0.01, polyp: 0.01], and the continuous numerical score for severity is 0.45. Combining the continuous numerical score for severity with the area size and shape complexity of the tissue compliance abnormality region, a pre-calibrated quantification formula is used to calculate the contribution weight of the tissue compliance abnormality region in the overall intrauterine health assessment. The formula is:

[0134]

[0135] in: The contribution weight representing a region of abnormal compliance in a particular organization. A continuous numerical score representing the severity of the area. This represents the size of the area. This represents the shape complexity score of the region. The function represents the natural logarithm, and the denominator is the sum of all identified tissue compliance abnormalities. The sum of values ​​ensures that the sum of all weights equals 1. For areas of adhesion, a continuous numerical score is used to determine the severity. Area size Shape complexity score For polyp areas, a continuous numerical score is used to assess severity. Area size Shape complexity score Calculate the contribution weight of the adhesion region. The contribution weight of the polyp region is approximately 0.72. It is approximately 0.28.

[0136] See Figure 5 In the urgency allocation stage of virtual scanning window acquisition, the correlation analysis between image information entropy and expected value of feature abundance is the core basis. Specifically, the distribution patterns of image information entropy and expected feature abundance values ​​of virtual scanning windows corresponding to different tissue compliance abnormalities were presented: high-urgency windows (such as windows 1 and 2, corresponding to adhesion areas) exhibited higher image information entropy (approximately 4.25~4.50) and expected feature abundance values ​​(approximately 0.93~0.95), because adhesion areas have high texture complexity and large potential feature increments, resulting in significant diagnostic information gain through enhanced scanning; medium-urgency windows (such as windows 3, 4, and 5, corresponding to polyp and edema areas) had image information entropy (approximately 3.00~3.80) and expected feature abundance values ​​(approximately 0.75~0.90) in the middle range, reflecting that their feature manifestation ratio and information potential were weaker than those of adhesion areas; low-urgency windows (such as window 6, corresponding to edema areas) had lower image information entropy (approximately 2.75) and expected feature abundance values ​​(approximately 0.70), indicating that the diagnostic information in this area was relatively sufficient, and enhanced scanning had the lowest priority. The positive correlation trend line in the figure further verifies the association between the two: the higher the image information entropy, the higher the expected value of feature abundance, and the higher the corresponding level of data collection urgency.

[0137] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for image recognition enhancement in hysteroscopic diagnosis of common intrauterine diseases, characterized in that, include: By simultaneously acquiring continuous raw image streams and physical motion trajectories using hysteroscopic equipment, and performing spatiotemporal continuity analysis, dynamic deformation mapping maps are generated. Based on the dynamic deformation mapping atlas, regions with abnormal tissue compliance are identified, and the boundary contours of these regions are marked in the three-dimensional uterine cavity space. Based on the shape complexity and area size of the boundary contour, multiple virtual scanning windows with different field of view are dynamically generated; Based on the type of the abnormal tissue compliance region and the spatial location of the virtual scanning window, calculate the image information entropy and expected value of feature abundance for each virtual scanning window; The image information entropy and the expected value of feature abundance are combined with a preset diagnostic confidence threshold to assign a collection urgency level to each virtual scanning window; Based on the urgency level of the data acquisition and the physical movement constraints of the hysteroscopic equipment, an ordered virtual scanning window execution queue is constructed. The hysteroscopic device is driven to perform pause-focus scanning on the target uterine cavity region sequentially according to the virtual scanning window execution queue, thereby acquiring a set of high-resolution locally enhanced images; The high-resolution locally enhanced image is feature-aligned and stitched with the continuous original image stream to synthesize an enhanced fusion diagnostic image covering the entire uterine cavity; The step of calculating the image information entropy and expected feature abundance value corresponding to each virtual scanning window based on the type of the abnormal tissue compliance region and the spatial position of the virtual scanning window is as follows: Based on the type of abnormal tissue compliance region, the image texture feature library and morphological feature library of typical diseases are called from the pre-stored knowledge base; For each virtual scanning window, historical image segments located within the window area are extracted from the continuous raw image stream based on its spatial location parameters; Multi-scale filtering and feature point detection are performed on the historical image segments. The types of textures and the number of morphological features that have been revealed are counted. The results are compared with the typical image texture feature library and morphological feature library to calculate the feature manifestation ratio. Based on the feature display ratio and the basic field of view size of the virtual scanning window, the number of unknown or fuzzy features captured within the window range is estimated as a potential feature increment. Calculate the image information entropy by combining the grayscale distribution of the historical image fragments themselves; The image information entropy and the latent feature increment are weighted and fused to obtain the expected value of the feature abundance of the virtual scanning window, wherein the weight of the latent feature increment is higher than the weight of the historical image information entropy. For newly discovered virtual scan windows that are not covered by historical image fragments, the average feature abundance of the corresponding tissue compliance anomaly region type is used as the initial estimate of its feature abundance expected value.

