Intelligent keratoconjunctival tumor identification method based on self-supervised learning
By establishing an intelligent identification model for corneal and conjunctival tumors through self-supervised learning, the problem of insufficient equipment in underdeveloped areas has been solved, achieving efficient identification and assisted diagnosis of corneal and conjunctival tumors, improving diagnostic accuracy, and reducing inconvenience and costs for patients.
Patent Information
- Application Number
- CN202511013965.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-23
- Publication Date
- 2025-11-21
AI Technical Summary
In existing technologies, intelligent identification methods for corneal and conjunctival tumors are highly dependent on doctors' experience and specialized equipment, making it difficult to achieve timely diagnosis and treatment in underdeveloped areas, and increasing patient inconvenience and costs during the tissue biopsy process.
A self-supervised learning method was adopted to establish an intelligent identification model for corneal and conjunctival tumors by acquiring patients' historical medical data and corneal images. The model was trained and identified using a self-supervised learning network, and risk assessment values were calculated by combining clinical digital features and image features to assist in diagnosis.
It improves the accuracy of corneal and conjunctival tumor identification, reduces reliance on physician experience, compensates for equipment deficiencies, reduces unnecessary biopsies, improves diagnostic efficiency in underdeveloped areas, and alleviates the burden on patients.
Smart Images
Figure CN120998486A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent identification technology for corneal and conjunctival tumors, and specifically to an intelligent identification method for corneal and conjunctival tumors based on self-supervised learning. Background Technology
[0002] Corneal and conjunctival tumors are uncommon eye tumors, typically referring to tumors that occur in the corneal and conjunctival regions; these tumors can include both benign and malignant tumors, and their symptoms, treatments, and prognoses may vary depending on the specific type and nature of the tumor.
[0003] In existing technologies, ophthalmologists make preliminary diagnoses of corneal and conjunctival tumors based on medical history and slit-lamp microscopy. If the nature of the tumor is uncertain, a histopathological biopsy is required to determine the final tumor type (malignant, benign, or precancerous). Currently, there is no intelligent method for identifying corneal and conjunctival tumor types. However, for intelligent identification methods of other ocular surface tumors (such as eyelid tumors and melanoma), artificial intelligence (AI) technology has been successfully applied. Typically, ImageNet datasets are used for pre-training, and then fine-tuning (retraining) is performed using eye images to improve recognition accuracy.
[0004] However, existing intelligent identification methods rely heavily on doctors' experience and professional skills, as well as the availability of specialized equipment such as slit-lamp microscopes in medical facilities. In many regions, especially underdeveloped areas, there is a severe shortage of professional ophthalmologists. This not only affects the timely diagnosis and treatment of various eye diseases, including corneal and conjunctival tumors, but also exacerbates patients' suffering and financial burden. In addition to human resources, medical institutions in these regions also face the dilemma of lacking key diagnostic and treatment equipment such as slit-lamp microscopes. Furthermore, due to inadequate infrastructure, even if a tumor is suspected, it is difficult to perform more in-depth examinations such as tissue biopsies. Moreover, when the specific nature of a tumor cannot be determined through preliminary examination, further histopathological biopsy is usually required to clarify the diagnosis. This process not only increases the inconvenience and medical expenses for patients, but may also cause anxiety while waiting for the results. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent identification method for corneal and conjunctival tumors based on self-supervised learning, and to solve the following technical problems.
[0006] The objective of this invention can be achieved through the following technical solutions: A self-supervised learning-based intelligent identification method for corneal and conjunctival tumors includes the following steps: Step S1: Obtain the patient's historical medical data, which includes the patient's historical corneal images and patient characteristics. The patient characteristics are several feature indicators, including eye medical history, age, and gender. Step S2: Based on the patient's characteristics, obtain several clinical numerical features; obtain the total number of all patients in the historical disease data, and obtain the clinical weight of the clinical numerical feature based on the proportion of the number of patients corresponding to the clinical numerical feature in the total number; Step S3: Based on a self-supervised learning network, establish a corneal and conjunctival tumor intelligent identification model, and according to the corneal and conjunctival tumor intelligent identification model, obtain the proportion of corneal and conjunctival tumor regions in each historical corneal image, obtain several disease stages, and determine the corneal and conjunctival tumor image features and disease weights of each disease stage. Step S4: Obtain the current corneal image of the current patient, determine the target clinical digital features and clinical weights of the current patient, determine the target disease stage and disease weight of the current corneal image, and obtain the risk assessment value of the current patient.
