CN112053321A - Artificial intelligence system for identifying high myopia retinopathy - Google Patents
Artificial intelligence system for identifying high myopia retinopathy Download PDFInfo
- Publication number
- CN112053321A CN112053321A CN202010752871.4A CN202010752871A CN112053321A CN 112053321 A CN112053321 A CN 112053321A CN 202010752871 A CN202010752871 A CN 202010752871A CN 112053321 A CN112053321 A CN 112053321A
- Authority
- CN
- China
- Prior art keywords
- retinopathy
- myopia
- image
- lesion
- optical coherence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B3/00—Apparatus for testing the eyes; Instruments for examining the eyes
- A61B3/10—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions
- A61B3/102—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions for optical coherence tomography [OCT]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B3/00—Apparatus for testing the eyes; Instruments for examining the eyes
- A61B3/10—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions
- A61B3/12—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions for looking at the eye fundus, e.g. ophthalmoscopes
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B3/00—Apparatus for testing the eyes; Instruments for examining the eyes
- A61B3/10—Objective types, i.e. instruments for examining the eyes independent of the patients' perceptions or reactions
- A61B3/14—Arrangements specially adapted for eye photography
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10101—Optical tomography; Optical coherence tomography [OCT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30041—Eye; Retina; Ophthalmic
Abstract
The invention relates to the technical field of medical image processing, in particular to an artificial intelligence system for identifying high myopia retinopathy, which comprises: the image acquisition module is used for acquiring an optical coherence tomography image and fundus reference of a patient to be identified; the first identification module is used for inputting the optical coherence tomography image into a high myopia retinopathy identification model, judging whether the optical coherence tomography image has pathological changes or not and obtaining a pathological change type result; the second identification module is used for inputting the fundus color photograph into the high myopia retinopathy stage model, judging the retinopathy stage of the fundus color photograph and obtaining a lesion stage judgment result; and the report generating module is used for generating a diagnosis and treatment suggestion report according to the lesion type result and the lesion stage judgment result. The invention can quickly and effectively identify the optical coherence tomography image and the eyeground color photograph of the patient to identify the common pathological changes of the retina with high myopia.
Description
Technical Field
The invention relates to the technical field of medical image processing, in particular to an artificial intelligence system for identifying high-myopia retinopathy.
Background
Currently, the examination procedure for high myopia retinopathy is: firstly, an ophthalmologist performs optometry and eye axis measurement on a patient, judges whether the patient has high myopia and related risk factors, and then judges whether retinopathy related to the high myopia (including retinal cleavage, macular hole, retinal detachment and choroidal neovascularization) exists according to Optical Coherence Tomography (OCT) image analysis of the patient. As seen in the examination of highly myopic retinopathy, it has the following major disadvantages:
1. the possibility of implementing large-scale population screening is low: at present, China has a large number of myopia population, medical resources are limited, the height of myopia patients is 6 hundred million, wherein the total myopia rate of teenagers is as high as 55%, the myopia rate is the main force of the current myopia, and the height of myopia patients accounts for about 21% among all myopia groups, and the patients are easy to have complications such as retinal atrophy, cleavage, macular hole, retinal detachment, choroidal neovascularization and the like, so that irreversible damage to vision is caused, and the living level of the young and the young people is seriously influenced.
2. Highly myopic fundus manifestations are complex, demanding high on the inspectors' conditions: the retinal atrophy of patients with high myopia is serious, the complications are many, the clinical experience of inspectors is high, the interpretation of high myopia retinopathy requires professional training of ophthalmologists and the accumulation of long-time experience, and generally, community hospitals or comprehensive hospitals with lower levels and physical examination centers do not have professional ophthalmologists and cannot perform comprehensive detection and evaluation on the retinopathy.
3. Patients with high myopia require lifelong follow-up: highly myopic patients need follow-up observation every 3-6 months according to different disease conditions, however, effective medical resources are lacked in partial remote areas and primary hospitals, the patients need to visit in comprehensive hospitals at other places, and long-term repeated running waves can bring large burden to the patients, and compliance reduction and missed visits of the patients are easily caused.
OCT images (optical coherence tomography images) contain more information, but the image representation is more complex: compared with other fundus imaging examinations such as fundus color photography and the like, OCT can display more lesion information, but images are more complex, the requirement on film reading of doctors is higher, general doctors do not have the film reading capability of OCT, and specialized doctors need to be trained for a longer time to obtain higher film reading accuracy. Meanwhile, the deep learning training of the OCT image also puts higher requirements on a corresponding artificial intelligence model.