2. The method for enhancing hysteroscopic images for common intrauterine diseases according to claim 1, characterized in that, The process of simultaneously acquiring continuous raw image streams and physical motion trajectories using a hysteroscopic device, performing spatiotemporal continuity analysis, and generating a dynamic deformation mapping map includes: The hysteroscopic device acquires a continuous stream of raw images of the target uterine cavity while in continuous motion, and simultaneously records the physical motion trajectory of the hysteroscopic device. Spatiotemporal continuity analysis is performed on the continuous raw image stream to extract the parallax changes and tissue texture drift features between image frames, and the physical motion trajectory is fused to generate a dynamic deformation mapping map of the endometrial surface. The process involves performing spatiotemporal continuity analysis on the continuous raw image stream, extracting disparity changes and tissue texture drift features between image frames, fusing the physical motion trajectory, and generating a dynamic deformation mapping map of the endometrial surface. Specifically: The continuous raw image stream is decomposed into continuous image frame pairs. Dense optical flow algorithm is applied to each image frame pair to calculate the motion vector of each pixel between consecutive frames, forming dense optical flow field data. Motion pattern decomposition was performed on the dense optical flow field data to separate the background flow field caused by the global motion of the hysteroscopic equipment and the characteristic flow field caused by the local deformation of the endometrium. Cluster analysis is performed on the characteristic flow field to identify flow field regions with consistent motion directions, and each region is defined as a local deformation unit. Synchronously read the physical motion trajectory data corresponding to the timestamp of each image frame, the physical motion trajectory data including the three-dimensional spatial coordinates and attitude angle of the end of the hysteroscopic device; Using inverse kinematics algorithm, the physical motion trajectory data is converted into a theoretical observation vector of the endometrial surface by the lens of the hysteroscopic device; The motion vector of the local deformation unit is compared with the theoretical observation vector, and the deviation between the actual motion and the theoretical observation of each local deformation unit is calculated. Based on the magnitude and direction of the deviation, each local deformation unit is assigned a deformation intensity coefficient and a deformation mode label; The deformation intensity coefficients and deformation mode labels of all local deformation units are mapped back to the initial spatial position on the surface of the endometrium, constructing a dynamic deformation mapping map containing spatiotemporal dimensions. The dynamic deformation mapping map records the dynamic response characteristics of each point on the surface of the endometrium during the movement of the device.

3. The method for enhancing hysteroscopic images for common intrauterine diseases according to claim 1, characterized in that, The step of identifying tissue compliance abnormalities based on the dynamic deformation mapping map and marking the boundary contours of the tissue compliance abnormalities in the three-dimensional uterine cavity space specifically involves: In the dynamic deformation mapping map, a threshold for the normal variation range of the deformation intensity coefficient is set, and continuous regions where the deformation intensity coefficient continuously exceeds the threshold for the normal variation range are initially screened as candidate abnormal regions. Analyze the distribution consistency of deformation pattern labels within each candidate anomaly region. If a candidate anomaly region contains more than a preset proportion of distinct deformation pattern labels, then the candidate anomaly region is divided into multiple sub-regions. For each candidate anomaly region or its sub-region, calculate the statistical distribution of the deformation intensity coefficients of all points within it, including the mean, variance, and gradient rate of change. Based on the statistical distribution, a multidimensional feature vector is constructed, which is used to characterize the deformation inhomogeneity and dynamic stability of the candidate anomaly region. The multidimensional feature vector is input into a pre-trained tissue compliance classification model, which outputs the probability that each region belongs to a rigid region, an adhesion region, an edematous region, or a normal region. Areas classified as rigid, adhesion, or edema areas are formally defined as areas of abnormal tissue compliance. Obtain the initial set of two-dimensional image coordinates corresponding to the abnormal tissue compliance region in the continuous raw image stream; Using the depth sensor data from the hysteroscopic device, the initial two-dimensional image coordinate set is back-projected onto a three-dimensional uterine cavity space model constructed from physical motion trajectories; In the three-dimensional uterine cavity space model, the concave hull algorithm is used to process the back-projected three-dimensional point set to generate the minimum closed surface that can enclose all three-dimensional points. The minimum closed surface is the boundary contour of the tissue compliance abnormal region in the three-dimensional uterine cavity space.