[0007] As a further aspect of the present invention: the process of obtaining the clinical digital features includes: The index values of the feature indicators are obtained. The index values of the eye disease history include keratitis, pterygium surgery history, family history of tumors, and long-term ultraviolet radiation exposure history. The index values of the gender include male and female. The index values of the age include several preset age ranges. The index values of the feature indicators are assigned based on binary encoding. The assigned index values are concatenated in a preset order to obtain the clinical digital features of the feature indicators.
[0008] As a further aspect of the present invention: the clinical weight Cw of the i-th clinical numerical feature i =n i / N, where n i Let N be the number of patients in the historical disease data whose numerical feature is the i-th clinical numerical feature, and let N be the total number of all patients in the historical disease data.
[0009] As a further aspect of the present invention: the process of establishing the intelligent identification model for corneal and conjunctival tumors includes: Unlabeled corneal images are acquired, and all unlabeled corneal images are cleaned to obtain an unlabeled dataset. Corneal images of patients with corneal and conjunctival tumors are obtained from medical institutions, and the pathological results of the patients are obtained. The corneal images are filtered and labeled according to the pathological results to obtain a labeled dataset. A self-supervised learning network is established, and the unlabeled corneal images in the unlabeled dataset are divided into several image patches, and all image patches are further divided into occluded image patches and visible image patches; the self-supervised learning network is trained based on the occluded image patches and visible image patches to obtain a general model of the ocular surface. A recognition network is constructed based on the general model of the ocular surface. All corneal images in the labeled dataset are divided into a training set, a validation set, and a test set. The recognition network is trained using the training set, validation set, and test set to obtain the final intelligent recognition model for corneal and conjunctival tumors.
[0010] As a further aspect of the present invention, the process of data cleaning for all unlabeled corneal images includes: A slit-lamp image quality monitoring model was established. The model was used to analyze the quality of all unlabeled corneal images, classifying them into low-quality and high-quality images. All low-quality images were filtered out, and all high-quality images were retained. The high-quality images were then adjusted to a preset standard size and format using cubic difference to obtain the unlabeled dataset.
[0011] As a further aspect of the present invention: the process of establishing the slit lamp image quality monitoring model includes: 12,000 slit-lamp corneal images were selected, which are corneal images of patients obtained through a slit-lamp microscope. Ophthalmologists labeled each slit-lamp corneal image, classifying all images into high-quality and low-quality images, and deleting all low-quality images. The pixel values of all remaining slit-lamp corneal images were normalized to between 0 and 1, and the size of all slit-lamp corneal images was adjusted to 224×224 pixels to obtain the image dataset. An image quality monitoring model was constructed using the ResNet50 deep learning algorithm. An adaptive moment estimation optimizer was used to optimize the image quality monitoring model. The image dataset was used to iteratively train the image quality monitoring model for 180 rounds, and finally the slit lamp image quality monitoring model was obtained.
[0012] As a further aspect of the present invention: the process for determining the image features and disease weights of corneal and conjunctival tumors at the disease stage includes: The intelligent corneal and conjunctival tumor recognition model identifies corneal and conjunctival tumor regions in the historical corneal images. The corneal and conjunctival tumor region is the area occupied by the corneal and conjunctival tumor in the historical corneal images. The area ratio of the corneal and conjunctival tumor region in the historical corneal images is obtained and denoted as the corneal and conjunctival tumor region ratio. Based on the proportion of corneal and conjunctival tumor regions in each historical corneal image, several region proportion intervals are set, and the average value of the corneal and conjunctival tumor region proportions at both ends of the region proportion interval is obtained. The region proportion intervals are sorted from smallest to largest according to the size of the average value and numbered. Each region proportion interval is recorded as a disease stage, and the number corresponding to the region proportion interval is recorded as the disease weight of the region proportion interval. Based on the intelligent identification model for corneal and conjunctival tumors, image features of all historical corneal images within the disease stage are obtained and denoted as the corneal and conjunctival tumor image features of the disease stage.
