Gynecological clinical examination diagnosis and treatment process field of view clarity enhancement system and implementation method

Abstract

The present application relates to the technical field of gynecological examination, in particular to a gynecological clinical examination and diagnosis process field of view definition enhancement system and implementation method thereof. It includes: image acquisition and preprocessing unit real-time acquisition of gynecological examination process field of view image, and the field of view image is pretreated; feature extraction unit extracts key features from the field of view image after pretreatment; definition evaluation and optimization unit generates definition score based on key features, uses polarization difference image and polarization average image to decouple and correct feature interference caused by mucus scattering and tissue reflection, generates optimized definition score, and generates definition optimization parameters through multi-task learning model; field of view enhancement execution unit enhances the definition of field of view image based on definition optimization parameters. The present application carries out feature correction from the imaging physical mechanism, so that the definition score can more truly reflect the intrinsic details of the tissue, and provide reliable basis for subsequent accurate enhancement.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gynecological examination, in particular to a gynecological clinical examination and diagnosis process field of view clarity enhancement system and implementation method thereof. BACKGROUND

[0002] In the process of gynecological clinical examination and diagnosis, especially in the operation of endoscopes such as colposcopy and hysteroscopy, doctors highly rely on real-time images to observe key anatomical structures such as the cervix, endometrium and fallopian tube openings, in order to achieve early identification and diagnosis of diseases such as inflammation, polyps and precancerous lesions. However, due to the special physiological environment of the female reproductive tract, the surface of the tissue to be examined is often covered with mucus, blood or secretions, which can cause multiple scattering of light, resulting in blurred images, reduced contrast and loss of details; at the same time, the strong specular reflection (high light) produced by the wet tissue surface is easy to cause local overexposure, covering the lesion edge and small structures. The above interference factors seriously reduce the clarity and distinguishability of the examination field of view, increasing the risk of missed diagnosis or misdiagnosis by doctors;

[0003] Existing image enhancement techniques mostly use general algorithms (such as histogram equalization, sharpening filtering, etc.), lack understanding of the unique interference physical mechanism of gynecology, and are difficult to effectively distinguish and suppress the mixed effects of mucus scattering and tissue reflection, often leading to problems such as noise amplification, color distortion or over-enhancement. In addition, traditional methods usually only process based on image statistical features, without differentiation optimization combined with the anatomical target of the current examination (such as focusing on overall visual perception for cervical observation, and paying attention to texture details for hysteroscopy), lacking the ability to perceive the intention of clinical workflow. More importantly, since the interference can significantly distort the brightness, contrast and edge of the image, direct use for clarity evaluation will lead to scoring bias, and thus mislead the subsequent enhancement strategy. Therefore, a gynecological clinical examination and diagnosis process field of view clarity enhancement system and implementation method thereof are provided. SUMMARY

[0004] The purpose of the present application is to provide a gynecological clinical examination and diagnosis process field of view clarity enhancement system and implementation method thereof, to solve the problem of image quality degradation caused by mucus scattering and tissue reflection, and the lack of precise separation of traditional image enhancement methods, the clarity evaluation is easy to be distorted by interference, and the optimization process does not combine with the intention of clinical diagnosis, resulting in poor enhancement effect.

[0005] To achieve the above purpose, on the one hand, the present application provides a gynecological clinical examination and diagnosis process field of view clarity enhancement system, comprising:

[0006] An image acquisition and preprocessing unit, which acquires the field of view image in real time during gynecological examination, and pre-processes the field of view image;

[0007] a feature extraction unit configured to extract key features from the preprocessed field of view image;

[0008] a definition evaluation and optimization unit configured to generate a definition score based on the key features, decouple and correct feature interference caused by mucus scattering and tissue reflection using the polarization difference image and the polarization average image, generate an optimized definition score, and generate definition optimization parameters through a multi-task learning model;

[0009] a field of view enhancement execution unit configured to enhance the definition of the field of view image based on the definition optimization parameters.

[0010] As a further improvement of the technical solution, the image acquisition and preprocessing unit comprises an image acquisition module and an image processing module;

[0011] The image acquisition module acquires the field of view image during the gynecological examination based on the optical illumination module, and synchronously acquires two polarization images of the same field of view and ;

[0012] The image processing module pre-processes the acquired field of view image.

[0013] As a further improvement of the technical solution, the feature extraction unit comprises a brightness distribution module, a contrast module, an edge definition module, and a blur degree module;

[0014] The brightness distribution module is configured to statistically analyze the overall brightness histogram of the preprocessed field of view image to obtain brightness distribution features reflecting the overall brightness of the image;

[0015] The contrast module is configured to analyze the gray scale difference of the local area of the field of view image to extract contrast features reflecting the distinction degree of the tissue details of the field of view image;

[0016] The edge definition module extracts edge definition features using an edge detection method to represent the definition degree of the tissue profile of the field of view image;

[0017] The blur degree module calculates blur degree features through a frequency domain analysis method to reflect the blur area in the field of view image due to inaccurate focusing.