In summary, due to the huge base number of people with high myopia in China, the requirement of long-term follow-up observation and the limitation of basic medical resources, the implementation of large-scale high myopia retinopathy screening and the guarantee of long-term effective follow-up work for high-risk patients have various difficulties and inconveniences, which may cause that some patients cannot be accurately diagnosed in early stage of the disease, thereby missing the optimal treatment opportunity, causing irreversible vision damage and finally bringing great loss to individuals, families and society. Therefore, the rapid and effective screening of the high-myopia retinopathy is realized, and doctors can be helped to diagnose the retinopathy at the early stage of the disease and even at the very early stage, so that the early discovery, the early diagnosis and the early intervention are realized. Meanwhile, the patient is subjected to long-term effective follow-up observation, so that the progress of the pathological changes can be closely monitored, and possible visual impairment can be prevented. In addition, no deep learning model specially aiming at the OCT image exists in the market at present, and in view of the complexity of the high-myopia OCT image, a high-efficiency deep learning model needs to be specially trained aiming at the situation so as to rapidly realize the screening of large-scale people.
Disclosure of Invention
The invention aims to overcome at least one defect of the prior art and provides an artificial intelligence system for identifying high myopia retinopathy, which is used for quickly and effectively identifying an optical coherence tomography image and an eyeground color photograph of a patient to identify common high myopia retinopathy.
The invention adopts the technical scheme that an artificial intelligence system for identifying high myopia retinopathy comprises: the image acquisition module is used for acquiring an optical coherence tomography image and an eyeground color photograph of a patient to be identified;
the first identification module is used for inputting the optical coherence tomography image into a high myopia retinopathy identification model, identifying whether the optical coherence tomography image has pathological changes or not and obtaining a pathological change type result;
the second identification module is used for inputting the fundus color photograph into the high myopia retinopathy stage model, judging the retinopathy stage of the fundus color photograph and obtaining a lesion stage judgment result;
and the report generating module is used for generating a diagnosis and treatment suggestion report according to the lesion type result and the lesion stage judgment result.
The invention relates to an artificial intelligence system for identifying high myopia retinopathy, which is characterized in that an optical coherence tomography image and an eyeground color photograph of a patient to be identified are obtained through an obtaining module; inputting the OCT image into a high myopia retinopathy recognition model through a first image recognition module, judging whether common lesions exist in the OCT image, and judging the common lesion types to include one or more of retinal cleavage, macular hole, retinal detachment and choroidal neovascularization; inputting the corresponding fundus color photograph into the high myopia retinopathy stage model through the second image recognition module, and judging the pathological change stage grade of the input fundus color photograph; and finally, the report generation module generates a corresponding diagnosis and treatment suggestion report according to the identification result of the high myopia lesion type of the OCT image and the stage result of the fundus color photograph. According to the invention, by means of the high sensitivity and accuracy of artificial intelligence deep learning, the optical coherence tomography image of the user is analyzed, so that the early screening of common retinopathy related to high myopia (retinal cleavage, macular hole, retinal detachment and choroidal neovascularization) is more accurate, intelligent and portable, the screening efficiency is improved, doctors can be helped to diagnose the pathological changes in early or even ultra-early stage of diseases, early discovery, early diagnosis and early intervention are realized, and irreversible damage to the eyesight of people caused by the pathological changes is reduced. In addition, the invention judges whether one or more of four common lesion types exist or not by inputting the OCT image into the high myopia retina recognition model, and the four lesion types have different risk levels: retinal cleavage is a risk of stroke, macular hole and choroidal neovascularization are high risk, and retinal detachment is a critical risk; the fundus color photograph is input into the high myopia retina staging model, and the staging result of the fundus color photograph is judged according to the internationally recognized staging standard, wherein the higher the staging is, the more serious the disease condition is. And (4) correspondingly generating a diagnosis and treatment suggestion report by integrating the risk grade corresponding to the type of the pathological changes of the OCT image and the severity of the pathological conditions corresponding to the stage result of the fundus color photograph. When a large number of high-myopia OCT images and fundus color photographs are specially trained in a deep learning model, the ophthalmologist with profound seniority reads the images and gives an authoritative diagnosis and treatment suggestion aiming at the disease condition, so that in the image recognition process, the decision idea of the retina specialist is simulated according to the recognition result to provide a corresponding diagnosis and treatment suggestion report for the patient, the patient without the related professional background can further recognize the retinopathy condition, the professional film reading capability is not needed, the pathological change can be diagnosed only by depending on the suggestion report, the early-finding early diagnosis is realized, and the examination time is greatly saved.
