A method for constructing a colposcope intelligent auxiliary diagnosis decision system
By constructing an intelligent colposcopy-assisted diagnostic decision-making system, and utilizing multimodal data and deep learning technology, the system addresses the uneven level of cervical precancerous lesion identification in primary healthcare institutions. It achieves efficient and accurate cervical lesion diagnosis and recommends biopsy sites, thereby improving the diagnostic capabilities of primary healthcare institutions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-09
- Publication Date
- 2026-04-10
AI Technical Summary
Differences exist in the training and clinical practice of colposcopy professionals across different regions, resulting in uneven levels of cervical precancerous lesion identification among primary healthcare institutions. There is a need to develop a colposcopy-assisted diagnostic system that can assist in diagnosis and improve the identification rate of high-grade lesions.
A colposcopy-assisted intelligent diagnostic decision-making system was constructed. Data was collected from a multimodal cervical disease database, and data preprocessing and feature learning were performed. Feature fusion was carried out using convolutional neural networks and Transformer structures to optimize the network structure. The model was trained by combining clinical data and pathological features, and the probability distribution of cervical lesion types was output, and biopsy sites were recommended.
It has significantly improved the efficiency and accuracy of diagnosing precancerous cervical lesions, reduced reliance on doctors' experience, enhanced the diagnostic capabilities of primary hospitals, and improved the quality of medical services.
Smart Images
Figure CN121122665B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image recognition model algorithm, in particular to a construction method of a colposcope intelligent auxiliary diagnosis decision system. BACKGROUND
[0002] Colposcopy is an important tool for further examination of abnormal cervical cancer screening and an important indicator for risk assessment of cervical precancerous lesions and infiltration. Therefore, high-quality colposcopy is of great significance in the early detection and prevention of cervical cancer. However, due to the differences in social and economic development in different regions, leading to unbalanced distribution of medical resources, there are great differences in colposcope professional training and clinical practice in different regions, resulting in uneven overall diagnosis and treatment level, especially in primary medical institutions. Therefore, developing a colposcope auxiliary diagnosis system that can assist in diagnosis and improve the recognition rate of high-grade lesions has become a pressing problem. SUMMARY
[0003] The technical problem to be solved by the present application is to provide a construction method of a colposcope intelligent auxiliary diagnosis decision system in view of the shortcomings of the prior art.
[0004] The technical problem to be solved by the present application is solved by the following technical scheme, a construction method of a colposcope intelligent auxiliary diagnosis decision system, characterized by comprising the following steps, step S1, based on a multi-modal cervical disease database, a large amount of image data covering clinical medical record information, HPV detection, cytological pathology sections and different lesion degrees are collected, the image data is annotated by experienced doctors, the high-risk area of cervical high-grade lesions is accurately determined and the 3-year long-term HSIL+ disease risk is labeled, and then a target detection annotation data set is formed;
[0005] Step S2, data preprocessing is performed on the data obtained in step S1, including labeling clinical medical record information, cytological pathology section feature analysis and original colposcope image systematic cleaning and enhancement processing, and detecting and processing the preprocessed data;
[0006] Step S3, input the multi-modal data detected in step S2 into a feature learning network, the network realizes the complementary advantages between local space modeling and global context understanding by fusing convolutional neural network (CNN) and Transformer structure;
[0007] Step S4, the network structure is optimized for colposcopy image features. Through repeated training of a large amount of structured clinical data, HPV test results, cytological pathology specimen features and colposcopy picture annotation data, including model training through feature-level fusion strategy, the model learns the feature patterns of the lesion area and continuously adjusts the network parameters while calculating the long-term disease risk;
[0008] Step S5, the fused features are output through a fully connected layer and a Softmax classifier to output the probability distribution of seven types of cervical lesions, including high-grade squamous intraepithelial lesion (HSIL), low-grade squamous intraepithelial lesion (LSIL), chronic inflammation (Y), cervical cancer (CA), glandular epithelial metaplasia (LZ), polyp (P), and adenocarcinoma in situ (AIS).
[0009] Further, the step S1 includes the process of collecting more than 6000 colposcopy images and synchronously collecting iodine-stained images and acetic acid-stained images to construct a double-view image pair. The data collected in step S1 is multi-modal data, including clinical medical record information, HPV test results, cytological medical record section features and different degrees of lesion colposcopy images.
[0010] Further, the step S2 includes removing the mirror interference area of the image by using a Top-Hat filter image repair algorithm, enhancing the contrast of the image by using a histogram equalization method, and unifying the image size to 640x640 to adapt to the input requirements of the target detection network YOLOv7.