[0013] As a further aspect of the present invention: the process of obtaining the current patient's risk assessment value includes: The patient characteristics of the current patient are obtained and denoted as current patient characteristics. Based on the patient characteristics, the clinical numerical characteristics of the current patient are obtained and denoted as current clinical numerical characteristics. The numerical characteristic similarity between the current clinical numerical characteristics and each clinical numerical characteristic is obtained. The clinical numerical characteristic with the highest numerical characteristic similarity to the current clinical numerical characteristic is selected and denoted as the target clinical numerical characteristic of the current patient. The clinical weight corresponding to the target clinical numerical characteristic is obtained and denoted as the clinical weight of the current patient. The image features of the current corneal image are obtained and denoted as the current image features. The image feature similarity between the current image features and the image features of each corneal and conjunctival tumor image is obtained. The corneal and conjunctival tumor image feature with the highest image feature similarity to the current image features is selected and denoted as the target corneal and conjunctival tumor image feature. The disease stage of the target corneal and conjunctival tumor image feature is obtained and denoted as the target disease stage of the current corneal image. The disease weight corresponding to the target disease stage is obtained and denoted as the disease weight of the current corneal image. The current patient's risk assessment value Rav = K × (Cw × Sim) Df +Q×Dw×Sim If ), where K is the preset first correction coefficient and K > 0, Sim Df The numerical feature similarity between the current clinical numerical features and the target clinical numerical features is given by Cw, where Cw is the clinical weight of the current patient, and Sim is the numerical feature similarity. If The image feature similarity between the current image features and the target corneal and conjunctival tumor image features is given by Dw, where Dw is the disease weight of the current corneal image, and Q is a preset second correction coefficient, where Q > 1.
[0014] The beneficial effects of this invention are: This method can identify dozens of corneal and conjunctival tumors, including carcinoma in situ, conjunctival lymphoma, squamous cell carcinoma, melanoma, actinic keratosis, conjunctival intraepithelial neoplasia, primary acquired atypia melanosis, nevus, conjunctival cyst, squamous papilloma, and dermoid tumor, covering a wide range. In internal and external test sets, the constructed corneal and conjunctival recognition model showed a receiver operating characteristic (AUC) exceeding 0.89 for each tumor state (benign, malignant, and precancerous), demonstrating high recognition accuracy.
[0015] Existing technologies heavily rely on doctors' experience and expertise, while this invention, through an intelligent recognition model, can assist in diagnosis to some extent, reducing over-reliance on doctors' experience. Many regions, especially underdeveloped areas, lack key diagnostic equipment such as slit-lamp microscopes. The method of this invention can, to some extent, compensate for this equipment shortage, facilitating timely diagnosis of corneal and conjunctival tumors. When the nature of a tumor cannot be determined through preliminary examination, a histopathological biopsy is usually required, increasing patient inconvenience and medical costs, and potentially causing anxiety. The intelligent recognition method of this invention can improve the accuracy of preliminary diagnosis, thereby reducing unnecessary biopsies. In areas with a shortage of professional ophthalmologists, this invention can assist in the rapid identification of corneal and conjunctival tumors, avoiding diagnostic delays caused by doctor shortages, helping patients receive timely treatment, and alleviating suffering and financial burden. Using a large amount of unlabeled ocular surface data to construct a pre-trained ocular surface model, compared to models trained on the ImageNet dataset, the AUC for each category is improved by 0.03-0.14 (3%–14%) on the external test set, further enhancing the recognition effect of corneal and conjunctival tumors. Attached Figure Description
[0016] The invention will now be further described with reference to the accompanying drawings.
[0017] Figure 1 This is a schematic diagram illustrating the steps of a self-supervised learning-based intelligent identification method for corneal and conjunctival tumors according to the present invention. Figure 2 This is a schematic diagram illustrating the process of establishing an intelligent identification model for corneal and conjunctival tumors in a self-supervised learning-based intelligent identification method for corneal and conjunctival tumors according to the present invention. Figure 3 This is a schematic diagram of the self-supervised learning network structure of the intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to the present invention. Figure 4 This is a schematic diagram of the intelligent identification model structure for corneal and conjunctival tumors based on self-supervised learning, according to the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Please see Figure 1 As shown, this invention is an intelligent identification method for corneal and conjunctival tumors based on self-supervised learning, comprising the following steps: Step S1: Obtain the patient's historical medical data, which includes the patient's historical corneal images and patient characteristics. The patient characteristics are several feature indicators, including eye medical history, age, and gender. Historical medical data of patients with corneal and conjunctival diseases who have previously visited medical institutions are collected through electronic medical