Further, still include: and the lesion positioning module is used for positioning a lesion part and generating a visual activation map when the lesion part is judged to be present according to the fact that whether the optical coherence tomography image has the lesion.
The invention also comprises a lesion positioning module which can judge whether the optical coherence tomography image has lesions through the high myopia retinopathy recognition model and judge the type of the lesions, and can position the lesion part of the image through the visual activation map if the optical coherence tomography image has the lesions. The lesion positioning module in the invention adopts the visual activation map, so that the patient can intuitively make preliminary judgment on the self lesion condition without professional film reading capability, thereby realizing the judgment of early treatment.
Further, the high myopia retinopathy recognition model is trained by adopting the following steps:
classifying a plurality of optical coherence tomography image samples of high myopia retinopathy as a sample image set according to retinal cleavage, macular hole, retinal detachment and choroidal neovascularization;
preprocessing the classified sample image set;
and inputting the preprocessed sample image set into a convolutional neural network for deep learning training to obtain a high myopia retinopathy recognition model based on the optical coherence tomography image.
The invention trains a high myopia retinopathy recognition model based on an optical coherence tomography image through deep learning, and the process of the deep learning is as follows: firstly, an ophthalmologist with profound seniors reads a large number of optical coherence tomography image samples as a sample image set, classifies each image according to internationally recognized diagnostic standards, and attaches corresponding disease labels including retinal cleavage, macular hole, retinal detachment and choroidal neovascularization to each OCT image; secondly, preprocessing the classified sample image set, unifying the attributes of the images and eliminating interference factors except lesion features; and finally, inputting the preprocessed sample image set into a convolutional neural network for model training according to a training method of deep learning, and obtaining an optimal model as a high myopia retinopathy recognition model through verification. On the premise of having a large number of optical coherence tomography images (OCT images) with definite retinopathy types, the deep learning technology is utilized to carry out deep learning on the convolutional neural network, and a deep learning model capable of efficiently identifying common retinopathy in high-myopia crowds through the optical coherence tomography images is obtained through training and verification, namely the high-myopia retinopathy identification model based on the optical coherence tomography images effectively solves the problem of high-myopia retinopathy in the screening process, and greatly improves the possibility of screening the high-myopia retinopathy of large-scale crowds.
Further, the preprocessing the classified sample image set includes:
cutting the classified optical coherence tomography image of the high myopia retinopathy, and cutting off a color photograph part of the sample image set;
and unifying the sample image sets after the cutting processing into black-background white images.
In the deep learning and training process of the high myopia retina recognition model, the classified optical coherence tomography image sample image set needs to be preprocessed, wherein the specific preprocessing process comprises the following steps: firstly, cutting a classified sample image set, cutting off a color photograph part on the left side of an image, and reserving a black and white tomographic image part; secondly, standardizing the size of the cut image, and uniformly converting the size of the cut image into the size of 500 × 764; finally, in order to improve the training effect, the background color of the image is unified, the white background image is subjected to black-white inversion, and the background color of the image is uniformly set to be black. Through the preprocessing process, useful image information in the optical coherence tomography image is reserved, the size, color and other relevant attributes of the image are processed in a unified mode, other interference factors except lesion features are eliminated, and the deep learning efficiency and the training effect are effectively improved.
Further, the step of inputting the preprocessed sample image set into a convolutional neural network for deep learning training to obtain a high myopia retinopathy recognition model based on the optical coherence tomography image includes:
dividing the preprocessed sample image set into a training set, a verification set and a test set;
performing image processing on the training set to increase the number of the training sets to n times, wherein n is an integer greater than or equal to 2, and the image processing comprises adjusting contrast and/or brightness and/or image rotation and/or mirror inversion;
and performing deep learning training and verification on the convolutional neural network by adopting the training set and the verification set after image processing to obtain a high myopia retinopathy recognition model, and testing and evaluating the effect of the high myopia retinopathy recognition model by adopting the test set.
The specific learning and training process for obtaining the high myopia retinopathy recognition model by deep learning of the convolutional neural network comprises the following steps: firstly, dividing a sample image set after classification and pretreatment into a training set, a verification set and a test set; secondly, carrying out increment to n times on a series of methods such as image adjustment contrast, brightness, image rotation, mirror inversion and the like of a training set, then inputting the training set after increment into a convolutional neural network for model training, and verifying the trained model by using a verification set to obtain an optimal model as a high myopia retinopathy recognition model; and finally, testing and evaluating the effect of the high myopia retinopathy identification model through the test set. The invention obtains the high myopia retinopathy recognition model by carrying out targeted training on the convolutional neural network, and provides a powerful technical means for screening the high myopia retinopathy common diseases.