[0011] Further, the step S2 includes a lesion area detection stage, which uses YOLOv7 as the main target detection network and through end-to-end training, quickly and efficiently locates the suspicious areas in the cervical, cervical canal, vaginal and vulva images.
[0012] Further, the step S3 includes introducing a transfer learning strategy in the feature learning network, which initializes the network weights based on the pre-trained parameters on the ImageNet and COCO data sets, significantly shortens the model convergence time and improves the generalization ability of small sample classes.
[0013] Further, the step S3 includes using MobileNetv3_large as the CNN backbone network in the feature learning network to extract local spatial features, and introducing a Transformer module based on multi-head self-attention mechanism that can enhance the model's perception of the multi-scale and irregular morphology of the lesion area.
[0014] Further, the step S4 includes a feature level fusion strategy: the iodine staining image at the same position and the acetic acid staining image at different time are respectively input into two parallel branches, and after extracting features respectively, the two branches are fused through a connection layer, and further, the correlation and dependency of the tissue morphology under the two staining conditions are modeled through an attention module.
[0015] Further, the step S5 includes that based on the specific identified high-risk target area, the network system combines the morphology, size, spatial and temporal distribution information of the lesion area, and the clinical medical record information, HPV detection result, HPV infection duration and cytological pathological section characteristics, and generates a recommended biopsy site and labels the process of 3-year long-term risk of HSIL+ through a model algorithm.
[0016] Compared with the prior art, the present application has the following advantages:
[0017] (1) The method fully considers the spatial distribution characteristics of the lesion severity, preferentially recommends the area most likely to exist HSIL, thereby improving the targeting and diagnostic efficiency of biopsy, and through comparison of the recognition result and the pathological diagnosis result, the sensitivity, specificity and accuracy of the system in recognizing the HSIL high-risk target area are calculated, and the confidence reaches 80%;
[0018] (2) The system can accurately label the HSIL high-risk target area during the examination process, scientifically recommend the biopsy site, significantly improve the examination efficiency, effectively reduce the dependence of colposcopy examination on the experience of doctors, can alleviate the problem of insufficient ability of primary hospitals in recognizing cervical precancerous lesions, effectively improve the medical service quality and level, and provide strong technical support for early diagnosis and precise treatment of gynecological diseases;
[0019] (3) In the model training process, the transfer learning strategy is introduced, the network weight is initialized based on the pre-trained parameters on the ImageNet and COCO data sets, the model convergence time is significantly shortened, and the generalization ability on small sample categories is improved, which assists doctors in real-time recognition and positioning of cervical lesions in clinical examination, and improves the screening efficiency and accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0020] Figure 1 is the overall architecture layer structure diagram of the colposcope intelligent auxiliary diagnosis system of the present application;
[0021] Figure 2 is a data image schematic diagram of the acetic acid staining image branch at different time in the feature fusion strategy of the present application. DETAILED DESCRIPTION
[0022] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of the present application.
[0023] As shown in the accompanying drawings of Embodiment 1 of the present application, Figure 1 As shown in the accompanying drawings of Embodiment 1 of the present application, The present application discloses a construction method of a colposcope intelligent auxiliary diagnosis decision system. The method is based on a multi-modal cervical disease database, widely collects massive colposcope image data covering different lesion degrees, and is labeled by experienced colposcope doctors in our center, accurately delimits the high-risk area of HSIL, and constructs a high-quality labeled data set, laying a solid foundation for subsequent model training. The construction method comprises the following steps: S1, based on the multi-modal cervical disease database, clinical medical record information, HPV detection, cytological pathology sections and colposcope image data of different lesion degrees are widely collected, the image data is labeled by experienced doctors, the high-risk area of cervical high-grade lesion is accurately delimited, and 3-year long-term HSIL+ disease risk is labeled, and then a target detection labeled data set is formed; S2, data preprocessing is performed on the data obtained in step S1, including labeling clinical medical record information, cytological pathology section feature analysis and original colposcope image systematic cleaning and enhancement processing, and detecting the preprocessed data; S3, the multi-modal data detected in step S2 is input into a feature learning network, which realizes the complementary advantages between local space modeling and global context understanding by fusing convolutional neural network (CNN) and Transformer structure; S4, the network structure is optimized according to the colposcope image features, a large amount of structured clinical data, HPV detection results, cytological pathology specimen features and colposcope picture labeled data are repeatedly trained, including model training by feature level fusion strategy, while calculating the long-term disease risk, the model gradually learns the feature mode of the lesion area and continuously adjusts the network parameters; S5, the fused features are output through a full connection layer and a Softmax classifier to obtain the probability distribution of seven types of cervical lesions, including high-grade squamous intraepithelial lesion (HSIL), low-grade squamous intraepithelial lesion (LSIL), chronic inflammation (Y), cervical cancer (CA), glandular epithelial metaplasia (LZ), polyp (P) and adenocarcinoma in situ (AIS).