record systems, imaging databases, or clinical research data platforms. This data covers patients from different regions, age groups, and medical histories to ensure sample diversity. The device acquires corneal images of patients using a slit-lamp microscope, providing high-resolution images of ocular surface tissue structures to ensure that details of tumor areas are accurately captured; image types include color images and images at specific wavelengths (such as those after fluorescein staining) to cover imaging data in different diagnostic scenarios; Record the patient's history of keratitis, pterygium surgery, family history of cancer (e.g., immediate family members with eye or other tumors), and history of long-term ultraviolet radiation exposure (e.g., long-term outdoor work or lifestyle habits). Record each indicator as "yes / no" or a specific number of years (e.g., ultraviolet radiation exposure ≥ 5 years). Divide the patient's age into preset ranges (e.g., 0-18 years, 19-40 years, 41-60 years, over 60 years, etc.) to facilitate subsequent analysis of disease distribution characteristics by age group. Step S2: Based on the patient's characteristics, obtain several clinical numerical features; obtain the total number of all patients in the historical disease data, and obtain the clinical weight of the clinical numerical feature based on the proportion of the number of patients corresponding to the clinical numerical feature in the total number; In a preferred embodiment of the present invention, the process of obtaining the clinical digital features includes: The index values of the feature indicators are obtained. The index values of the eye disease history include keratitis, pterygium surgery history, family history of tumors, and long-term ultraviolet radiation exposure history. The index values of the gender include male and female. The index values of the age include several preset age ranges. The index values of the feature indicators are assigned based on binary encoding. The assigned index values are concatenated in a preset order to obtain the clinical digital features of the feature indicators. The process of assigning values to each indicator of the feature index based on binary encoding includes: Each indicator value is encoded with an independent binary bit, with "yes" or the corresponding interval recorded as "1" and "no" or the non-corresponding interval recorded as "0". If there are 4 indicators, namely keratitis, history of pterygium surgery, family history of cancer, and history of ultraviolet exposure, then the 4 indicators will each occupy 4 binary bits, in the order of [keratitis, history of pterygium surgery, family history of cancer, history of ultraviolet exposure]. For example, if a patient has keratitis and a history of ultraviolet exposure, then the code is "1001". In a preferred embodiment of the present invention, the clinical weight Cw of the i-th clinical numerical feature is... i =n i / N, where n i Let N be the number of patients in the historical disease data whose numerical feature is the i-th clinical numerical feature, and let N be the total number of all patients in the historical disease data. Clinical weights reflect the frequency of occurrence of the feature combination in historical patients. The higher the frequency (the greater the weight), the more statistically significant the association between the feature combination and corneal and conjunctival tumors. Step S3: Based on a self-supervised learning network, establish a corneal and conjunctival tumor intelligent identification model, and according to the corneal and conjunctival tumor intelligent identification model, obtain the proportion of corneal and conjunctival tumor regions in each historical corneal image, obtain several disease stages, and determine the corneal and conjunctival tumor image features and disease weights of each disease stage. In a preferred embodiment of the present invention, the process of establishing the intelligent identification model for corneal and conjunctival tumors includes: S1: Obtain unlabeled corneal images, perform data cleaning on all unlabeled corneal images to obtain an unlabeled dataset; obtain corneal images of patients with corneal and conjunctival tumors through medical institutions, and obtain the pathological results of the patients with corneal and conjunctival tumors; filter and label the corneal images according to the pathological results to obtain a labeled dataset. S2: Establish a self-supervised learning network, divide the unlabeled corneal images in the unlabeled dataset into several image blocks, and divide all image blocks into occluded image blocks and visible image blocks; train the self-supervised learning network based on the occluded image blocks and visible image blocks to obtain a general model of the ocular surface; S3: Construct a recognition network based on the general ocular surface model, and divide all corneal images in the labeled dataset into a training set, a validation set, and a test set; train the recognition network using the training set, validation set, and test set to obtain the final intelligent recognition model for corneal and conjunctival tumors; The data cleaning process for all unlabeled corneal images includes: A slit-lamp image quality monitoring model was established. The model was used to analyze the quality of all unlabeled corneal images, which were then divided into low-quality and high-quality images. All low-quality images were filtered out, and all high-quality images were retained. The high-quality images were adjusted to a preset standard size and format using cubic difference to obtain the unlabeled dataset. The standard size is 512×512 pixels, and the standard format is the lossless PNG format. The process of establishing the slit lamp image quality monitoring model includes: 12,000 slit-lamp corneal images were selected, which are corneal images of patients obtained through a slit-lamp microscope. Ophthalmologists labeled each slit-lamp corneal image, classifying all images into high-quality