Further, the high myopia retinopathy stage model is trained by adopting the following steps:
classifying the pathological change stages by taking a plurality of fundus color photograph samples with high-degree myopia retinopathy as a color photograph sample set according to a preset diagnosis standard;
preprocessing the classified color photo sample set;
and inputting the preprocessed color photograph sample set into a convolutional neural network for deep learning training to obtain a high myopia retinopathy stage model based on the fundus color photograph.
The invention trains a high myopia retinopathy stage model based on fundus color photography in deep learning, and the deep learning process comprises the following steps: firstly, a large number of fundus color photographs are read by ophthalmologists with profound seniors, and according to internationally recognized grading standards, highly myopic fundus lesions are divided into four stages according to the degree of atrophy: C1-C4, preprocessing a large number of fundus color photographs of high myopia retinopathy after the period division, unifying the attributes of the images, eliminating interference factors outside the period division of the pathological changes, finally inputting the preprocessed fundus color photographs into a convolutional neural network for model training according to a deep learning training method, and establishing a four-classification deep learning model. On the premise of having a large number of fundus color photographs marked with pathological change international stages, the invention utilizes the deep learning technology to carry out deep learning on the convolutional neural network, trains and verifies to obtain four international stages C1-C4 capable of judging high myopia pathological changes through the fundus color photographs, namely, a high myopia retinopathy stage model based on the fundus color photographs can automatically judge the retinal pathological changes stage in the current fundus color photographs and judge the severity and urgency of the disease, and can timely and effectively help patients to make diagnosis and treatment judgments so as not to delay the irreversible consequences caused by the disease.
Further, the locating the lesion site and generating the visualized activation-like map includes:
performing feature extraction on the optical coherence tomography image of the highly myopic retina, and generating a feature map;
calculating and positioning a lesion part in the characteristic diagram by adopting a class activation algorithm to generate a heat map;
and overlapping the characteristic graph and the heat graph to generate a visual activation-like graph.
The invention uses a high myopia retinopathy recognition model and a similar activation mapping method to position the pathological part of an OCT image, and the specific positioning operation process is as follows: inputting the OCT image into a high myopia retinopathy recognition model, extracting the output of the last convolutional layer of the model to obtain a feature map, calculating by using a class activation algorithm and generating a heat map to reveal an important region in the feature map, and finally overlapping the feature map and the heat map to generate a visual class activation map. The part marked by the heat map in the visual activation map represents the part of the lesion, and through the visual activation map, a patient can intuitively make preliminary judgment on the self lesion condition without professional film reading capacity, so that the judgment of early treatment is realized, a fundus disease specialist can also refer to the lesion part positioned by the visual activation map, the part needing to be verified is quickly positioned, and the examination time is greatly saved.
Further, the locating the lesion site and generating the visual activation map further comprises: when the optical coherence tomography image contains multiple different lesion types, different markers are used to locate the lesion site and provide a superimposed composite thermal map and a separate thermal map for each lesion, respectively.
In the case of optical coherence tomography images of highly myopic retinas, where multiple lesions may be present, when multiple lesions are present in the image, and two or more lesions are present at the same location or in close proximity, different lesions are marked with different colors during the generation of the heat map, and a visual activation map and a separate heat map for each lesion are generated. By outputting the positioning result graphs of various pathological changes, the judgment and analysis of various retinopathy are more accurate, and the treatment result is more accurately judged in the shortest time.
Further, at least the acquisition module is arranged on a cloud platform, and the optical coherence tomography image and the fundus color photograph are uploaded to the cloud platform by a user.
The invention integrates an artificial intelligence system for identifying the high myopia retinas into a scanning instrument or an artificial intelligence network based on cloud computing service, users such as community hospitals, primary hospitals, physical examination centers and other units with optical coherence tomography scanning instruments collect corresponding patient data and images according to inclusion and discharge standards, and each image is processed by the system for about 3-6s to give a preliminary examination suggestion.
Further, the convolutional neural network adopts an inclusion-ResNetV 2 network model.
Compared with the prior art, the invention has the beneficial effects that:
1. high efficiency: the invention has the sensitivity and specificity probability of identifying the common lesion of the high myopia retina in the OCT image as high as more than 90 percent, takes about 3-6s for analyzing each image, can simultaneously meet the requirement of uploading analysis photos in batches, and has the efficiency far higher than that of other inspection methods.