[0018] As a further improvement of the technical solution, the definition evaluation and optimization unit generates a definition score based on the key features, comprising the following steps:

[0019] S1.1, calculating a polarization difference image based on the two polarization images and ​and a polarization average image ;

[0020] S1.2, based on the difference in polarization characteristics of mucus and reflection light, respectively detecting mucus regions based on scattering entropy and detecting reflection light regions based on polarization difference, decoupling and correcting the extracted key features, and generating decoupled and corrected key features, wherein the extracted key features include brightness distribution features, contrast features, edge sharpness features and blur degree features;

[0021] S1.3, normalizing the decoupled and corrected key features;

[0022] S1.4, assigning a preset weight coefficient according to the importance of each decoupled and corrected key feature in image sharpness evaluation;

[0023] S1.5, calculating sub-feature scores of key features, including brightness balance score, contrast score, edge sharpness score and blur score, reflecting image quality in each single feature dimension;

[0024] S1.6, weighting and fusing the sub-feature scores of key features according to the weights, and outputting a comprehensive sharpness score , used to represent the overall sharpness level of the current field of view image.

[0025] As a further improvement of the technical solution, in S1.2, the decoupled and corrected key features are generated, including the following steps:

[0026] S1.21, based on the brightness threshold of the polarization difference image , extracting a reflection light mask ;

[0027] S1.22, calculating local scattering entropy of the polarization average image , and recording the region with local scattering entropy value higher than the mucus threshold as a mucus mask ;

[0028] S1.23, in the overall brightness histogram statistics, excluding the reflection light mask and the mucus mask region, and only using the normal region to calculate the reference brightness distribution;

[0029] S1.24, in the mucus mask region, using a scattering model for contrast compensation, and restoring the original gray difference based on the polarization difference value;

[0030] S1.25, in the reflection light mask The region utilizes the polarization image to repair the edge break caused by overexposure across the channel edge information.

[0031] As a further improvement of the technical solution, the sharpness evaluation and optimization unit generates sharpness optimization parameters through a multi-task learning model, including the following steps:

[0032] S2.1, receiving the decoupled corrected key features output by the feature extraction unit, and synthesizing the sharpness score ;

[0033] S2.2, real-time analysis of the current field of view image using a lightweight convolutional neural network, identifying the main anatomical structures therein, and taking the identification result as a workflow state signal vector;

[0034] S2.3, the feature vector of the decoupled corrected key features and the workflow state signal vector are jointly input into the multi-task learning model;

[0035] S2.4, based on the output of the multi-task learning model, generating an optimization strategy weight vector and a basic parameter adjustment amount;

[0036] S2.5, the basic parameter adjustment amount, the optimization strategy weight vector, and the sub-feature score of each key feature are fused and calculated to obtain a weighted final parameter adjustment amount The final parameter adjustment amount is the sharpness optimization parameter.

[0037] As a further improvement of the technical solution, in S2.2, the lightweight convolutional neural network is used to analyze the current field of view image in real time, and the main anatomical structures therein are identified, including the following steps:

[0038] S2.21, receiving the preprocessed field of view image output by the image acquisition and preprocessing unit, and standardizing the preprocessed field of view image;

[0039] S2.22, input the standardized field of view image into the lightweight convolutional neural network;

[0040] S2.23, extract multi-scale feature maps through convolution and pooling operations of the lightweight convolutional neural network;

[0041] S2.24, input the multi-scale feature maps into the classification module to identify the main anatomical structures;

[0042] S2.25, one-hot encoding the identification result to generate a workflow state signal vector.

[0043] As a further improvement of the technical solution, in S2.4, the generation of the optimization strategy weight vector is specifically: the decoupled and corrected key feature vector is spliced with the workflow state signal vector in the channel dimension to form a fused feature vector, and the fused feature vector is input into a shared deep neural network encoder to extract a deep shared representation containing the current image quality condition and the current clinical task target ; the deep shared representation is input into a clinical intent analysis decoder to generate an optimization strategy weight vector.

[0044] As a further improvement of the technical solution, the field of view enhancement execution unit receives the definition optimization parameter generated by the definition evaluation and optimization unit, maps the definition optimization parameter to an optimization scheme, executes the optimization scheme to generate an enhanced final field of view image, and outputs the final field of view image to a clinical display interface.

[0045] In another aspect, the present application provides a method for enhancing the definition of the field of view in the process of gynecological clinical examination and diagnosis and treatment, based on the gynecological clinical examination and diagnosis and treatment process field of view definition enhancement system of any one of the above, comprising the following steps:

[0046] S3.1, real-time acquisition of the field of view image in the process of gynecological examination, and pretreatment of the field of view image;

[0047] S3.2, extracting key features from the pretreated field of view image;

[0048] S3.3, generating a definition score based on the key features, decoupling and correcting the feature interference caused by mucus scattering and tissue reflection using polarized difference images and polarized average images, generating an optimized definition score, and generating a definition optimization parameter through a multi-task learning model;

[0049] S3.4, enhancing the definition of the field of view image based on the definition optimization parameter.

[0050] Compared with the prior art, the present application has the following advantages:

[0051] 1、In the present application, by integrating a switchable linear polarized light source and a polarization-sensitive camera, 0° and 90° polarization images are synchronously collected to construct polarization difference images and polarization average images, and the essential differences in polarization characteristics between mucus multiple scattering and tissue specular reflection are utilized to extract the light reflection mask and the mucus mask respectively. On this basis, key features such as brightness, contrast and edge are regionally shielded and physically model compensated (such as detecting mucus area based on scattering entropy and implementing contrast recovery, repairing overexposed edges using polarization channel edge consistency), effectively eliminating the coupling effect of physiological interference on image quality evaluation. Feature correction is carried out from the imaging physical mechanism, making the sharpness score more truly reflect the intrinsic details of the tissue, and providing a reliable basis for subsequent precise enhancement.