and low-quality images, and deleting all low-quality images. The pixel values of all remaining slit-lamp corneal images were normalized to between 0 and 1, and the size of all slit-lamp corneal images was adjusted to 224×224 pixels to obtain the image dataset. An image quality monitoring model was constructed using the ResNet50 deep learning algorithm. An adaptive moment estimation optimizer was used to optimize the image quality monitoring model. The image dataset was used to iteratively train the image quality monitoring model for 180 rounds, and finally the slit lamp image quality monitoring model was obtained. The low-quality images include non-ocular surface images, poor field of view, out-of-focus images, and poor positioning; The pathological results include carcinoma in situ, conjunctival lymphoma, squamous cell carcinoma, melanoma, actinic keratosis, conjunctival intraepithelial neoplasia, primary acquired atypia melanosis, nevus, conjunctival cyst, squamous papilloma, and dermoid tumor. The process of filtering the corneal image includes: Based on the pathological results, patients with inaccurate diagnoses are manually removed. After removal, the corneal images of all remaining patients are desensitized. The desensitization process involves removing the patient's personal information (usually, one patient has multiple images). Using the slit-lamp image quality monitoring model, low-quality images are deleted from all remaining corneal images, and the remaining corneal images are uniformly adjusted to standard size and format using cubic difference. The process of establishing the self-supervised learning network includes: The self-supervised learning network is built based on the MAE (Masked Autoencoders) algorithm. The encoder uses a 24-layer ViT (Vision Transformer) structure, the decoder uses an 8-layer ViT structure, and the vectors are learnable vectors. It is worth noting that the self-supervised learning network can also be based on beit2 (the second version of Bidirectional Encoder representation from Image Transformer) and DINOv2 (the second version of Data-agnostic Intrinsic Noise Online Learning). The process of dividing all image patches into occluded image patches and visible image patches includes: The unlabeled corneal images in the unlabeled dataset are divided into rectangular grids to obtain several image blocks. 85%n of the image blocks are randomly occluded, where n is the total number of all image blocks. The occluded image blocks are denoted as occluded image blocks, and the 15%n image blocks other than the occluded image blocks are denoted as visible image blocks. The process of training the self-supervised learning network includes: The visible image patch is used as input to the encoder to obtain its features, denoted as visible patch features. A vector is constructed to represent the features of the occluded image patch, denoted as mask patch features. The visible patch features and mask patch features are concatenated to obtain concatenated features, which are used as input to the decoder. The decoder reconstructs the image based on the concatenated features to obtain a reconstructed image. The reconstruction error between the reconstructed image and the unlabeled corneal image is obtained, and the training parameters of the self-supervised learning network are adjusted to minimize the reconstruction error, thus obtaining a general model for the ocular surface. The process of training the self-supervised learning network also includes: Before dividing the label-free corneal image into rectangular grids, image enhancement is performed on the label-free corneal image. The specific process includes: The image enhancement includes image normalization, random horizontal image flipping, and image resizing; the image resizing process involves adjusting the size of all unlabeled corneal images to 224×224. The embedding vector of the encoder is set to 1024, and the embedding vector of the decoder is set to 512; the size of the image patch is set to 16×16, the batch size is set to 1536, the training epoch is set to 800, and the first 40 epochs are used for warm-up learning rate, which is [0, 1.5×10-4]; The process of dividing all corneal images in the labeled dataset into training, validation, and test sets includes: Obtain the total number m of all corneal images in the labeled dataset; set the number ratio as 70%m:15%m:15%m, and divide the labeled dataset into three subsets according to the number ratio, which are respectively denoted as training set, validation set and test set; It is worth noting that the classification layer of the recognition network can also be replaced by FNN+softmax (FNN (FeedForward Network), SVM (Support Vector Machine), or LR (Logistic Regression). Before training the recognition network, image enhancement is performed on all corneal images, and the image enhancement process is the same as the image enhancement process for the unlabeled corneal images. The process of training the recognition network also includes: Set the batch size to 32, use stochastic gradient descent as the optimizer, set the initial learning rate to 0.01, set the weightdecay to 1e-4, and set the training epochs to 100. The process of training the recognition network also includes: The training set is input into the recognition network to train it, and the recognition network outputs prediction results. During training, the prediction results are evaluated using the validation set every epoch to obtain evaluation metrics. The parameters of the recognition network are adjusted based on the evaluation metrics and the optimizer to obtain the final intelligent recognition model for corneal and conjunctival tumors. The evaluation metrics include accuracy, precision, and recall. The intelligent recognition model for corneal and conjunctival tumors is then evaluated using the test set to obtain a test loss. The test loss is used to determine whether the intelligent recognition model for corneal and conjunctival tumors meets the requirements of practical applications. It should