4. The method for enhancing hysteroscopic images for common intrauterine diseases according to claim 1, characterized in that, Based on the shape complexity and area size of the boundary contour, multiple virtual scanning windows with different field of view are dynamically generated, specifically as follows: Geometric features are extracted from the boundary contour of each region with abnormal tissue compliance. The circularity, aspect ratio, and edge curvature of the boundary contour are calculated, and the shape complexity score of the boundary contour is obtained by combining them. Calculate the area of ​​the projected surface of the boundary contour in the three-dimensional uterine cavity space model, and use it as the area size of the abnormal tissue compliance region; Establish a table showing the correspondence between shape complexity score, area size, and basic field of view size of the virtual scanning window. The higher the shape complexity or the larger the area, the smaller the corresponding basic field of view. Based on the corresponding table, a basic field of view size for a virtual scanning window is determined for each region of tissue compliance abnormality. The geometric center of the boundary contour of each region with abnormal tissue compliance is used as the center point of the viewport; Based on the basic field of view size, in the three-dimensional uterine cavity space model, a square pyramid-shaped observation space is constructed with the center point of the window as the vertex. The size of the rectangular base of the observation space is determined by the basic field of view size. Adjust the observation direction of the observation space so that its axis is aligned with the optimal observation posture that the hysteroscopic device can reach at the center point of the window. The observation space volume is defined as a virtual scanning window covering the region of abnormal tissue compliance. The virtual scanning window includes parameters such as spatial location, observation direction, and field of view.

5. The method for enhancing hysteroscopic diagnostic images of common intrauterine diseases according to claim 1, characterized in that, The step of combining the image information entropy with the expected value of feature abundance and a preset diagnostic confidence threshold to assign a collection urgency level to each virtual scanning window is as follows: Multiple diagnostic confidence threshold intervals are set, and each threshold interval corresponds to the minimum amount of feature information required for a diagnosis. The current image information entropy of each virtual scanning window is compared with the minimum feature information amount. If the image information entropy is lower than the minimum feature information amount, it is determined that the current information of the virtual scanning window is insufficient. For virtual scanning windows with insufficient current information, the difference between their expected feature abundance and the current image information entropy is further calculated. This difference reflects the potential for additional information to be obtained through enhanced scanning. Based on the magnitude of the difference, the virtual scanning window is divided into a high information potential window, a medium information potential window, and a low information potential window; The type weights of the abnormal tissue compliance regions are combined, and these type weights are pre-set according to the clinical risk level of the disease. The urgency score is obtained by multiplying the information potential level by the type weight and then normalizing the result. Based on the distribution of urgency scores, the virtual scanning window is divided into three acquisition urgency levels: high, medium, and low. The high acquisition urgency level indicates the window that needs to be scanned with priority and high quality.

6. The method for enhancing hysteroscopic images for common intrauterine diseases according to claim 1, characterized in that, The step of constructing an ordered virtual scanning window execution queue based on the urgency level of data acquisition and the physical movement constraints of the hysteroscopic equipment is as follows: Obtain the physical motion constraints of the hysteroscopic device, including the maximum turning angle, minimum travel distance, device length limit, and safe obstacle avoidance space; Starting from the current position of the hysteroscopic device, all virtual scanning windows are considered as nodes to be accessed. A heuristic graph search algorithm is used for path planning, with the center point of each virtual scanning window as a node in the graph. The movement cost between nodes is determined by the actual movement distance of the device and the complexity of attitude adjustment. The collection urgency level is used as the priority weight for node access, and nodes with higher collection urgency levels receive higher access priority in path cost calculation. The heuristic graph search algorithm, under the premise of satisfying physical motion constraints, finds an approximately optimal path that starts from the starting point, visits all nodes in sequence, and finally covers all nodes with high acquisition urgency level. Based on the access order of nodes in the near-optimal path, a virtual scanning window execution queue is generated. The queue contains not only the execution order of the windows, but also a description of the transition trajectory of the hysteroscopic device moving from one window to the next.