[0052] 2、In the present application, key anatomical structures such as cervix, endometrium and fallopian tube opening are identified in real time by a lightweight convolutional neural network, and a workflow state signal is generated; combined with the decoupled image quality features, the basic adjustment amount and the optimization strategy weight vector are jointly predicted by a multi-task learning model. The final parameter adjustment amount takes into account the image quality defect degree (sub-feature score), the diagnostic focus of the current examination site (such as brightness equalization for cervical mode, and emphasis on contrast and edge for uterine cavity mode) and the clinical intention priority, realizing the intelligent enhancement of "different for different parts, different for different lesions". It breaks through the "one-size-fits-all" limitation of traditional image enhancement algorithms, and improves the visual experience and diagnosis efficiency of doctors at different examination stages, with high clinical adaptability and practicality. Figure One BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The figure is the overall flow chart of the present application;

[0054] The meanings of the various numbers in the figure are as follows:

[0055] 1, image acquisition and preprocessing unit; 2, feature extraction unit; 3, sharpness evaluation and optimization unit; 4, field of view enhancement execution unit. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] Embodiment 1: Please refer to Figure 1 as shown, a field of view sharpness enhancement system in gynecological clinical examination and diagnosis process is provided, which comprises:

[0058] ​The image acquisition and preprocessing unit 1 acquires the visual field image in the gynecological examination process in real time, and pre-processes the visual field image;

[0059] In this embodiment, the image acquisition and preprocessing unit 1 includes an image acquisition module and an image processing module;

[0060] The image acquisition module acquires the visual field image in the gynecological examination process based on an optical illumination module, wherein the optical illumination module provides a stable, uniform, and adjustable light environment, illuminates the tissue to be examined through the optical illumination module, and acquires the visual field image in the gynecological examination process through an endoscope lens based on the illuminated tissue to be examined, and integrates a switchable linearly polarized light source (0° and 90° polarization directions) and a polarization-sensitive camera (equipped with orthogonal polarizing plates) to synchronously acquire two polarization images of the same visual field and ;

[0061] The image processing module pre-processes the acquired visual field image, and the pre-processing includes denoising, brightness and contrast correction, gamma correction, edge preservation filtering, and shadow compensation to obtain a standardized image with higher quality.

[0062] The feature extraction unit 2 extracts key features from the pre-processed visual field image;

[0063] In this embodiment, the feature extraction unit 2 includes a brightness distribution module, a contrast module, an edge sharpness module, and a blur degree module;

[0064] The brightness distribution module is used to statistically analyze the overall brightness histogram of the pre-processed visual field image to obtain a brightness distribution feature reflecting the overall brightness of the image, specifically: the pre-processed visual field image is converted into a grayscale image, and then the overall brightness histogram of the image is calculated, that is, the number of pixels of each grayscale level appearing in the image is counted to obtain the grayscale distribution of the image; then the histogram can be normalized to obtain the probability distribution of each grayscale level, thereby forming a brightness distribution feature vector reflecting the overall brightness of the image;

[0065] The contrast module is used to analyze the gray difference of the local area of the visual field image, and extract a contrast feature reflecting the distinguishability of the tissue details of the visual field image, specifically: the pre-processed visual field image is divided into a plurality of local windows or sub-regions, the variance of the gray value or the gray difference index of each local region is calculated to quantify the degree of change of the brightness of the pixels in the region, and the distinguishability of the tissue details is reflected; then the contrast indexes of all local regions are summarized or weightedly averaged to form a whole contrast feature vector;

[0066] The edge sharpness module extracts edge sharpness features using an edge detection method, which is used to represent the sharpness of the tissue profile in the visual field image. Specifically, an edge detection operator (Sobel) is applied to the pre-processed visual field image to extract edge responses in the image and obtain edge intensity information for each pixel point. Then, the edge intensity is counted or normalized to form an edge sharpness feature vector.

[0067] The blur degree module calculates the blur degree feature through a frequency domain analysis method, which is used to reflect the blurred area in the visual field image due to inaccurate focusing. Specifically, the processing process of the blur degree module is as follows: first, a two-dimensional fast Fourier transform is performed on the pre-processed visual field image to convert the image from the spatial domain to the frequency domain; then, the proportion of the energy of the high-frequency component in the total energy in the frequency spectrum is calculated, and the high-frequency component reflects the details and sharpness of the image, while the low-frequency component mainly corresponds to the smooth area; the blur degree feature is obtained by inverse quantization of the proportion.