be noted that the intelligent recognition of this invention covers a more comprehensive range of tumor types, and can identify the status of dozens of tumors (carcinoma in situ, conjunctival lymphoma, squamous cell carcinoma, melanoma, actinic keratosis, conjunctival intraepithelial neoplasia, primary acquired atypia melanosis, nevus, conjunctival cyst, squamous epithelial papilloma, and dermoid tumor); in internal and external test sets, the area under the receiver operating characteristic curve (AUC) of the corneal and conjunctival recognition model constructed by this invention exceeds 0.89 for each status (benign, malignant, and precancerous); This project uses a large amount of unlabeled ocular surface data to build a pre-trained ocular surface model, which improves the recognition effect of corneal and conjunctival tumors. On the external test set, compared with the model trained on the ImageNet dataset, the AUC of each category is improved by 0.03-0.14 (3%~14%). In a preferred embodiment of the present invention, the process of determining the image features and disease weights of the corneal and conjunctival tumor at the disease stage includes: The intelligent corneal and conjunctival tumor recognition model identifies corneal and conjunctival tumor regions in the historical corneal images. The corneal and conjunctival tumor region is the area occupied by the corneal and conjunctival tumor in the historical corneal images. The area ratio of the corneal and conjunctival tumor region in the historical corneal images is obtained and denoted as the corneal and conjunctival tumor region ratio. Based on the proportion of corneal and conjunctival tumor regions in each historical corneal image, several region proportion intervals are set, and the average value of the corneal and conjunctival tumor region proportions at both ends of the region proportion interval is obtained. The region proportion intervals are sorted from smallest to largest according to the size of the average value and numbered. Each region proportion interval is recorded as a disease stage, and the number corresponding to the region proportion interval is recorded as the disease weight of the region proportion interval. Based on the intelligent identification model for corneal and conjunctival tumors, the image features of all historical corneal images within the disease stage are obtained and denoted as the corneal and conjunctival tumor image features of the disease stage. The proportion of corneal and conjunctival tumor regions is determined by outputting a tumor region mask through the segmentation branch of the recognition model, and the ratio of the mask area to the total area of the corneal image is calculated. The disease weights are assigned in ascending order of stage number (stage 1 weight 1, stage 2 weight 2, and so on), with a higher weight indicating a more severe disease. Step S4: Obtain the current corneal image of the current patient, determine the target clinical digital features and clinical weights of the current patient, determine the target disease stage and disease weights of the current corneal image, and obtain the risk assessment value of the current patient; By integrating current patients’ clinical characteristics with corneal image characteristics and combining them with historical data models, quantitative risk assessment results are output to provide data support for clinical decision-making. When acquiring current corneal images of the current patient, use a slit-lamp microscope consistent with historical data to ensure standardized illumination intensity (500-1000 lux), magnification (10-40x), and imaging distance (10-15cm); the entire corneal area should be covered, with a focus on areas suspected of having tumors, and at least three images should be acquired from different angles (anteroposterior, left 45°, and right 45°). Obtain the current patient's ocular history (including 4 items such as keratitis and pterygium surgery history), age range, and gender through electronic medical records or consultations to form an original feature vector; In a preferred embodiment of the present invention, the process of obtaining the current patient's risk assessment value includes: The patient characteristics of the current patient are obtained and denoted as current patient characteristics. Based on the patient characteristics, the clinical numerical characteristics of the current patient are obtained and denoted as current clinical numerical characteristics. The numerical characteristic similarity between the current clinical numerical characteristics and each clinical numerical characteristic is obtained. The clinical numerical characteristic with the highest numerical characteristic similarity to the current clinical numerical characteristic is selected and denoted as the target clinical numerical characteristic of the current patient. The clinical weight corresponding to the target clinical numerical characteristic is obtained and denoted as the clinical weight of the current patient. The image features of the current corneal image are obtained and denoted as the current image features. The image feature similarity between the current image features and the image features of each corneal and conjunctival tumor image is obtained. The corneal and conjunctival tumor image feature with the highest image feature similarity to the current image features is selected and denoted as the target corneal and conjunctival tumor image feature. The disease stage of the target corneal and conjunctival tumor image feature is obtained and denoted as the target disease stage of the current corneal image. The disease weight corresponding to the target disease stage is obtained and denoted as the disease weight of the current corneal image. The current patient's risk assessment value Rav = K × (Cw × Sim) Df +Q×Dw×Sim If ), where K is the preset first correction coefficient and K > 0, Sim Df The numerical feature similarity between the current clinical numerical features and the target clinical numerical features is given