2. The screening of the common high-myopia retinal pathological changes of large-scale population is realized: the invention provides a noninvasive examination, is convenient and rapid, the examination time is only a few minutes, no professional ophthalmologist is needed for analyzing pictures, the result can be rapidly obtained only by uploading the pictures to the cloud platform, and a foundation is laid for large-scale crowd screening.
3. The application range is wide: whether an individual or a group such as a community hospital, a primary hospital, a physical examination center, etc., can upload the obtained OCT image to the cloud platform without being limited by location.
4. Convenience: the user can be at any time, any place with a network. The wide-area fundus image is uploaded to an online downloading system, and then an analysis result can be quickly obtained.
5. The economic efficiency is as follows: the expenditure required for specially training or hiring ophthalmonogy doctors is saved.
Drawings
FIG. 1 is a system block diagram of the present invention.
Detailed Description
The drawings are only for purposes of illustration and are not to be construed as limiting the invention.
Examples
An artificial intelligence system for identifying highly myopic retinopathy, including retinoschisis, macular holes, retinal detachment, choroidal neovascularization, as shown in fig. 1, comprising:
the image acquisition module is used for acquiring an optical coherence tomography image and an eyeground color photograph of a patient to be identified;
the first identification module is used for inputting the optical coherence tomography image into a high myopia retinopathy identification model, identifying whether the optical coherence tomography image has pathological changes or not and obtaining a pathological change type result;
specifically, the training step of the high myopia retinopathy recognition model comprises the following steps:
s1, classifying the high myopia retina OCT image samples as a sample image set according to retinal cleavage, macular hole, retinal detachment and choroidal neovascularization;
s2, preprocessing the classified sample image set;
specifically, the pretreatment process is as follows:
s21, cutting the classified optical coherence tomography image of the high myopia retinopathy, and cutting off the color photograph part of the sample image set;
s22, unifying the sample image sets after the cropping processing into a black-and-white image with a size of 500 × 764.
S23, inputting the preprocessed sample image set into a convolutional neural network for deep learning training to obtain a high myopia retinopathy recognition model;
specifically, the convolutional neural network adopts an inclusion-ResNetV 2 network model, and the deep learning training process is as follows:
s231, dividing the preprocessed sample image set into a training set, a verification set and a test set;
s232, carrying out image processing on the training set to increase the number of the training set to n times, wherein n is an integer greater than or equal to 2, and the image processing comprises adjusting contrast and/or brightness and/or image rotation and/or mirror inversion;
and S233, carrying out deep learning training and verification on the convolutional neural network by adopting the training set and the verification set after image processing to obtain a high myopia retinopathy recognition model, and carrying out test evaluation on the high myopia retinopathy recognition model by adopting the test set.
The second identification module is used for inputting the fundus color photograph into the high myopia retinopathy stage model, judging the retinopathy stage of the fundus color photograph and obtaining a lesion stage judgment result;
specifically, the specific training steps of the stage model of the high myopia retinopathy of fundus color photography comprise:
s1', taking a plurality of fundus color photograph samples with high myopia retinopathy as a color photograph sample set to classify pathological changes according to a preset diagnosis standard;
s2', preprocessing the classified color picture sample set;
specifically, the preprocessing process is to uniformly convert the size of the fundus color photograph into the size of 500 × 764.
S3', inputting the preprocessed color photograph sample set into a convolutional neural network for deep learning training, and obtaining a high myopia retinopathy stage model based on fundus color photographs.
Specifically, the convolutional neural network adopts an inclusion-ResNetV 2 network model, and the deep learning training process is as follows:
firstly, a large number of fundus color photographs are read by ophthalmologists with profound seniors, and according to internationally recognized grading standards, highly myopic fundus lesions are divided into four stages according to the degree of atrophy: C1-C4, preprocessing a large number of fundus color photographs of high-myopia retinopathy after the period of the period, unifying the attributes of the images, eliminating interference factors outside the period of the pathological changes, and finally inputting the preprocessed fundus color photographs into an IncepotionResNetV 2 convolutional neural network for model training according to a training method of deep learning to establish a four-classification deep learning model.
And the report generating module is used for generating a diagnosis and treatment suggestion report according to the lesion type result and the lesion stage judgment result. Specifically, three diagnosis and treatment suggestion reports can be generated: comparing with the previous record, identifying the new type of lesion which does not exist in the past in the examination, and recommending the outpatient service special for the fundus disease for further examination; secondly, if the old lesion is found to have little change from the past record, continuous follow-up is recommended; thirdly, no pathological changes are found temporarily, and follow-up is recommended.