[0068] The clarity evaluation and optimization unit 3 generates a clarity score based on key features, decouples and corrects the feature interference caused by mucus scattering and tissue reflection using polarized difference images and polarized average images, generates an optimized clarity score, and generates clarity optimization parameters through a multi-task learning model;

[0069] In this embodiment, in gynecological clinical examination, mucus adhesion can cause a semi-transparent cover in the local image, causing light scattering and detail blur; tissue reflection can form a highlight area, causing local overexposure and detail loss. These two interferences can significantly distort the brightness distribution, contrast, and edge sharpness features, causing the clarity score to deviate; traditional methods usually consider both as a unified image quality degradation phenomenon and use a global enhancement strategy (such as histogram equalization or general deblurring), but cannot distinguish their physical causes, which instead leads to noise amplification or detail distortion; this system highlights reflection (specular reflection dominant) through polarized difference images and highlights mucus (multiple scattering dominant) through polarized average images, and quantifies the local confusion degree through scattering entropy, achieving physical separation and accurate positioning of the two types of interference areas, providing a reliable basis for subsequent feature correction; based on the decoupled masks (reflection mask, mucus mask), repair strategies are designed respectively: in the reflection area, the polarized cross-channel edge consistency is used to repair the broken profile, and in the mucus area, the scattering model is used to compensate for the loss of gray scale. This partition optimization strategy avoids the side effects of global processing and significantly improves the accuracy of detail recovery;

[0070] The clarity evaluation and optimization unit 3 generates a clarity score based on key features, including the following steps:

[0071] S1.1, based on two polarized images and Calculate the polarized difference image ​ ), and polarization-averaged images ( ), where polarization difference image Polarization-averaged image used to highlight reflective areas Used to highlight the mucus scattering area;

[0072] S1.2 Utilizing the difference in polarization characteristics between slime (multiple scattering) and reflection (specular reflection), slime regions are detected based on scattering entropy and reflection regions are detected based on polarization difference, respectively. The extracted key features are decoupled and corrected to generate decoupled and corrected key features. The extracted key features include brightness distribution features, contrast features, edge sharpness features, and blur degree features. The decoupled and corrected key features include optimized brightness distribution features, contrast features, edge sharpness features, and the original blur degree features.

[0073] The process of generating the decoupling-corrected key features includes the following steps:

[0074] S1.21, Based on polarization difference image Brightness threshold Segmentation and extraction of reflective mask It is used to locate and mark overexposed areas in the field of view caused by tissue reflection;

[0075] Segmentation and extraction of reflective mask The formula is (for polarization difference images) Each pixel in To make a judgment, the polarization difference image At pixel grayscale value ( If greater than or equal to the reflectivity threshold Then in the binary mask The reflective area is marked as 1 (reflective area) if it is not, and 0 (non-reflective area) otherwise; reflective threshold The reflectivity threshold in this embodiment can be preset using statistical experimental data. The value is 30. This formula segments bright areas by setting a brightness threshold on the polarization difference image. When the pixel gray value is greater than the threshold, it is determined to be a reflective area, thereby generating a reflective mask. Its function is to effectively identify overexposed interference areas on the tissue surface caused by specular reflection, preparing for subsequent correction.

[0076] ;

[0077] S1.22, Polarization-averaged image Calculate the local scattering entropy; the local scattering entropy value is higher than the mucus threshold. The area is denoted as the mucus mask. , for detecting the area blurred by mucus scattering;

[0078] The local scattering entropy is calculated as:

[0079] For each pixel point in the image , a local window (e.g. ) around it is taken , and the local scattering entropy value is calculated as: , where is the number of gray levels (e.g. 256), is the probability (i.e. frequency divided by the total number of pixels in the window) of the gray value appearing in the local window , and is the local scattering entropy value, reflecting the degree of gray level confusion in the local area. The mucus-covered area will have an increased entropy value due to multiple scattering effects. The purpose is to accurately detect the scattering blur area caused by mucus coverage, so as to isolate such interference;

[0080] The mucus mask is generated by comparing the local scattering entropy value of each pixel point with the mucus threshold value . If it is higher than the mucus threshold value , it is considered to be a mucus area and marked as 1 in the mask. In this embodiment, the mucus threshold value is 3, calibrated by experiment:

[0081] ;

[0082] S1.23, in the overall brightness histogram statistics, exclude the areas of the glare mask and the mucus mask , and only use the normal area to calculate the reference brightness distribution. This step optimizes the brightness feature and avoids interference with the brightness feature, making the brightness evaluation more realistic and reflecting the inherent situation of the tissue;

[0083] S1.24, in the mucus mask area, use the scattering model for contrast compensation, and restore the original gray difference based on the polarization difference value. This step optimizes the contrast feature based on the linear scattering model of the polarization physical mechanism, and compensates for the scattering loss in the polarization average image by extracting direct component information from the polarization difference image, thereby accurately and efficiently reversing the problem of reduced contrast caused by multiple scattering of biological mucosa;

[0084] The scattering model is (this model restores the direct gray information weakened by scattering to improve the degree of detail differentiation and correct the contrast reduction problem caused by mucus):

[0085] ;

[0086] For each pixel in the mucus region, its compensated gray value is obtained by adding a compensation term to the value of the polarization average image ; the compensation term is linearly related to the polarization difference value with a gain coefficient and an offset , in this embodiment, is less than 1 to avoid over-compensation, is set to 0 or a small positive value, and are calibrated by experiments); because reflects the direct component information that is not contaminated by scattering (i.e. the part of the light signal that is directly reflected or transmitted to the camera after being irradiated on the tissue surface without multiple scattering through the mucus or other semi-transparent medium), it can be used to partially reverse the contrast loss caused by scattering; the compensated image will be used to re-calculate the local contrast feature of the region;