7. The method for enhancing hysteroscopic images for common intrauterine diseases according to claim 1, characterized in that, The hysteroscopic device sequentially performs pause-focus scanning on the target uterine cavity region according to the virtual scanning window execution queue, acquiring a set of high-resolution locally enhanced images, specifically: The hysteroscopy device controller receives the virtual scanning window execution queue and parses the spatial position and observation direction parameters of the first virtual scanning window in the queue; Control the hysteroscopic equipment to move from its current position along the transition trajectory specified in the queue to the observation position of the first virtual scanning window; Adjust the orientation of the hysteroscopy equipment so that its lens optical axis is aligned with the observation direction of the virtual scanning window; In the pause-focus scanning mode, the hysteroscopic device remains stationary, and the lens focal length and illumination intensity are adjusted to the optimal values ​​preset for the virtual scanning window. These optimal values ​​are pre-configured according to the urgency level of the acquisition and the tissue type. While the device is stationary, multiple local images within the window area are acquired at a frame rate and resolution higher than that of the continuous raw image stream. Real-time sharpness evaluation and motion artifact detection are performed on multiple acquired local images, and the image with the highest sharpness and no artifacts is selected as the high-resolution local enhancement image of the virtual scanning window. Store the high-resolution locally enhanced image and its corresponding virtual scanning window identifier; The hysteroscopic device is controlled to move to the next virtual scanning window in the queue order, and the pause-focusing scanning process, from adjusting the posture of the hysteroscopic device to aligning the observation direction to selecting and storing high-resolution local enhancement images, is repeated until all virtual scanning windows in the queue have been scanned, and finally a set of high-resolution local enhancement images corresponding one-to-one with the virtual scanning windows are obtained.

8. The method for enhancing hysteroscopic diagnostic images of common intrauterine diseases according to claim 1, characterized in that, The step of aligning and stitching the high-resolution locally enhanced image with the continuous original image stream to synthesize an enhanced fusion diagnostic image covering the entire uterine cavity specifically involves: Extract scale-invariant feature transform feature points and their descriptors for each high-resolution locally enhanced image; In the continuous raw image stream, locate the reference image frame that is closest to each high-resolution locally enhanced image in time and space; Extract feature points and descriptors of the same type from the reference image frame; A correspondence between feature points in a high-resolution locally enhanced image and feature points in a reference image frame is established using a feature descriptor matching algorithm. Based on the correspondence pairs, the homography matrix transformation model is used to align each high-resolution local enhancement image to the global image coordinate system formed by the continuous original image stream; A multi-band fusion algorithm is used to seamlessly stitch the aligned high-resolution local enhanced image with the corresponding region in the global image coordinate system, achieving a smooth transition of texture and color at the stitching boundary. For the uterine cavity region not covered by the high-resolution local enhancement image, the image content of the corresponding region in the continuous original image stream is preserved; All image blocks after stitching and fusion are integrated onto a unified image canvas to generate an enhanced fusion diagnostic image that is spatially continuous and has high-resolution details in key areas. The enhanced fusion diagnostic image preserves the complete uterine cavity topology.

9. The method for enhancing hysteroscopic images for common intrauterine diseases according to claim 1, characterized in that, The method further includes a step of performing quantitative analysis of intrauterine diseases based on the enhanced fusion diagnostic image, specifically: Load the enhanced fusion diagnostic image covering the entire uterine cavity, and load the associated three-dimensional uterine cavity spatial model and the boundary contour data of the tissue compliance abnormality area; On the enhanced fusion diagnostic image, image sub-regions of each tissue compliance abnormality region are segmented based on boundary contour data; For each image sub-region, intensive feature extraction is performed. The extracted features include, but are not limited to, color histogram statistics, multi-directional Gabor filter response, local binary pattern histogram, and deep convolutional neural network activation features. The extracted dense features are input into a multi-task analysis network, which performs two tasks in parallel: disease type classification and disease severity regression. The multi-task analysis network outputs a disease type probability distribution vector for each region of tissue compliance abnormality, as well as a continuous numerical score representing the severity. Based on the disease type probability distribution vector, the type with the highest probability is taken as the preliminary diagnostic type of the tissue compliance abnormality region; By combining the continuous numerical score of the severity with the area size and shape complexity of the tissue compliance abnormality region, the contribution weight of the tissue compliance abnormality region in the overall intrauterine health assessment is calculated using a pre-calibrated quantitative formula. By summarizing the diagnostic types, severity continuous numerical scores, and contribution weights of all areas with abnormal tissue compliance, a structured quantitative analysis report of intrauterine diseases is generated. This report describes the distribution, type, and severity of intrauterine diseases in data form.

Images