Preferably, the method further comprises the following steps: and the lesion positioning module is used for positioning a lesion part and generating a visual activation map when the lesion part is judged to be present according to the fact that whether the optical coherence tomography image has the lesion.
Specifically, the process of locating the lesion site and generating the visual activation map includes:
performing feature extraction on related features of retinal cleavage, macular hole, retinal detachment and choroidal neovascularization in an optical coherence tomography image of the high myopia retina, and specifically obtaining a feature map by extracting the output of the last convolutional layer of the high myopia retinopathy recognition model;
calculating and positioning a lesion part in the characteristic diagram by adopting a class activation algorithm to generate a heat map;
and (3) superposing the characteristic map and the heat map to generate a visual activation-like map, wherein when the optical coherence tomography image contains a plurality of different lesion types, or the positions of the lesions are very close to each other, even the same position can show the characteristics of two lesions, different color marks are adopted to position the lesion positions, and the superposed composite heat map and the separate heat map of each lesion are respectively provided.
Preferably, at least the acquisition module is disposed on a cloud platform, and the optical coherence tomography image and the fundus color photograph are uploaded to the cloud platform by a user.
An artificial intelligence system for identifying high-myopia retinopathy based on deep learning is integrated into a scanning instrument and an artificial intelligence network platform based on cloud computing service, units such as community hospitals, primary hospitals, physical examination centers and the like, which have OCT instruments and fundus cameras, collect corresponding patient data and images according to inclusion and exclusion standards, and after each image is processed by a screening system for about 3-6 seconds, a preliminary screening suggestion is given. If suspicious lesions are found, the system uploads the original pictures and the heat map to a clinical diagnosis and treatment center through the cloud end for the expert to recheck. This platform mainly includes four hurdles: the method comprises the following steps of (1) briefly introducing a platform, wherein the platform is used for describing the structure of the platform and available resources; the use instruction is used for guiding the user to rapidly learn the use of the platform according to the steps of the process; image analysis for optical coherence tomography image one-key analysis and uploading; fourthly, when the user has difficulty or what opinion on the use, the user can leave a message through the column. We consult this column periodically and make a quick feedback on the user's needs, optimizing the use of the platform step by step. The specific use steps are as follows:
a user enters the network cloud platform, and obtains the use right of the cloud platform after registering an application account;
the user uploads images with independent numbers (generally generated automatically by an instrument) to a system platform through the instrument independently or in batch by the account number of the user, and clicks a one-key analysis button after setting a position for storing an analysis result, each image is processed for about 3-6 seconds, and finally the analyzed images are stored in the position appointed to be stored in a user computer;
the analysis result is that corresponding suggestions are given according to the specific classification of the pictures, and if the lesion is found on the image, the part of the lesion is marked on the picture by using the heat map.
In the embodiment, the established AI system software is integrated with the scanning equipment and is accessed to the network cloud platform, so that user groups in different areas can complete preliminary screening and regular follow-up nearby through nearby community hospitals. When the system finds abnormal conditions, corresponding images and heat maps can be uploaded to a clinical diagnosis and treatment center in time through the network, and ophthalmic experts with abundant experience can manually review suspicious images and provide diagnosis and treatment suggestions for patients through a network platform.
The beneficial effect of this embodiment does:
1. high efficiency: the invention has the sensitivity and specificity probability of identifying the common lesion of the high myopia retina in the OCT image as high as more than 90 percent, takes about 3-6s for analyzing each image, can simultaneously meet the requirement of uploading analysis photos in batches, and has the efficiency far higher than that of other inspection methods.
2. The screening of the common high-myopia retinal pathological changes of large-scale population is realized: the invention provides a noninvasive examination, is convenient and rapid, the examination time is only a few minutes, no professional ophthalmologist is needed for analyzing pictures, the result can be rapidly obtained only by uploading the pictures to the cloud platform, and a foundation is laid for large-scale crowd screening.
3. The application range is wide: whether an individual or a group such as a community hospital, a primary hospital, a physical examination center, etc., can upload the obtained OCT image to the cloud platform without being limited by location.
4. Convenience: the user can be at any time, any place with a network. The wide-area fundus image is uploaded to an online downloading system, and then an analysis result can be quickly obtained.
5. The economic efficiency is as follows: the expenditure required for specially training or hiring ophthalmonogy doctors is saved.
It should be understood that the above-mentioned embodiments of the present invention are only examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific embodiments of the present invention. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention claims should be included in the protection scope of the present invention claims.
