Uterine muscular layer injury rating method based on double-branch feature fusion and edge calibration

Abstract

The invention discloses a uterine muscular layer injury rating method based on double-branch feature fusion and edge calibration, which solves the core problems of difficulty in uterine muscular layer injury detection and subjective rating in a hysteroscope scene through technical innovation of image preprocessing, double-branch segmentation, edge calibration, multi-dimensional feature fusion and hierarchical modeling. The method is adaptive to the special environment of the hysteroscope image and the clinical features of the injured area, real-time processing in the operation is achieved while the detection accuracy is guaranteed, and objective and reliable technical support is provided for early warning of the water poisoning risk in the operation.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer vision, medical image analysis and minimally invasive gynecological diagnosis and treatment. Specifically, it relates to an automatic detection and grading assessment method for uterine myometrial damage in hysteroscopic surgery, which can be integrated into hysteroscopic surgical equipment to provide technical support for early warning of intraoperative water intoxication risk. Background Technology

[0002] Hysteroscopic surgery is one of the core technologies in minimally invasive gynecological diagnosis and treatment. However, intraoperative damage to the myometrium can lead to serious complications such as water intoxication, threatening patient safety. Currently, the assessment of myometrial damage mainly relies on the surgeon's visual observation and clinical experience, which has significant limitations: First, it is highly subjective, with poor consistency in diagnosis among physicians of different experience, easily leading to missed diagnoses or misjudgments; second, hysteroscopic images are affected by uneven lighting during surgery, interference from tissue secretions, and instrument obstruction, resulting in blurred edges and indistinct features of the damaged area, making it difficult to accurately identify with the naked eye; third, there is a lack of quantitative damage rating standards, which cannot provide an objective basis for intraoperative risk assessment.

[0003] While existing image recognition technologies have made progress in the medical field, they still have shortcomings in the detection and rating of uterine myometrial injuries under hysteroscopy: general medical image segmentation models are not adapted to the special noise and tissue texture characteristics of hysteroscopic images, resulting in insufficient segmentation accuracy; lesion detection methods are mostly aimed at common tumors or polyps, and have poor adaptability to lesions with diverse shapes and blurred boundaries, such as myometrial injuries; existing rating algorithms rely on large-scale labeled data, while uterine myometrial injury samples are scarce, resulting in limited model generalization ability. Therefore, developing a method for detecting and rating uterine myometrial injuries that is adapted to hysteroscopy scenarios and balances accuracy and practicality has become crucial for improving the safety of hysteroscopic surgery. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies and provide a method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration. Environmental interference is eliminated through adaptive denoising and illumination equalization; a bi-branch network is used to extract global and local features of the myometrium for accurate segmentation; an edge calibration algorithm is combined to refine the boundaries of the damaged area; a rating model is constructed by fusing multi-dimensional features such as texture and morphology; and finally, an objective damage level is output, providing reliable support for intraoperative water intoxication risk warning.

[0005] The technical solution adopted in this invention is: a method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration, comprising the following steps: Step 1: Hysteroscopic image acquisition: Real-time images of the uterine cavity during the procedure are acquired using the built-in camera of the hysteroscopic device, and imaging parameters are recorded simultaneously; Step 2: Image preprocessing: Adaptive denoising and illumination equalization are performed on the acquired images to eliminate environmental interference factors; Step 3: Two-branch segmentation of the uterine myometrium: Using a Transformer and CNN two-branch network, global and local features of the myometrium are extracted to generate a myometrium region mask; Step 4: Damage area detection and edge calibration: Based on the segmented muscle layer region, potential damage sites are identified, and the damage boundary is optimized through an edge calibration algorithm; Step 5: Multi-dimensional feature fusion: Extract texture, shape, and grayscale features of the damaged area, and enhance the hierarchical recognition through weighted fusion; Step 6: Damage severity rating: Input the fused features into the grading model and output the damage level: mild, moderate, severe; Step 7: Result Output: Simultaneously display the location of the damaged area, edge contour, and damage level, providing the surgeon with an intuitive reference.

[0006] Furthermore, in step 1, a high-definition RGB camera built into the hysteroscope is used to acquire real-time images of the uterine cavity during the procedure. The image resolution is adapted to the device's hardware specifications. During acquisition, it is ensured that there are no serious instruments obstructing the myometrium, and the image includes a complete view of the uterine cavity wall. The acquired raw image is denoted as [image name missing]. ,in The coordinates are pixel coordinates, and the image channels are R, G, and B channels.