[0087] S1.25, in the highlight mask region, use the polarization image to repair the edge breakage caused by overexposure using the edge consistency (e.g. the consistency of the edges of and ); this step is to optimize the edge sharpness feature and restore the tissue profile that is lost due to overexposure;

[0088] The specific process of repairing the edge breakage caused by overexposure is as follows: use the Sobel operator to extract the edges of the two polarization images and respectively: , , define the edge consistency function ( , in which is a very small constant to prevent division by zero error); in the highlight region , if a pixel point has a weak edge response in but has a high edge consistency, it is considered that the edge is truly present but is suppressed by overexposure and should be enhanced or repaired; the repaired edge map will be used to generate a more accurate edge sharpness feature:

[0089] , ;

[0090] in which is the edge consistency threshold, in this embodiment, is 0.8; the repaired edge intensity map.

[0091] S1.3, normalizing the decoupled corrected key features;

[0092] S1.4, assigning a preset weight coefficient according to the importance of each decoupled corrected key feature in the image definition evaluation;

[0093] S1.5, calculating the sub-feature scores of the key features respectively, the sub-feature scores of the key features including the brightness balance score, the contrast score, the edge definition score and the blur score, reflecting the image quality of each single feature dimension;

[0094] The brightness balance score is: ;

[0095] The contrast score is: ;

[0096] The edge definition score is: ;

[0097] The blur score is: ;

[0098] In the formula, is the global gray mean value of the field of view image, which comes from the area excluding the areas marked by the light shielding mask and the mucus mask, is the ideal brightness value (the empirical value is the middle value of the gray scale range), is the adjustment constant for controlling the decay rate (determined by experiment, set to 50 in the embodiment), is the global gray standard deviation of the field of view image, which comes from the field of view image after recovery by the scattering model compensation (the recovery by the scattering model compensation is specifically step S1.24), is the adjustment constant for controlling the growth rate (determined by experiment, set to 20 in the embodiment), is the repaired edge intensity map, is the total number of edge pixels in the field of view image, used for normalization to avoid the influence of the number of edges on the average intensity, is the energy of high frequency components (usually defined as the part outside the central area of the spectrum, i.e. the sum of squares of the spectrum amplitude), which is the energy of high frequency components in the spectrum obtained by fast Fourier transform of the field of view image, is the total energy of the entire spectrum, is the two-dimensional coordinate in the spectrum, representing the frequency domain position of the image after Fourier transform; ​​​​

[0099] S1.6, the sub-feature scores of the key features are weighted and fused according to the weights, and a comprehensive definition score is output , for representing the overall definition level of the current field image;

[0100] The comprehensive definition score is:

[0101] ;

[0102] In the formula, is the weight coefficient of the brightness balance degree, is the weight coefficient of the contrast, is the weight coefficient of the edge definition, is the weight coefficient of the blur, , , , , ,

[0103] In this embodiment, the score is used to quantify the quality level of the current field image in four dimensions of brightness balance, contrast, edge definition and blur. By calculating the sub-feature scores and fusing them into a comprehensive definition score, the overall definition condition of the image can be intuitively reflected. In the optimization link, the system will sort the defects in each dimension according to the score results, and the lower the score value, the more serious the problem in that dimension, and the higher the corresponding adjustment priority.

[0104] Further, the definition evaluation and optimization unit 3 generates definition optimization parameters through a multi-task learning model, including the following steps:

[0105] S2.1, receiving the decoupled corrected key features and the comprehensive definition score output by the feature extraction unit 2 ;

[0106] S2.2, using a lightweight convolutional neural network (CNN) to perform real-time analysis on the current field image, identifying the main anatomical structures (including the cervix, endometrial cavity, and fallopian tube opening) therein, and taking the identification results as a workflow state signal vector;

[0107] Among them, the traditional image enhancement system adopts a one-size-fits-all fixed optimization strategy, which cannot adapt to the different requirements of different anatomical sites (such as cervix, uterine cavity, fallopian tube) in gynecological clinical examination, and takes the clinical diagnosis intention (workflow state) as the key input, realizes a scene-aware adaptive optimization through a multi-task learning model. The prior art mostly only mechanically enhances according to the low-level features of the image itself (such as low contrast and dark brightness). The system identifies the main anatomical structure in the current field of view in real time through a lightweight CNN, and fuses this workflow state signal with the image features, and infers through a deep learning model which quality dimensions should be prioritized for enhancement (such as giving higher weights to contrast and edge definition in the uterine cavity mode), so that the optimization strategy changes from blind generalization to targeted precision, truly matching the doctor's diagnosis needs;

[0108] Real-time analysis of the current field of view image using a lightweight convolutional neural network (CNN) to identify the main anatomical structure, including the following steps:

[0109] S2.21, receive the pre-processed field of view image output by the image acquisition and preprocessing unit 1, and standardize the pre-processed field of view image to adjust it to a standard size (e.g. 224x224 pixels) and channel number suitable for CNN input;