Claims (10)
1. An artificial intelligence system for identifying highly myopic retinopathy comprising:
the image acquisition module is used for acquiring an optical coherence tomography image and an eyeground color photograph of a patient to be identified;
the first identification module is used for inputting the optical coherence tomography image into a high myopia retinopathy identification model, identifying whether the optical coherence tomography image has pathological changes or not and obtaining a pathological change type result;
the second identification module is used for inputting the fundus color photograph into the high myopia retinopathy stage model, judging the retinopathy stage of the fundus color photograph and obtaining a lesion stage judgment result;
and the report generating module is used for generating a diagnosis and treatment suggestion report according to the lesion type result and the lesion stage judgment result.
2. The artificial intelligence system for identifying high-myopia retinopathy of claim 1, further comprising:
and the lesion positioning module is used for positioning a lesion part and generating a visual activation map when judging whether a lesion exists according to the identification of whether the optical coherence tomography image has the lesion.
3. The artificial intelligence system for identifying high-myopia retinopathy of claim 1, wherein the high-myopia retinopathy identification model is trained by:
classifying a plurality of optical coherence tomography image samples of high myopia retinopathy as a sample image set according to retinal cleavage, macular hole, retinal detachment and choroidal neovascularization;
preprocessing the classified sample image set;
and inputting the preprocessed sample image set into a convolutional neural network for deep learning training to obtain a high myopia retinopathy recognition model based on the optical coherence tomography image.
4. The artificial intelligence system for identifying high-myopia retinopathy of claim 3, wherein the preprocessing the classified sample image set comprises:
cutting the classified sample image set to cut off the color part of the sample image set;
and unifying the sample image sets after the cutting processing into black-background white images.
5. The artificial intelligence system for identifying high-myopia retinopathy according to claim 3, wherein the inputting of the preprocessed sample image set into a convolutional neural network for deep learning training to obtain a high-myopia retinopathy identification model based on the optical coherence tomography image comprises:
dividing the preprocessed sample image set into a training set, a verification set and a test set;
performing image processing on the training set to increase the number of the training sets to n times, wherein n is an integer greater than or equal to 2, and the image processing comprises adjusting contrast and/or brightness and/or image rotation and/or mirror inversion;
and performing deep learning training and verification on the convolutional neural network by adopting the training set and the verification set after image processing to obtain a high myopia retinopathy recognition model, and testing and evaluating the effect of the high myopia retinopathy recognition model by adopting the test set.
6. The artificial intelligence system for identifying high-myopia retinopathy of claim 1, wherein the high-myopia retinopathy stage model is trained by the following steps:
classifying the pathological change stages by taking a plurality of fundus color photograph samples with high-degree myopia retinopathy as a color photograph sample set according to a preset diagnosis standard;
preprocessing the classified color photo sample set;
and inputting the preprocessed color photograph sample set into a convolutional neural network for deep learning training to obtain a high myopia retinopathy stage model based on the fundus color photograph.
7. The artificial intelligence system for identifying high-myopia retinopathy according to claim 2, wherein the locating a lesion site and generating a visual class activation map comprises:
performing feature extraction on the optical coherence tomography image of the highly myopic retina, and generating a feature map;
calculating and positioning a lesion part in the characteristic diagram by adopting a class activation algorithm to generate a heat map;
and overlapping the characteristic graph and the heat graph to generate a visual activation-like graph.
8. The artificial intelligence system for identifying high-myopia retinopathy of claim 7, wherein the locating a lesion site and generating a visual class activation map further comprises:
when the optical coherence tomography image contains multiple different lesion types, different markers are used to locate the lesion site and provide a superimposed composite thermal map and a separate thermal map for each lesion, respectively.
9. The artificial intelligence system for identifying high-myopia retinopathy of claim 1 wherein at least the acquisition module is disposed on a cloud platform to which the OCT image and fundus color photograph are uploaded by a user.