by Cw, where Cw is the clinical weight of the current patient, and Sim is the numerical feature similarity. If The image feature similarity between the current image features and the target corneal and conjunctival tumor image features is given by Dw, where Dw is the disease weight of the current corneal image, and Q is a preset second correction coefficient with Q > 1. The process of determining the image features of the target corneal and conjunctival tumor includes: Hamming distance is used to calculate the difference between current clinical data features and historical clinical data features. All historical clinical data features are iterated through, and Sim is selected. Df The highest-ranking feature is used as the target clinical numerical feature; Cosine similarity was used to calculate the similarity between the current image features and the image features in the feature databases for each disease stage; Sim was selected. If The highest disease stage is used as the target disease stage. If the regional proportion matching and feature matching results conflict (e.g., the proportion belongs to stage 2, and the highest feature similarity is stage 3), then the feature matching result shall prevail (with higher weight). It should be noted that the disease weight (Dw) in the risk assessment value is directly related to the proportion of the tumor area. The larger the proportion (the higher the disease stage), the larger the Dw value, and the higher the Rav value. This can intuitively reflect the extent of tumor involvement in corneal images and help doctors judge the progression of the disease. The clinical weight (Cw) reflects the frequency of patient characteristics (such as eye history, age, gender, etc.) in historical data. If the combination of clinical characteristics corresponding to a high Cw value is superimposed with a high-risk disease stage, the patient's comprehensive risk level can be quantified by Rav, providing a basis for developing treatment strategies. By calculating the current patient's risk assessment value, a quantitative reference is provided for ophthalmologists, helping them to more accurately determine whether a patient has corneal and conjunctival tumors and the stage of the tumor, thereby assisting the clinical diagnostic process. The level of the risk assessment value helps doctors determine whether the current patient has corneal and conjunctival tumors. For example, if Rav exceeds a preset threshold, it can prompt doctors to pay further attention or take diagnostic measures. The risk assessment value comprehensively considers factors such as the patient's clinical digital characteristics and the disease stage of corneal images, and can reflect the severity of the patient's condition. Doctors can formulate more appropriate treatment plans based on this value. Based on historical disease data and current patient characteristics, the risk assessment value can predict the development trend of corneal and conjunctival tumors to a certain extent, providing support for early intervention and treatment, and helping to improve the patient's prognosis. By correlating the Rav of a large number of patients with actual diagnostic results (such as pathological type and prognosis), the correction coefficients (K, Q) or feature weights in the risk assessment formula can be further optimized to improve the accuracy of the model.
[0020] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.
Claims
1. A method for intelligent identification of corneal and conjunctival tumors based on self-supervised learning, characterized in that, Includes the following steps: Step S1: Obtain the patient's historical medical data, which includes the patient's historical corneal images and patient characteristics. The patient characteristics are several feature indicators, including eye medical history, age, and gender. Step S2: Obtain several clinical numerical characteristics based on the patient's patient characteristics; Obtain the total number of all patients in the historical disease data, and obtain the clinical weight of the clinical digital feature based on the proportion of the number of patients corresponding to the clinical digital feature in the total number; Step S3: Based on a self-supervised learning network, establish a corneal and conjunctival tumor intelligent identification model, and according to the corneal and conjunctival tumor intelligent identification model, obtain the proportion of corneal and conjunctival tumor regions in each historical corneal image, obtain several disease stages, and determine the corneal and conjunctival tumor image features and disease weights of each disease stage. Step S4: Obtain the current corneal image of the current patient, determine the target clinical digital features and clinical weights of the current patient, determine the target disease stage and disease weight of the current corneal image, and obtain the risk assessment value of the current patient.
2. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 1, characterized in that, In step S2, the process of obtaining the clinical digital features includes: The index values of the feature indicators are obtained. The index values of the eye disease history include keratitis, pterygium surgery history, family history of tumors, and long-term ultraviolet radiation exposure history. The index values of the gender include male and female. The index values of the age include several preset age ranges. The index values of the feature indicators are assigned based on binary encoding. The assigned index values are then concatenated in a preset order to obtain the clinical digital features of the feature indicators.
3. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 1, characterized in that, In step S2, the clinical weight Cw of the i-th clinical numerical feature i =n i / N, where n i Let N be the number of patients in the historical disease data whose numerical feature is the i-th clinical numerical feature, and let N be the total number of all patients in the historical disease data.
4. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 1, characterized in that, In step S3, the process of establishing the intelligent identification model for corneal and conjunctival tumors includes: Unlabeled corneal images are acquired, and all unlabeled corneal images are cleaned to obtain an unlabeled dataset. Corneal images of patients with corneal and conjunctival tumors are obtained from medical institutions, and the pathological results of the patients are obtained. The corneal images are filtered and labeled according to the pathological results to obtain a labeled dataset. A self-supervised learning network is established, and the unlabeled corneal images in the unlabeled dataset are divided into several image patches, and all image patches are further divided into occluded image patches and visible image patches; the self-supervised learning network is trained based on the occluded image patches and visible image patches to obtain a general model of the ocular surface. A recognition network is constructed based on the general model of the ocular surface. All corneal images in the labeled dataset are divided into a training set, a validation set, and a test set. The recognition network is trained using the training set, validation set, and test set to obtain the final intelligent recognition model for corneal and conjunctival tumors.
5. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 4, characterized in that, In step S3, the data cleaning process for all unlabeled corneal images includes: A slit-lamp image quality monitoring model was established. The model was used to analyze the quality of all unlabeled corneal images, classifying them into low-quality and high-quality images. All low-quality images were filtered out, and all high-quality images were retained. The high-quality images were then adjusted to a preset standard size and format using cubic difference to obtain the unlabeled dataset.
6. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 5, characterized in that, In step S3, the process of establishing the slit lamp image quality monitoring model includes: 12,000 slit-lamp corneal images were selected, which are corneal images of patients obtained through a slit-lamp microscope. Ophthalmologists labeled each slit-lamp corneal image, classifying all images into high-quality and low-quality images, and deleting all low-quality images. The pixel values of all remaining slit-lamp corneal images were normalized to between 0 and 1, and the size of all slit-lamp corneal images was adjusted to 224×224 pixels to obtain the image dataset. An image quality monitoring model was constructed using the ResNet50 deep learning algorithm. An adaptive moment estimation optimizer was used to optimize the image quality monitoring model. The image dataset was used to iteratively train the image quality monitoring model for 180 rounds, and finally the slit lamp image quality monitoring model was obtained.
7. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 1, characterized in that, In step S3, the process of determining the image features and disease weights of the corneal and conjunctival tumor at the disease stage includes: The intelligent corneal and conjunctival tumor recognition model identifies corneal and conjunctival tumor regions in the historical corneal images. The corneal and conjunctival tumor region is the area occupied by the corneal and conjunctival tumor in the historical corneal images. The area ratio of the corneal and conjunctival tumor region in the historical corneal images is obtained and denoted as the corneal and conjunctival tumor region ratio. Based on the proportion of corneal and conjunctival tumor regions in each historical corneal image, several region proportion intervals are set, and the average value of the corneal and conjunctival tumor region proportions at both ends of the region proportion interval is obtained. The region proportion intervals are sorted from smallest to largest according to the size of the average value and numbered. Each region proportion interval is recorded as a disease stage, and the number corresponding to the region proportion interval is recorded as the disease weight of the region proportion interval. Based on the intelligent identification model for corneal and conjunctival tumors, image features of all historical corneal images within the disease stage are obtained and denoted as the corneal and conjunctival tumor image features of the disease stage.
8. The intelligent identification method for corneal and conjunctival tumors based on self-supervised learning according to claim 1, characterized in that, In step S4, the process of obtaining the current patient's risk assessment value includes: The patient characteristics of the current patient are obtained and denoted as current patient characteristics. Based on the patient characteristics, the clinical numerical characteristics of the current patient are obtained and denoted as current clinical numerical characteristics. The numerical characteristic similarity between the current clinical numerical characteristics and each clinical numerical characteristic is obtained. The clinical numerical characteristic with the highest numerical characteristic similarity to the current clinical numerical characteristic is selected and denoted as the target clinical numerical characteristic of the current patient. The clinical weight corresponding to the target clinical numerical characteristic is obtained and denoted as the clinical weight of the current patient. The image features of the current corneal image are obtained and denoted as the current image features. The image feature similarity between the current image features and the image features of each corneal and conjunctival tumor image is obtained. The corneal and conjunctival tumor image feature with the highest image feature similarity to the current image features is selected and denoted as the target corneal and conjunctival tumor image feature. The disease stage of the target corneal and conjunctival tumor image feature is obtained and denoted as the target disease stage of the current corneal image. The disease weight corresponding to the target disease stage is obtained and denoted as the disease weight of the current corneal image. The current patient's risk assessment value Rav = K × (Cw × Sim) Df +Q×Dw×Sim If ), where K is the preset first correction coefficient and K > 0, Sim Df The numerical feature similarity between the current clinical numerical features and the target clinical numerical features is given by Cw, where Cw is the clinical weight of the current patient, and Sim is the numerical feature similarity. If The image feature similarity between the current image features and the target corneal and conjunctival tumor image features is given by denoting ...