[0024] Further explanation is needed: (1) The data of step S1 is multi-modal data, which includes clinical medical record information, HPV test results, cytological medical record slice features, and colposcopy images of different degrees of lesions. (2) The model in step S3 is established by deep learning. The data in S1 is processed by step S2, and multi-modal deep learning is performed in an unsupervised learning manner. (3) After step S4 is trained by S3, the model can have the following characteristics: ① Precise positioning: In step S1, the colposcopy picture not only labels the lesions of the cervix, but also labels the lesions of the cervical canal, vaginal wall, vulva and other parts. Multi-modal deep learning makes the model not limited to colposcopy pictures. The model also adds clinical medical record information (such as age, marital history, and risk factors such as smoking), HPV test results, and cytological pathological features and other multiple influencing factors. The colposcopy pictures collected in step S1 include iodine-stained images and acetic acid-stained images at different time points, forming a multi-temporal dual-view image pair.
[0025] ② Label long-term disease risk: Through deep learning of the clinical medical record information, HPV test results, and pathological specimen characteristics in S1, the model can give the risk of patients developing HSIL+ lesions in the next 3 years.
[0026] Specifically, based on the multi-modal cervical disease database, we have collected a large number of colposcopy image data covering different degrees of lesions, totaling more than 6000 images, and simultaneously collected iodine-stained images and acetic acid-stained images to construct dual-view image pairs. All images were annotated by experienced colposcopy doctors with more than ten years of clinical experience, accurately delineating high-risk areas of cervical high-grade intraepithelial neoplasia (HSIL) and other lesions, forming a high-quality, multi-class target detection annotation dataset, laying a solid foundation for subsequent model training.
[0027] In the data processing stage, the original image is first preprocessed, including removing mirror reflection, enhancing image contrast, and uniform image size, etc. operations, aiming to improve image quality, reduce noise interference, and make the data more suitable for model training requirements.
[0028] In the data preprocessing stage, we systematically clean and enhance the original images, including: using image restoration algorithms such as Top-Hat filtering to remove specular reflection interference areas, using histogram equalization to enhance image contrast, and unifying image size to 640x640 to adapt to the input requirements of the target detection network YOLOv7. (This stage aims to improve image quality, reduce training interference caused by low-quality images, and improve the model's perception of abnormal areas.) In the lesion area detection stage, we use YOLOv7 as the main target detection network and through end-to-end training, we quickly and efficiently locate suspicious areas in cervical images. (YOLOv7 has excellent detection accuracy and real-time performance, with mAP@0.5 reaching 92.6% and recall rate reaching 90.1%, effectively locking HSIL, LSIL, and other lesion areas.)
[0029] Subsequently, we use a detection network in deep learning to preliminarily locate the areas in the image that may have abnormalities and lock potential lesion areas. On this basis, we use a convolutional neural network (CNN) as the core algorithm framework and optimize the network structure according to the characteristics of colposcopy images. Through repeated training of more than 6000 colposcopy image annotation data, the model gradually learns the feature patterns of lesion areas and continuously adjusts network parameters, thereby gradually acquiring the ability to accurately identify HSIL high-risk target areas. In the above process, we introduce a transfer learning strategy, initialize network weights based on pre-trained parameters on ImageNet and COCO datasets, thereby significantly shortening the model convergence time (reducing the training period by about 30%) and improving the generalization ability on small sample categories (such as AIS).
[0030] Specifically, the detected candidate regions are then input into our designed feature learning network. This network combines convolutional neural networks (CNN) and Transformer structures, achieving a complementary advantage between local spatial modeling and global context understanding. Further, we use MobileNetv3_large as the backbone network of CNN to extract local spatial features and introduce a Transformer module based on multi-head self-attention mechanism to enhance the model's perception of lesion areas with multiple scales and irregular shapes. At the same time, to fully utilize the information of double-view images, we design a feature-level fusion strategy: the iodine-stained image and the acetic acid-stained image at the same position are input into two parallel branches respectively, and after extracting features, they are fused through a concatenation layer, and then the attention module is used to model the correlation and dependency of tissue morphology under different staining conditions. (This dual-channel fusion network can effectively capture the differences in the same lesion area under different processing conditions and from 28x28 to 224x224 pixels, improving the model's recognition accuracy.)
[0031] Finally, the fused features are output through a fully connected layer and a Softmax classifier to obtain the probability distribution of seven types of cervical lesions, including high-grade squamous intraepithelial lesion (HSIL), low-grade squamous intraepithelial lesion (LSIL), chronic inflammation (Y), cervical cancer (CA), squamous metaplasia (LZ), polyp (P), and adenocarcinoma in situ (AIS). The classification confidence of the model on the validation set reaches 81%, showing high clinical practical value.