[0007] Furthermore, step 2, through adaptive denoising and illumination equalization processing, provides high-quality input for subsequent segmentation and detection, including the following steps: (1) Adaptive noise suppression A context-aware weighted filtering algorithm is used to suppress noise while preserving tissue edge details. The formula is as follows: Symbol explanation: These are the pixel values ​​of the denoised image; The width of the filter window is half its width. For adaptive weighting coefficients, derived from pixels With neighboring pixels ( The RGB grayscale difference and gradient information are calculated together. Regions with small grayscale differences and gentle gradients have higher weights to avoid edge blurring. (2) Light equalization The partitioned adaptive gamma correction algorithm is used to correct uneven illumination in images. The formula is as follows: Symbol explanation: Represents the R, G, and B channels; The c-channel pixel value after illumination correction; These are pixel-level adaptive gamma coefficients, calculated from the mean and variance of the grayscale value of the local region where the pixel is located, for dark areas. To enhance brightness, bright areas To reduce brightness and achieve global illumination balance.

[0008] Furthermore, step 3 employs a dual-branch network structure of Transformer and CNN to extract and fuse global and local features respectively, achieving accurate segmentation of the muscle layer region. This includes two steps: feature extraction and fusion. (1) Dual-branch feature extraction Transformer branch: For Perform block encoding and extract global semantic features, using the following formula: Symbol explanation: For image segmentation, the image is divided into feature blocks of fixed size; It is a Transformer encoder that captures global feature associations through a self-attention mechanism; For global feature maps; CNN branch: Employs convolutional neural networks to extract local texture and edge features, using the following formula: Symbol explanation: The convolutional module extracts local detail features through multi-scale convolutional kernels; This is a local feature map; (2) Feature fusion and segmentation Muscle layer segmentation mask is generated by fusing bi-branch features using attention weighting, as shown in the formula: Symbol explanation: This is a channel attention mechanism used to enhance effective information in local features; To integrate weights, the contributions of global and local features are adaptively balanced; To fuse feature maps; The activation function maps features to probability values ​​in the interval [0,1]. This is a muscle layer segmentation mask. Pixels with a probability higher than a set threshold are identified as muscle layer regions, otherwise they are considered background.

[0009] Furthermore, step 4, which locates the damaged area based on the muscle layer segmentation mask and optimizes the boundary accuracy through edge calibration, includes two steps: (1) Initial detection of the damaged area The formula for screening potential injury sites based on abnormal features in the muscle layer region is as follows: Symbol explanation: To perform pixel-by-pixel multiplication, extract the muscle layer region image. ; , These represent the global grayscale mean and standard deviation of the muscle layer region, respectively. This is an abnormal feature map. The larger the value, the higher the probability that the pixel is a damaged area. Areas with values ​​higher than the set threshold are potential damaged areas. (2) Edge calibration The edge of the damaged region is refined using a multi-scale gradient weighted algorithm, as shown in the formula: Symbol explanation: Three different scale gradient detection kernels are used. Each edge is captured with a different thickness; This is a gradient calculation operation; This is the gradient map for the k-th scale; The gradient weights are adaptively assigned based on edge sharpness. The gradient map after calibration; Thresholding is applied to generate a binarized damage region mask. .

[0010] Furthermore, in step 5, three core features—texture, morphology, and grayscale—are extracted from the damaged area. These features are then weighted and fused to enhance the hierarchical recognition. The formula is as follows: Symbol explanation: Image of the damaged area; Use the Gabor filter operator to extract texture features. ; Morphological feature extraction operators are used to extract features such as area and roundness of the damaged region. ; Extract grayscale features for mean calculation. ; , , For feature fusion weights, satisfying This is used to balance the hierarchical contributions of the three types of features; This is the final fused feature vector.

[0011] Furthermore, in step 6, the fused features are input into the hierarchical model, and the damage level is output using the following formula: Symbol explanation: The hierarchical model is implemented using a lightweight fully connected network. Output the damage level, with values ​​of {0, 1, 2}, corresponding to mild damage, moderate damage, and severe damage, respectively. Mild damage represents slight damage to the surface of the muscle layer, moderate damage represents partial damage to the thickness of the muscle layer, and severe damage represents deep damage or discontinuity of the muscle layer.

[0012] Furthermore, in step 7, the damaged area is masked. Calibrated edge contour and damage level The image is overlaid on the original hysteroscopic image and simultaneously output to the display interface of the surgical device, providing the surgeon with a real-time and intuitive reference for damage assessment.