[0110] S2.22, input the standardized field of view image into a lightweight convolutional neural network (CNN), which includes 3 to 5 convolutional layers, with the number of convolutional kernels varying from 16, 32 to 64, and after convolution, batch normalization and ReLU activation layers are connected in turn, and a maximum pooling layer is added after part of the convolutional layers to reduce the size of the feature map and extract multi-scale features; the network ends with one or two fully connected layers to map the multi-scale features to the main anatomical structure class output, the entire CNN can be optionally trained from scratch or pre-trained on ImageNet and other datasets to speed up convergence, and supports standardized processing of input images with a size of 224x224x3;

[0111] S2.23, extract multi-scale feature maps through convolution and pooling operations of the lightweight convolutional neural network (CNN), and retain texture and edge information of the main anatomical structure;

[0112] S2.24, input the multi-scale feature map into the classification module to identify the main anatomical structure, which includes the cervix, endometrium in the uterine cavity, and fallopian tube opening;

[0113] S2.25, one-hot encoding of the recognition result to generate a workflow state signal vector: the workflow state signal vector is: [1, 0, 0] represents the cervix mode, [0, 1, 0] represents the uterine cavity mode, and [0, 0, 1] represents the fallopian tube mode;

[0114] S2.3, input the feature vector of the decoupled and corrected key features and the workflow state signal vector together into a multi-task learning model, wherein the architecture of the multi-task learning model is an end-to-end neural network structure, the input is the decoupled and corrected key feature vector and the one-hot encoded workflow state signal, common feature representations are extracted through a plurality of shared convolutional or fully connected layers, and then divided into a plurality of task branches: one branch is used for clarity regression task to output a comprehensive clarity score, one branch is used for interference discrimination task to identify reflective or mucus areas, and one branch is used for optimization parameter prediction task to output parameter adjustment amounts of brightness, contrast, edge clarity and blurring; each task branch is jointly trained through shared features and task-specific layers to achieve multi-dimensional visual field image clarity optimization;

[0115] S2.4, based on the output of the multi-task learning model, an optimization strategy weight vector and a basic parameter adjustment amount are generated, the optimization strategy weight vector is obtained from the multi-task learning model, and reflects the relative importance priority of the four optimization dimensions of brightness, contrast, edge clarity and blurring under the current clinical workflow state, for example, in the cervical mode, the weight vector may be [brightness: 0.5, contrast: 0.3, edge: 0.1, blur: 0.1], and the body view is adjusted; and in the uterine cavity mode, the vector becomes [brightness: 0.2, contrast: 0.4, edge: 0.4, blur: 0.0], and in the uterine cavity mode, the endometrial texture and opening details are usually paid more attention to, so the weights of the contrast and edge features are higher, and the blur correction weight is lower; the basic parameter adjustment amount is a general, unweighted optimization suggestion, including brightness adjustment value , contrast adjustment value , edge adjustment value , and deblurring adjustment value , which is a standard adjustment scheme given by the model based on the quality condition of the image features (brightness, contrast, etc.), for example, the model detects that the image is overall dark, and it will output a basic adjustment value of increasing brightness;

[0116] Further, the generation of the optimization strategy weight vector is specifically: the decoupled and corrected key feature vector and the workflow state signal vector are spliced in the channel dimension to form a fused feature vector, the fused feature vector is input into a shared deep neural network encoder (a multi-layer fully connected network) to extract deep shared representations containing the current image quality condition and the current clinical task target ; the deep shared representations are input into a clinical intent analysis decoder to generate an optimization strategy weight vector (the clinical intent analysis decoder is a simple fully connected network, and a Softmax activation layer is connected after the fully connected network, and the clinical intent analysis decoder extracts the shared representations The four-dimensional vector is mapped by one to two full connection layers, corresponding to four dimensions of brightness, contrast, edge definition and blurriness respectively; a Softmax function is applied to the four-dimensional vector for normalization processing to ensure that the sum of the four weights is 1, so as to obtain a standard, interpretable probability distribution form of the optimization strategy weight vector, and the optimization strategy weight vector includes the weight of brightness adjustment, the weight of contrast adjustment, the weight of edge adjustment and the weight of blurriness adjustment;

[0117] The architecture of the shared deep neural network encoder is a multi-layer fully connected network, and the input is the spliced fusion feature vector, which includes the key feature feature vector after decoupling correction and the current workflow state signal vector; after a plurality of fully connected layers (each layer is followed by normalization and nonlinear activation function (such as ReLU) activation), high-dimensional features are extracted and deep shared representations are generated step by step; the intermediate layer contains a dimension reduction layer or a residual connection to enhance the feature expression ability; the output is a deep shared representation; the representation is usually a lower-dimensional continuous vector, for example, 128-dimensional or 256-dimensional, which is used to uniformly encode image quality and clinical task information;

[0118] S2.5, the basic parameter adjustment amount, the optimization strategy weight vector, and the sub-feature scores of each key feature are fused and calculated to obtain a weighted final parameter adjustment amount The final parameter adjustment amount is the definition optimization parameter;

[0119] The final parameter adjustment amount is:

[0120] ;

[0121] In the formula, represents the weight of brightness adjustment, represents the weight of contrast adjustment, represents the weight of edge adjustment, represents the weight of blurriness adjustment;

