Intelligent cervical cancer screening method based on Unet3 +-Repvg model

Abstract

The invention relates to an intelligent cervical cancer screening method based on a Unet3 +-Repvang model, which comprises the following steps: collecting colposcope image data, and marking disease positions and categories of images; performing fine tuning and segmentation training by using the pre-trained Unet3 network, and segmenting the image; restoring the segmented binary image corresponding to the original image; performing classification training by utilizing a Repvg model to obtain a symptom to which the colposcope belongs; and extracting features of a newly shot colposcope image through the segmentation model, and finally obtaining final disease prediction through the classification model. According to the method, a reasonable judgment basis is provided for helping a doctor to accurately judge the cervical symptom, relieving the pressure of the doctor and judging whether the patient needs to be further treated and diagnosed or not.

Description

Technical Field

[0001] The present invention relates to the technical field of computer image processing, and in particular to an intelligent screening method for cervical cancer based on a Unet3+-Repvgg model. Background Art

[0002] In the field of intelligent cervical cancer screening, traditional screening methods mainly rely on cervical cytology (Papsmear), high-risk human papillomavirus (HPV) testing and colposcopy. Although these methods have improved the detection rate of cervical cancer to a certain extent, they still have some limitations.

[0003] 1. Pap smear: This is a method of examining cervical cells under a microscope to detect abnormal cells. However, this method has high requirements for sample collection and processing, and is highly dependent on the operator's technical level, which may lead to missed or misdiagnosis.

[0004] 2. High-risk HPV testing: HPV is the main cause of cervical cancer. The risk of cervical cancer can be predicted by testing high-risk HPV. However, HPV testing cannot directly diagnose cervical cancer, and some HPV-infected people may not develop cervical cancer.

[0005] 3. Colposcopy: This is a method of magnifying the surface of the cervix to observe its shape and color changes. Colposcopy can detect tiny lesions that are difficult to detect with the naked eye, but it requires high experience and judgment from the operator and is labor-intensive.

[0006] With the development of artificial intelligence technology, more and more researchers have begun to explore the use of deep learning models to assist in the screening and diagnosis of cervical cancer. Deep learning models can learn the characteristics of cervical cancer by analyzing a large amount of colposcopy image data, thereby improving the accuracy and efficiency of screening. However, existing deep learning models still face some challenges when processing colposcopy images: 1. Image quality differences: The quality and clarity of colposcopic images are affected by many factors, such as shooting angle, lighting conditions, etc., which may lead to unstable performance of the model in practical applications.

[0007] 2. Difficulty in data labeling: Labeling of colposcopy images requires professional medical knowledge, and the labeling process is time-consuming and labor-intensive, which limits the acquisition of large-scale training data.

[0008] 3. Model generalization ability: Existing models often perform well on specific data sets, but may perform poorly on new, unknown data, which limits the generalization ability of the model. Summary of the invention

[0009] The purpose of the present invention is to overcome the problems existing in the prior art and provide a method for intelligent screening of cervical cancer based on the Unet3+-Repvgg model, which realizes efficient processing and accurate classification of colposcopy images by combining the Unet3+ image segmentation model and the Repvgg classification model. The Unet3+ model can accurately segment the key feature areas in the image, while the Repvgg model can accurately classify these feature areas. In addition, the present invention also adopts data enhancement and preprocessing technology to improve the adaptability of the model to different image changes, and improves the generalization ability of the model by optimizing model training and parameter adjustment. It provides a new technical means for the intelligent screening of cervical cancer, which is expected to significantly improve the accuracy and efficiency of screening.

[0010] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions: An intelligent screening method for cervical cancer based on the Unet3+-Repvgg model, the method comprising the following steps: Step S1: While performing colposcopy detection, collect colposcopy images, use closed curves to mark the sampling location area, and after obtaining the detection results, correspond the detection results to the original image and mark the disease category; Step S2: Based on the disease annotation position of the closed-loop curve, the image is segmented using the pre-trained Unet3+ image segmentation network; Step S3: segment the colposcopy image using the trained segmentation model, and restore the segmented binary image to the original image, retaining only the extracted important feature areas; Step S4: For the remaining feature areas, the Repvgg network model is used to perform disease classification training to classify the colposcopy symptoms; Step S5: For the newly captured colposcopy image, the trained segmentation model is used to divide the feature area, and the trained classification model is used to classify the image, and finally the classification prediction result of the new image is obtained.