[0013] The innovation of this invention lies in the fact that, based on the characteristics of hysteroscopic images and the clinical features of uterine myometrial damage, it integrates multi-scale feature extraction and edge optimization technology. While adapting to small-sample medical data scenarios, it ensures detection accuracy and meets the needs of real-time processing during surgery, thereby achieving objective and efficient damage assessment.

[0014] This invention proposes a uterine myometrial injury rating method based on bi-branch feature fusion and edge calibration. Through technical innovations in image preprocessing, bi-branch segmentation, edge calibration, multi-dimensional feature fusion, and hierarchical modeling, it solves the core problems of difficult detection and subjective rating of uterine myometrial injuries in hysteroscopic scenarios. This method is adapted to the special environment of hysteroscopic images and the clinical characteristics of the injury area. While ensuring detection accuracy, it enables real-time intraoperative processing, providing objective and reliable technical support for intraoperative water intoxication risk warning.

[0015] The beneficial effects of this invention are: High adaptability: The preprocessing module specifically addresses issues such as uneven illumination and noise interference in hysteroscopic images, improving robustness in complex scenarios.

[0016] Accurate detection: Dual-branch segmentation and edge calibration work together to effectively identify damaged areas with blurred boundaries, reducing missed detections and false positives.

[0017] Objective rating: The grading system, constructed by integrating multi-dimensional features, avoids the subjectivity of manual evaluation and provides quantitative standards.

[0018] Real-time and efficient: Lightweight network architecture and integrated processing links adapt to the needs of real-time analysis during surgery without affecting the surgical procedure.

[0019] Highly practical: No large-scale labeled data is required; it can be quickly deployed on existing hysteroscopic equipment without additional hardware modifications. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall process of the hysteroscopic method for detecting and evaluating uterine myometrial damage according to the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: like Figure 1 As shown, a method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration includes the following steps: Step 1: Hysteroscopic image acquisition A high-definition RGB camera built into the hysteroscope was used to acquire real-time images of the uterine cavity during the procedure. The image resolution was adapted to the device's hardware specifications. During acquisition, it was ensured that there were no significant instrument obstructions in the myometrial area, and the images included a complete view of the uterine cavity wall. The acquired raw images were recorded as follows: ,in The coordinates are pixel coordinates, and the image channels are R, G, and B channels.

[0022] Step 2: Image Preprocessing Through adaptive denoising and illumination equalization, high-quality input is provided for subsequent segmentation and detection. The core process includes two steps: (1) Adaptive noise suppression A context-aware weighted filtering algorithm is used to suppress noise while preserving tissue edge details. The formula is as follows: Symbol explanation: These are the pixel values ​​of the denoised image; The width of the filter window is half its width. For adaptive weighting coefficients, derived from pixels With neighboring pixels ( The RGB grayscale difference and gradient information are calculated together. Regions with small grayscale differences and gentle gradients have higher weights to avoid edge blurring.

[0023] (2) Light equalization The partitioned adaptive gamma correction algorithm is used to correct uneven illumination in images. The formula is as follows: Symbol explanation: Represents the R, G, and B channels; The c-channel pixel value after illumination correction; These are pixel-level adaptive gamma coefficients, calculated from the mean and variance of the grayscale value of the local region where the pixel is located, for dark areas. To enhance brightness, bright areas To reduce brightness and achieve global illumination balance.

[0024] Step 3: Bibranching of the uterine myometrium A dual-branch network structure combining Transformer and CNN is employed to extract and fuse global and local features respectively, achieving accurate segmentation of the muscle layer region. The core process includes two steps: feature extraction and fusion. (1) Dual-branch feature extraction Transformer branch: For Perform block encoding and extract global semantic features, using the following formula: Symbol explanation: For image segmentation, the image is divided into feature blocks of fixed size; It is a Transformer encoder that captures global feature associations through a self-attention mechanism; This is the global feature map.

[0025] CNN branch: Employs convolutional neural networks to extract local texture and edge features, using the following formula: Symbol explanation: The convolutional module extracts local detail features through multi-scale convolutional kernels; This is a local feature map.

[0026] (2) Feature fusion and segmentation Muscle layer segmentation mask is generated by fusing bi-branch features using attention weighting, as shown in the formula: Symbol explanation: This is a channel attention mechanism used to enhance effective information in local features; To integrate weights, the contributions of global and local features are adaptively balanced; To fuse feature maps; The activation function maps features to probability values ​​in the interval [0,1]. This is a muscle layer segmentation mask. Pixels with a probability higher than a set threshold are identified as muscle layer regions, otherwise they are considered background.