[0122] The final parameter adjustment amount includes a brightness adjustment coefficient, a contrast gain coefficient, an edge enhancement coefficient, and a deblurring parameter. The final parameter adjustment amount is a personalized and weighted final execution scheme. It not only considers the quality of the field image itself, but also incorporates the current clinical intention (workflow state) and the current score of each feature. Different examination sites (cervix, uterine cavity, and fallopian tube) have different requirements for the field image quality: cervical examination: more attention is paid to overall appearance and color uniformity (high brightness weight); uterine cavity examination: more attention is paid to endometrial texture and small lesion details (high contrast and edge definition weight). If only the general parameter adjustment amount output by the model is used, all scenes will be optimized without distinction, which cannot best assist the doctor in the current key diagnosis. The sub-feature score quantitatively reflects the current status of the image in each quality dimension. The system needs to determine the optimization intensity according to the status. For example: if the contrast score of the current image is already high (such as 0.9), even if the uterine cavity mode weight indicates that attention should be paid to contrast, the system will only fine-tune or maintain it because it is already good enough, and excessive optimization will introduce noise; on the contrary, if the blurriness score of the current image is low (the image is very blurred), and the parameter adjustment amount also suggests deblurring, then the system will use a stronger deblurring filter.

[0123] The field enhancement execution unit 4 enhances the sharpness of the field image based on the sharpness optimization parameter;

[0124] In this embodiment, the field enhancement execution unit 4 receives the sharpness optimization parameter generated by the sharpness evaluation and optimization unit 3, maps the sharpness optimization parameter to an optimization scheme, executes the optimization scheme to generate an enhanced final field image, and outputs the final field image to the clinical display interface;

[0125] The optimization scheme is: according to the deblurring parameter, Wiener filtering is performed on the image to reverse the out-of-focus blur; then according to the brightness and contrast adjustment coefficients, linear transformation is performed to overall improve the picture brightness and the gray difference between tissues; finally, based on the edge enhancement coefficient, adaptive non-sharpening masking technology is used to sharpen the tissue contour and texture details, and the intensity of each operation is weighted, fused, and adaptively adjusted according to the specific scene of gynecological examination (such as uterine cavity examination mode);

[0126] The optimization scheme is mapped to a series of specific device instructions and algorithm parameters: for example, the deblurring parameters are converted into instructions for controlling the kernel size (e.g., 5x5 pixels) and noise power ratio (NSR = 0.02~0.1) of a real-time Wiener filter; the contrast gain coefficient is converted into a curve slope for adjusting a programmable lookup table (LUT) that is mapped in a dynamic range (e.g., 12-bit data of 0-4096); and the edge enhancement coefficient is converted into the weight and intensity of a convolution kernel (e.g., a high-pass filter kernel) for controlling a spatial convolution filter. Finally, all the processed image data is output to a clinical display in a standard medical video format (e.g., 1080p, 30fps) through a high-definition video interface (e.g., HDMI or SDI), thereby completing the entire process from optimization parameters to physical image enhancement.

[0127] Embodiment 2: The difference between Embodiment 2 and Embodiment 1 of the present application is that the embodiment introduces the clarity enhancement method used by the field of view clarity enhancement system in the gynecological clinical examination and diagnosis process.

[0128] The field of view clarity enhancement implementation method in the gynecological clinical examination and diagnosis process is based on the field of view clarity enhancement system in the gynecological clinical examination and diagnosis process described above, and includes the following steps:

[0129] S3.1, real-time acquisition of the field of view image in the gynecological examination process, and pre-processing of the field of view image;

[0130] S3.2, extracting key features from the pre-processed field of view image;

[0131] S3.3, generating a clarity score based on the key features, using a polarization difference image and a polarization average image to decouple and correct feature interference caused by mucus scattering and tissue reflection, generating an optimized clarity score, and generating clarity optimization parameters through a multi-task learning model;

[0132] S3.4, enhancing the clarity of the field of view image based on the clarity optimization parameters.