[0011] Furthermore, in step S1, the specific steps of colposcopy data collection and processing are as follows: Step S1.1: While performing the colposcopy test, colposcopy images are collected. These images include normal images and images of different diseases. The data comes from several university-affiliated hospitals specializing in cervical cancer early screening to ensure the representativeness and scientificity of the data; Step S1.2: While testing, use a closed curve to mark the sampling location area. After obtaining the test results, correspond the test results to the original image, and ask a doctor of corresponding specialty to mark the disease category based on the test results.

[0012] Furthermore, in step S2, the specific steps of performing segmentation training on the image are as follows: Step S2.1: Download the open-source cervical cancer segmentation image and perform pre-training operations on it on the Unet3+ model. The Unet3+ model is used to capture the context information and detail features in the image. During the segmentation process, the Unet3+ model separates the region of interest in the image from the background and generates a clear binary image. Step S2.2: Fine-tune the pre-trained model, and train the fine-tuned model on the colposcopy dataset.

[0013] Furthermore, in step S3, the specific steps of restoring the binary image to the original image are as follows: Step S3.1: After the binary image is obtained by segmentation, the binary image is matched to the corresponding original colposcopy image, and the black area in the corresponding binary image, that is, the irrelevant noise area in the colposcopy image, is removed; The formula corresponding to the original image is: , In the formula, It is the pixel value of a point in the colposcopy image, and the numerical value is expressed using RGB. Represents the value of the pixel of the binary image, white is (255,255,255), black is (0,0,0). If the corresponding position of the binary image is a black area, the corresponding position of the original image is fitted to white. If the corresponding position of the binary image is a white area, the image of the corresponding position of the original image remains unchanged.

[0014] Furthermore, in step S4, the specific steps of disease classification training are as follows: Step S4.1: When performing classification training on the fused image, the image is rotated, scaled, and color adjusted to simulate different observation conditions, so as to enhance the generalization ability of the model and cope with the image variability in practical applications, and improve the adaptability of the model to new samples; Step S4.2: In the preprocessing stage, ensure that all images are standardized to the same size and color space to facilitate model processing. In addition, image quality enhancement techniques are used to improve the clarity and readability of the images. Step S4.3: During the disease classification training process, the model is first initialized; then supervised learning is performed using a labeled colposcopy image dataset; through the back-propagation algorithm, the model learns the ability to map image features to specific disease categories; in the classification detection stage, the Repvgg model analyzes the extracted feature areas and learns their associations with different disease categories; the model uses multiple layers of convolution and fully connected layers to ultimately output classification prediction results for new images; these results include the disease category to which the image may belong, as well as the confidence level of the prediction, providing doctors with more comprehensive diagnostic information.

[0015] Furthermore, in step S5, the specific steps of predicting the new image are as follows: Step S5.1: For a new colposcopy image, use the trained Unet3+ segmentation model to segment the image and obtain feature regions; Step S5.2: The segmented image enters the trained Repvgg model for classification detection, and finally obtains the classification prediction result of the new image.