[0027] Step 4: Damage Area Detection and Edge Calibration The core process involves two steps: Localizing the damaged area using a muscle layer segmentation mask and optimizing boundary accuracy through edge calibration. (1) Initial detection of the damaged area The formula for screening potential injury sites based on abnormal features in the muscle layer region is as follows: Symbol explanation: To perform pixel-by-pixel multiplication, extract the muscle layer region image. ; , These represent the global grayscale mean and standard deviation of the muscle layer region, respectively. This is an abnormal feature map. The larger the value, the higher the probability that the pixel is a damaged area. Areas with values ​​higher than the set threshold are potential damaged areas.

[0028] (2) Edge calibration The edge of the damaged region is refined using a multi-scale gradient weighted algorithm, as shown in the formula: Symbol explanation: For gradient detection kernels of three different scales ( ), capturing edges of different thicknesses respectively; This is a gradient calculation operation; This is the gradient map for the k-th scale; The gradient weights are adaptively assigned based on edge sharpness. The gradient map after calibration; Thresholding is applied to generate a binarized damage region mask. .

[0029] Step 5: Multi-dimensional feature fusion Extract three core features from the damaged area: texture, morphology, and grayscale. Then, enhance the hierarchical recognition through weighted fusion. The formula is: Symbol explanation: Image of the damaged area; Use the Gabor filter operator to extract texture features. ; Morphological feature extraction operators are used to extract features such as area and roundness of the damaged region. ; Extract grayscale features for mean calculation. ; , , For feature fusion weights, satisfying This is used to balance the hierarchical contributions of the three types of features; This is the final fused feature vector.

[0030] Step 6: Damage Severity Assessment The fused features are input into the hierarchical model, and the damage level is output using the following formula: Symbol explanation: The hierarchical model is implemented using a lightweight fully connected network. Output the damage level, with values ​​of {0, 1, 2}, corresponding to mild damage, moderate damage, and severe damage, respectively. Mild damage represents slight damage to the surface of the muscle layer, moderate damage represents partial damage to the thickness of the muscle layer, and severe damage represents deep damage or discontinuity of the muscle layer.

[0031] Step 7: Output Results Mask the damaged area Calibrated edge contour and damage level The image is overlaid on the original hysteroscopic image and simultaneously output to the display interface of the surgical device, providing the surgeon with a real-time and intuitive reference for damage assessment.

[0032] This invention offers significant advantages in terms of adaptability, practicality, and ease of deployment. It can be directly integrated into existing hysteroscopic surgical equipment without additional hardware investment, significantly lowering the barrier to clinical application. Through standardized damage detection and rating procedures, it reduces diagnostic errors caused by differences in physician experience, improving the safety and standardization of hysteroscopic surgery. In the future, it can be further extended to other gynecological endoscopic diagnostic and treatment scenarios, providing technical reference for the intelligent upgrading of minimally invasive surgery, and possessing broad clinical application value and social significance. The above description is merely a preferred embodiment of the invention and is not intended to limit the invention. For those skilled in the art, the invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration, characterized in that: Includes the following steps: Step 1: Hysteroscopic image acquisition: Real-time images of the uterine cavity during the procedure are acquired using the built-in camera of the hysteroscopic device, and imaging parameters are recorded simultaneously; Step 2: Image preprocessing: Adaptive denoising and illumination equalization are performed on the acquired images to eliminate environmental interference factors; Step 3: Two-branch segmentation of the uterine myometrium: Using a Transformer and CNN two-branch network, global and local features of the myometrium are extracted to generate a myometrium region mask; Step 4: Damage area detection and edge calibration: Based on the segmented muscle layer region, potential damage sites are identified, and the damage boundary is optimized through an edge calibration algorithm; Step 5: Multi-dimensional feature fusion: Extract texture, shape, and grayscale features of the damaged area, and enhance the hierarchical recognition through weighted fusion; Step 6: Damage severity rating: Input the fused features into the grading model and output the damage level: mild, moderate, severe; Step 7: Result Output: Simultaneously display the location of the damaged area, edge contour, and damage level, providing the surgeon with an intuitive reference.

2. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 1, characterized in that: In step 1, a high-definition RGB camera built into the hysteroscope is used to acquire real-time images of the uterine cavity during the procedure. The image resolution is adapted to the device's hardware specifications. During acquisition, it is ensured that there are no serious instruments obstructing the myometrium, and the image includes a complete view of the uterine cavity wall. The acquired raw image is recorded as follows: ,in The coordinates are pixel coordinates, and the image channels are R, G, and B channels.

3. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 2, characterized in that: Step 2, through adaptive denoising and illumination equalization, provides high-quality input for subsequent segmentation and detection, and includes the following steps: (1) Adaptive noise suppression A context-aware weighted filtering algorithm is used to suppress noise while preserving tissue edge details. The formula is as follows: Symbol explanation: These are the pixel values ​​of the denoised image; The width of the filter window is half its width. For adaptive weighting coefficients, derived from pixels With neighboring pixels ( The RGB grayscale difference and gradient information are calculated together. Regions with small grayscale differences and gentle gradients have higher weights to avoid edge blurring. (2) Light equalization The partitioned adaptive gamma correction algorithm is used to correct uneven illumination in images. The formula is as follows: Symbol explanation: Represents the R, G, and B channels; The c-channel pixel value after illumination correction; These are pixel-level adaptive gamma coefficients, calculated from the mean and variance of the grayscale value of the local region where the pixel is located, for dark areas. To enhance brightness, bright areas To reduce brightness and achieve global illumination balance.

4. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 3, characterized in that: Step 3 employs a dual-branch network structure combining Transformer and CNN to extract and fuse global and local features respectively, achieving accurate segmentation of the muscle layer region. This includes two steps: feature extraction and fusion. (1) Dual-branch feature extraction Transformer branch: For Perform block encoding and extract global semantic features, using the following formula: Symbol explanation: For image segmentation, the image is divided into feature blocks of fixed size; It is a Transformer encoder that captures global feature associations through a self-attention mechanism; For global feature maps; CNN branch: Employs convolutional neural networks to extract local texture and edge features, using the following formula: Symbol explanation: The convolutional module extracts local detail features through multi-scale convolutional kernels; This is a local feature map; (2) Feature fusion and segmentation Muscle layer segmentation mask is generated by fusing bi-branch features using attention weighting, as shown in the formula: Symbol explanation: This is a channel attention mechanism used to enhance effective information in local features; To integrate weights, the contributions of global and local features are adaptively balanced; To fuse feature maps; The activation function maps features to probability values ​​in the interval [0,1]. This is a muscle layer segmentation mask. Pixels with a probability higher than a set threshold are identified as muscle layer regions, otherwise they are considered background.

5. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 4, characterized in that: Step 4, which locates the damaged area based on the muscle layer segmentation mask and optimizes the boundary accuracy through edge calibration, includes two steps: (1) Initial detection of the damaged area The formula for screening potential injury sites based on abnormal features in the muscle layer region is as follows: Symbol explanation: To perform pixel-by-pixel multiplication, extract the muscle layer region image. ; , These represent the global grayscale mean and standard deviation of the muscle layer region, respectively. This is an abnormal feature map. The larger the value, the higher the probability that the pixel is a damaged area. Areas with values ​​higher than the set threshold are potential damaged areas. (2) Edge calibration The edge of the damaged region is refined using a multi-scale gradient weighted algorithm, as shown in the formula: Symbol explanation: Three different scale gradient detection kernels are used. Each edge is captured with a different thickness; This is a gradient calculation operation; This is the gradient map for the k-th scale; The gradient weights are adaptively assigned based on edge sharpness. The gradient map after calibration; Thresholding is applied to generate a binarized damage region mask. .

6. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 5, characterized in that: Step 5 extracts three core features of the damaged area: texture, shape, and grayscale. Weighted fusion is then used to enhance hierarchical recognition. The formula is: Symbol explanation: Image of the damaged area; Use the Gabor filter operator to extract texture features. ; Morphological feature extraction operators are used to extract features such as the area and roundness of the damaged region. ; Extract grayscale features for mean calculation. ; , , For feature fusion weights, satisfying This is used to balance the hierarchical contributions of the three types of features; This is the final fused feature vector.

7. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 6, characterized in that: In step 6, the fused features are input into the hierarchical model, and the damage level is output using the following formula: Symbol explanation: The hierarchical model is implemented using a lightweight fully connected network. Output the damage level, with values ​​of {0, 1, 2}, corresponding to mild damage, moderate damage, and severe damage, respectively. Mild damage represents slight damage to the surface of the muscle layer, moderate damage represents partial damage to the thickness of the muscle layer, and severe damage represents deep damage or discontinuity of the muscle layer.

8. The method for rating uterine myometrial damage based on bi-branch feature fusion and edge calibration according to claim 7, characterized in that: In step 7, the damaged area is masked. Calibrated edge contour and damage level The image is overlaid on the original hysteroscopic image and simultaneously output to the display interface of the surgical device, providing the surgeon with a real-time and intuitive reference for damage assessment.

Images