[0133] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A system for enhancing visual clarity during gynecological clinical examination and diagnosis, characterized in that, The application relates to a gynecological examination field of view enhancement method and device. The method comprises the following steps: An image acquisition and preprocessing unit (1) acquires a field of view image in a gynecological examination process in real time and pre-processes the field of view image; A feature extraction unit (2) extracts key features from the pre-processed field of view image; A definition evaluation and optimization unit (3) generates a definition score based on the key features, decouples and corrects feature interference caused by mucus scattering and tissue reflection by using a polarization difference image and a polarization average image, generates an optimized definition score, and generates definition optimization parameters by using a multi-task learning model; A field of view enhancement execution unit (4) enhances the definition of the field of view image based on the definition optimization parameters; The image acquisition module acquires the visual field image in the gynecological examination process based on the optical illumination module, and synchronously acquires two polarization images of the same visual field and ; The image acquisition and preprocessing unit (1) comprises an image acquisition module and an image processing module; The image processing module pre-processes the acquired field of view image; The feature extraction unit (2) comprises a brightness distribution module, a contrast module, an edge definition module and a blur degree module; The brightness distribution module is used for counting the overall brightness histogram of the pre-processed field of view image to obtain brightness distribution features reflecting the overall brightness of the image; The contrast module is used for analyzing the gray scale difference of a local region of the field of view image to extract contrast features reflecting the distinction degree of the tissue details of the field of view image; The edge definition module extracts edge definition features by using an edge detection method to represent the definition degree of the tissue profile of the field of view image; The blur degree module calculates blur degree features by using a frequency domain analysis method to reflect the blur region in the field of view image caused by inaccurate focusing; S1.1, based on two polarization images and calculating a polarization difference image and a polarization average image ; The definition evaluation and optimization unit (3) generates a definition score based on the key features, and the steps comprise the following steps: S1.2, the polarization characteristics of mucus and reflection are different, and the key features are decoupled and corrected based on scattering entropy detection of the mucus region and polarization difference detection of the reflection region, respectively, to generate decoupled and corrected key features, wherein the key features include brightness distribution features, contrast features, edge definition features and blur degree features; S1.3, the decoupled and corrected key features are normalized; S1.4, according to the importance of each decoupled and corrected key feature in image definition evaluation, a preset weight coefficient is allocated; S1.6, score of sub-features of key features are fused by weighting, and output comprehensive definition score , for representing overall definition level of current field image; S1.5, the sub-feature scores of the key features are calculated, the sub-feature scores of the key features include brightness balance degree scores, contrast scores, edge definition scores and blur scores, and reflect the image quality of each single feature dimension; S1.21, based on the polarized difference image of the luminance threshold segmentation extraction retro-reflective mask ; S1.22, polarization average image Compute local scattering entropy, regions with local scattering entropy values above a mucus threshold are marked as mucus mask ;​ S1.23, exclude specular mask in overall brightness histogram statistics and mucus masks regions, only use normal regions to compute baseline brightness distribution; S1.24, in the mucus mask Within the region, contrast compensation is performed using a scattering model to recover the original gray scale difference based on the polarization difference value; S1.25, in the reflective mask Region uses polarized image cross-channel edge information to repair edge breaks caused by overexposure.

2. The system for intensifying the field of vision during gynecological clinical examination, diagnosis and treatment process according to claim 1, characterized in that In the S1.2, the decoupled and corrected key features are generated, and the steps comprise the following steps: S2.1, receiving the decoupling-corrected key features output by the feature extraction unit (2), and synthesizing the intelligibility score ; The definition evaluation and optimization unit (3) generates definition optimization parameters by using a multi-task learning model, and the steps comprise the following steps: S2.2, a light-weight convolutional neural network is used to analyze the current field of view image in real time, the main anatomical structures are identified, and the identification results are used as a work flow state signal vector; S2.3, the feature vector of the key feature decoupled and corrected and the workflow state signal vector are jointly input into the multitask learning model as inputs; S2.4, based on the output of the multitask learning model, an optimization strategy weight vector and a basic parameter adjustment amount are generated; S2.5, the base parameter adjustment amount, the optimization strategy weight vector, and the sub-feature scores of each key feature are fused and calculated to obtain a weighted final parameter adjustment amount the final parameter adjustment amount which is the clarity optimization parameter.

3. The system for intensifying the field of vision during gynecological clinical examination, diagnosis and treatment process according to claim 2, characterized in that it comprises: In S2.2, the current field of view image is analyzed in real time by using a lightweight convolutional neural network to identify the main anatomical structures therein, including the following steps: S2.21, receiving the preprocessed field of view image output by the image acquisition and preprocessing unit (1), and standardizing the preprocessed field of view image; S2.22, inputting the standardized field of view image into the lightweight convolutional neural network; S2.23, extracting multi-scale feature maps through convolution and pooling operations of the lightweight convolutional neural network; S2.24, inputting the multi-scale feature maps into a classification module to identify the main anatomical structures; S2.25, the identification results are one-hot encoded to generate a workflow state signal vector.

4. The system for intensifying the field of vision during gynecological clinical examination, diagnosis and treatment process according to claim 3, characterized in that it comprises: In S2.4, the generating of the optimization strategy weight vector is specifically: the decoupled and corrected key feature vector and the workflow state signal vector are spliced in the channel dimension to form a fused feature vector, and the fused feature vector is input into a shared deep neural network encoder to extract a deep shared representation containing the current image quality condition and the current clinical task target ; and the deep shared representation is input into a clinical intention analysis decoder to generate the optimization strategy weight vector.

5. The gynecological clinical examination diagnosis and treatment process field of view definition enhancement system according to claim 4, characterized in that: The field of view enhancement execution unit (4) receives the sharpness optimization parameters generated by the sharpness evaluation and optimization unit (3), maps the sharpness optimization parameters to an optimization scheme, executes the optimization scheme to generate an enhanced final field of view image, and outputs the final field of view image to a clinical display interface.

6. A method for enhancing the field of view clarity in gynecological clinical examination and treatment process, based on the system for enhancing the field of view clarity in gynecological clinical examination and treatment process according to any one of claims 1-5, characterized in that, Including the following steps: S3.1, real-time acquisition of field of view images during gynecological examination, and preprocessing of the field of view images; S3.2, extracting key features from the preprocessed field of view images; S3.3, generating a sharpness score based on the key features, decoupling and correcting feature interference caused by mucus scattering and tissue reflection using polarized difference images and polarized average images, generating an optimized sharpness score, and generating sharpness optimization parameters through a multitask learning model; S3.4, enhancing the sharpness of the field of view image based on the sharpness optimization parameters.

Images