[0016] The beneficial effects of the present invention are: The method of the present invention significantly improves the accuracy of diagnosis through deep learning technology, and reduces misdiagnosis and missed diagnosis caused by human factors. It greatly improves the screening efficiency through automated image processing and classification processes, reduces the workload of doctors, and enables them to devote more time and energy to the diagnosis and treatment of complex cases. In addition, the method optimizes the allocation of medical resources, especially in areas with limited resources, through fast and accurate screening results, it reduces the waiting time of patients and improves patients' satisfaction with medical services. It also provides doctors with data-driven decision support, helping them to formulate more accurate treatment plans, while promoting the development of scientific research and education. The application scope of this technology is not limited to cervical cancer screening, but can also be extended to other types of cancer screening and medical image analysis, providing patients with personalized diagnosis and treatment recommendations, and promoting the development of personalized medicine. In general, the present invention provides strong technical support for the early detection, diagnosis and treatment of cervical cancer, and is expected to significantly improve the survival rate and treatment effect of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is an overall framework diagram for the implementation of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0019] Reference Figure 1 As shown, a cervical cancer intelligent screening method based on the Unet3+-Repvgg model comprises the following steps: Step S1: While performing colposcopy detection, collect colposcopy images, use closed curves to mark the sampling location area, and after obtaining the detection results, correspond the detection results to the original image and mark the disease category; Step S1.1: While performing the colposcopy test, colposcopy images are collected. These images include normal images and images of different diseases. The data comes from several university-affiliated hospitals specializing in cervical cancer early screening to ensure the representativeness and scientificity of the data; Step S1.2: While testing, use a closed curve to mark the sampling location area. After obtaining the test results, correspond the test results to the original image, and ask doctors of corresponding specialties to mark the disease categories according to the test results; Step S2: Based on the disease annotation position of the closed-loop curve, the image is segmented using the pre-trained Unet3+ image segmentation network; Step S2.1: Download the open-source cervical cancer segmentation image and perform pre-training on it on the Unet3+ model. The pre-trained model has more perceptual information. The Unet3+ model can effectively capture the contextual information and detail features in the image through its unique encoder-decoder structure. During the segmentation process, the Unet3+ model separates the region of interest (such as potential lesion area) in the image from the background to generate a clear binary image. Step S2.2: fine-tune the pre-trained model, and train the fine-tuned model on the colposcopy dataset; Step S3: segment the colposcopy image using the trained segmentation model, and restore the segmented binary image to the original image, retaining only the extracted important feature areas; Step S3.1: After the binary image is obtained by segmentation, the binary image is matched to the corresponding original colposcopy image, and the black area in the corresponding binary image, that is, the irrelevant noise area in the colposcopy image, is removed; The formula corresponding to the original image is: , In the formula, It is the pixel value of a point in the colposcopy image, and the numerical value is expressed using RGB. Represents the value of the pixel of the binary image. White is (255,255,255) and black is (0,0,0). If the corresponding position of the binary image is a black area, the corresponding position of the original image is fitted to white. If the corresponding position of the binary image is a white area, the image of the corresponding position of the original image remains unchanged. Step S4: For the remaining feature areas, the Repvgg network model is used to perform disease classification training to classify the colposcopy symptoms; Step S4.1: When training the fused images for classification, rotate, scale, and adjust the colors of the images to simulate different viewing conditions, in order to enhance the generalization ability of the model and cope with the image variability in practical applications. These techniques not only increase the diversity of training data, but also improve the adaptability of the model to new samples. Step S4.2: In the preprocessing stage, ensure that all images are normalized to the same size and color space for easy model processing. In addition, image quality enhancement techniques such as denoising and sharpening are used to improve the clarity and readability of the images. Step S4.3: During the disease classification training process, the model is first initialized; then supervised learning is performed using the labeled colposcopy image dataset; the model learns to map image features to specific disease categories through the back propagation algorithm; in the classification detection stage, the Repvgg model analyzes the extracted feature regions and learns their associations with different disease categories; the model outputs the classification prediction results for the new image through multiple layers of convolution and fully connected layers; these results include the disease category to which the image may belong and the confidence of the prediction, providing doctors with more comprehensive diagnostic information. This method can provide doctors with fast and accurate auxiliary diagnosis, thereby improving the efficiency and accuracy of cervical cancer screening; Step S5: for the newly captured colposcopy image, the trained segmentation model is used to divide the feature area, and the trained classification model is used to classify the image, and finally the classification prediction result of the new image is obtained; Step S5.1: For a new colposcopy image, use the trained Unet3+ segmentation model to segment the image and obtain feature regions; Step S5.2: The segmented image is subjected to classification detection using the trained Repvgg model, and finally the classification prediction result of the new image is obtained; Through these two steps, not only the synergistic effect of Unet3+ and Repvgg models in intelligent cervical cancer screening is combined, but also their high efficiency and accuracy in processing new images are emphasized. This strategy of combining image segmentation and classification provides strong technical support for the early diagnosis and treatment of cervical cancer.