[0032] After training, the model is deployed to the colposcopy system. During the actual examination process, the system receives image data collected by the colposcope in real time, and uses the trained model to quickly analyze and process the image, accurately identify the high-risk area of HSIL, and visually mark the red area in the image for immediate reference by the examining doctor. Based on the identified high-risk target area, the system further combines the shape, size, and spatial distribution of the lesion area, and generates recommended biopsy sites through specific algorithms.
[0033] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A method for constructing a colposcopy intelligent auxiliary diagnosis decision system, characterized in that: Comprise the following steps, Step S1, based on the multi-modal cervical disease database, a large number of clinical information, HPV detection, cytological pathology sections and colposcopy image data of different lesion degrees are collected, and the image data is annotated by experienced doctors, the high-risk area of cervical high-grade lesion is accurately delineated, and the 3-year long-term HSIL+ disease risk is labeled, and then the target detection annotation dataset is formed; Step S2, data preprocessing is performed on the data obtained in step S1, including labeling clinical medical records, analyzing cytological pathology section features, and systematically cleaning and enhancing the original colposcopy image, and detecting the preprocessed data; Step S3, the multi-modal data detected in step S2 is input into the feature learning network, which realizes the complementary advantages between local space modeling and global context understanding by fusing convolutional neural network and Transformer structure, and step S3 further includes introducing a transfer learning strategy in the feature learning network, initializing the network weights based on the pre-trained parameters on ImageNet and COCO dataset, to significantly shorten the model convergence time and improve the generalization ability of small sample categories; Step S4, the network structure is optimized for colposcopy image features, through repeated training of a large amount of structured clinical data, HPV detection results, cytological pathology specimen features and colposcopy picture annotation data, including model training through feature-level fusion strategy, while calculating the long-term disease risk, the model gradually learns the feature mode of the lesion area and continuously adjusts the network parameters, step S4 further includes a feature-level fusion strategy: iodine-stained images and acetic acid-stained images at different times are input into two parallel branches respectively, and after extracting features respectively, they are fused through a connection layer, and further through an attention module to model the correlation and dependency of tissue morphology under two staining conditions; Step S5, the fused features are output through a fully connected layer and a Softmax classifier to obtain the probability distribution of seven types of cervical lesions, including high-grade squamous intraepithelial lesion, low-grade squamous intraepithelial lesion, chronic inflammation, cervical cancer, glandular epithelial metaplasia, polyp, and adenomatous intraepithelial tumor.
2. The method for constructing a colposcope intelligent auxiliary diagnostic decision system according to claim 1, characterized in that: The step S1 includes the process of collecting more than 6000 colposcopy images and synchronously collecting iodine-stained images and acetic acid-stained images to construct a double-view image pair, and the step S1 collects multi-modal data including clinical medical records, HPV detection results, cytological pathology section features and colposcopy images of different lesion degrees.
3. The method for constructing a colposcope intelligent auxiliary diagnostic decision system according to claim 1, characterized in that: The step S2 includes removing the image mirror interference area by using the Top-Hat filter image repair algorithm, enhancing the image contrast by using the histogram equalization method, and unifying the image size to 640x640 to adapt to the input requirements of the target detection network YOLOv7.
4. The method for constructing a colposcope intelligent auxiliary diagnostic decision system according to claim 3, characterized in that: The step S2 includes a lesion area detection stage, YOLOv7 is used as the main target detection network, and through end-to-end training, the suspicious areas in the cervical, cervical canal, vagina and vulva images are quickly and efficiently located.
5. The method for constructing a colposcope intelligent auxiliary diagnostic decision system according to claim 1, characterized in that: The step S3 comprises adopting MobileNetv3_large as the backbone network of CNN in the feature learning network to extract local spatial features, and introducing a Transformer module based on a multi-head self-attention mechanism which can enhance the perception ability of the model to the multi-scale and irregular morphology of the lesion area.
6. The method for constructing a colposcope intelligent auxiliary diagnostic decision system according to claim 1, characterized in that: The step S5 comprises that, based on the specific identified high-risk target area, the network system combines the morphology, size, spatial and temporal distribution information of the lesion area, and the clinical medical record information, HPV detection result, HPV infection duration and cytological pathological section characteristics, and generates recommended biopsy sites and labels the process of 3-year far-term risk of HSIL+ through model algorithm.
Citation Information
Patent Citations
Multi-modal cervical cancer anterior lesion region segmentation method based on Swin Transform
CN119107450A
Image segmentation system via graph or multiscale cascaded attention decoding
US20250139775A1