[0020] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent screening method for cervical cancer based on the Unet3+-Repvgg model, characterized in that: The method comprises the following steps: Step S1: While performing colposcopy detection, collect colposcopy images, use closed curves to mark the sampling location area, and after obtaining the detection results, correspond the detection results to the original image and mark the disease category; Step S2: Based on the disease annotation position of the closed-loop curve, the image is segmented using the pre-trained Unet3+ image segmentation network; Step S3: segment the colposcopy image using the trained segmentation model, and restore the segmented binary image to the original image, retaining only the extracted important feature areas; Step S4: For the remaining feature areas, the Repvgg network model is used to perform disease classification training to classify the colposcopy symptoms; Step S5: For the newly captured colposcopy image, the trained segmentation model is used to divide the feature area, and the trained classification model is used to classify the image, and finally the classification prediction result of the new image is obtained.

2. The intelligent screening method for cervical cancer based on the Unet3+-Repvgg model according to claim 1, characterized in that: In step S1, the specific steps of colposcopy data collection and processing are as follows: Step S1.1: The colposcopy images collected include normal images and images containing different diseases. The data comes from several university-affiliated hospitals specializing in cervical cancer early screening to ensure the representativeness and scientificity of the data; Step S1.2: While testing, use a closed curve to mark the sampling location area. After obtaining the test results, correspond the test results to the original image, and ask a doctor of corresponding specialty to mark the disease category based on the test results.

3. The intelligent screening method for cervical cancer based on the Unet3+-Repvgg model according to claim 1, characterized in that: In step S2, the specific steps of performing segmentation training on the image are as follows: Step S2.1: Download the open-source cervical cancer segmentation image and perform pre-training operations on it on the Unet3+ model. The Unet3+ model is used to capture the context information and detail features in the image. During the segmentation process, the Unet3+ model separates the region of interest in the image from the background and generates a clear binary image. Step S2.2: Fine-tune the pre-trained model, and train the fine-tuned model on the colposcopy dataset.

4. The intelligent screening method for cervical cancer based on the Unet3+-Repvgg model according to claim 1, characterized in that: In step S3, the specific steps of restoring the binary image to the original image are as follows: Step S3.1: After the binary image is obtained by segmentation, the binary image is matched to the corresponding original colposcopy image, and the black area in the corresponding binary image, that is, the irrelevant noise area in the colposcopy image, is removed; The formula corresponding to the original image is: ， In the formula, is the pixel value of a point in the colposcopy image, and the numerical value is represented by RGB. Represents the value of the pixel of the binary image, white is (255,255,255), black is (0,0,0). If the corresponding position of the binary image is a black area, the corresponding position of the original image is fitted to white. If the corresponding position of the binary image is a white area, the image of the corresponding position of the original image remains unchanged.

5. The intelligent screening method for cervical cancer based on the Unet3+-Repvgg model according to claim 1, characterized in that: In step S4, the specific steps of disease classification training are as follows: Step S4.1: When performing classification training on the fused image, the image is rotated, scaled, and color adjusted to simulate different observation conditions, so as to enhance the generalization ability of the model and cope with the image variability in practical applications, and improve the adaptability of the model to new samples; Step S4.2: In the preprocessing stage, ensure that all images are standardized to the same size and color space to facilitate model processing. In addition, image quality enhancement techniques are used to improve the clarity and readability of the images. Step S4.3: During the disease classification training process, the model is first initialized; then supervised learning is performed using the labeled colposcopy image dataset; Through the back-propagation algorithm, the model learns the ability to map image features to specific disease categories. In the classification detection stage, the Repvgg model analyzes the extracted feature regions and learns their associations with different disease categories. The model goes through multiple layers of convolution and fully connected layers, and ultimately outputs classification prediction results for new images; these results include the disease category to which the image may belong, as well as the confidence level of the prediction, providing doctors with more comprehensive diagnostic information.

6. The intelligent screening method for cervical cancer based on the Unet3+-Repvgg model according to claim 1, characterized in that: In step S5, the specific steps of predicting the new image are as follows: Step S5.1: For a new colposcopy image, use the trained Unet3+ segmentation model to segment the image and obtain feature regions; Step S5.2: The segmented image enters the trained Repvgg model for classification detection, and finally obtains the classification prediction result of the new image.

Images