10. The artificial intelligence system for identifying high-myopia retinopathy of claim 3, wherein the convolutional neural network employs an inclusion-ResNetV 2 network model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010752871.4A CN112053321A (en) | 2020-07-30 | 2020-07-30 | Artificial intelligence system for identifying high myopia retinopathy |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010752871.4A CN112053321A (en) | 2020-07-30 | 2020-07-30 | Artificial intelligence system for identifying high myopia retinopathy |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112053321A true CN112053321A (en) | 2020-12-08 |
Family
ID=73602368
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010752871.4A Pending CN112053321A (en) | 2020-07-30 | 2020-07-30 | Artificial intelligence system for identifying high myopia retinopathy |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112053321A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113558564A (en) * | 2021-09-23 | 2021-10-29 | 中山大学中山眼科中心 | Data processing system based on simple high myopia database construction |
| CN113706518A (en) * | 2021-09-01 | 2021-11-26 | 吉林大学 | Femoral head necrosis stage analysis method and system based on artificial intelligence |
| CN115082414A (en) * | 2022-07-08 | 2022-09-20 | 深圳市眼科医院 | Portable detection method and device based on visual quality analysis |
| TWI779731B (en) * | 2021-07-21 | 2022-10-01 | 長庚醫療財團法人基隆長庚紀念醫院 | Retinopathy report inspection method using artificial intelligence and retinopathy image integration method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108615051A (en) * | 2018-04-13 | 2018-10-02 | 博众精工科技股份有限公司 | Diabetic retina image classification method based on deep learning and system |
| CN108876775A (en) * | 2018-06-12 | 2018-11-23 | 广州图灵人工智能技术有限公司 | The rapid detection method of diabetic retinopathy |
| CN109191451A (en) * | 2018-09-11 | 2019-01-11 | 百度在线网络技术（北京）有限公司 | Method for detecting abnormality, device, equipment and medium |
| CN109493954A (en) * | 2018-12-20 | 2019-03-19 | 广东工业大学 | A kind of SD-OCT image retinopathy detection system differentiating positioning based on classification |
| CN109829882A (en) * | 2018-12-18 | 2019-05-31 | 苏州比格威医疗科技有限公司 | A kind of stages of DR prediction technique |
| CN110010219A (en) * | 2019-03-13 | 2019-07-12 | 杭州电子科技大学 | Optical coherence tomography image retinopathy intelligent checking system and detection method |
| CN111179258A (en) * | 2019-12-31 | 2020-05-19 | 中山大学中山眼科中心 | Artificial intelligence method and system for identifying retinal hemorrhage image |
-
2020
- 2020-07-30 CN CN202010752871.4A patent/CN112053321A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108615051A (en) * | 2018-04-13 | 2018-10-02 | 博众精工科技股份有限公司 | Diabetic retina image classification method based on deep learning and system |
| CN108876775A (en) * | 2018-06-12 | 2018-11-23 | 广州图灵人工智能技术有限公司 | The rapid detection method of diabetic retinopathy |
| CN109191451A (en) * | 2018-09-11 | 2019-01-11 | 百度在线网络技术（北京）有限公司 | Method for detecting abnormality, device, equipment and medium |
| CN109829882A (en) * | 2018-12-18 | 2019-05-31 | 苏州比格威医疗科技有限公司 | A kind of stages of DR prediction technique |
| CN109493954A (en) * | 2018-12-20 | 2019-03-19 | 广东工业大学 | A kind of SD-OCT image retinopathy detection system differentiating positioning based on classification |
| CN110010219A (en) * | 2019-03-13 | 2019-07-12 | 杭州电子科技大学 | Optical coherence tomography image retinopathy intelligent checking system and detection method |
| CN111179258A (en) * | 2019-12-31 | 2020-05-19 | 中山大学中山眼科中心 | Artificial intelligence method and system for identifying retinal hemorrhage image |
Non-Patent Citations (1)
| Title |
|---|
| JUAN WANG: "Simultaneous Diagnosis of Severity and Features of Diabetic Retinopathy in Fundus Photography Using Deep Learning", 《IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS》, vol. 24, no. 12, pages 3397 - 3407, XP011824868, DOI: 10.1109/JBHI.2020.3012547 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI779731B (en) * | 2021-07-21 | 2022-10-01 | 長庚醫療財團法人基隆長庚紀念醫院 | Retinopathy report inspection method using artificial intelligence and retinopathy image integration method |
| CN113706518A (en) * | 2021-09-01 | 2021-11-26 | 吉林大学 | Femoral head necrosis stage analysis method and system based on artificial intelligence |
| CN113558564A (en) * | 2021-09-23 | 2021-10-29 | 中山大学中山眼科中心 | Data processing system based on simple high myopia database construction |
| CN115082414A (en) * | 2022-07-08 | 2022-09-20 | 深圳市眼科医院 | Portable detection method and device based on visual quality analysis |
| CN115082414B (en) * | 2022-07-08 | 2023-01-06 | 深圳市眼科医院 | Portable detection method and device based on visual quality analysis |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112053321A (en) | Artificial intelligence system for identifying high myopia retinopathy | |
| Zapata et al. | Artificial intelligence to identify retinal fundus images, quality validation, laterality evaluation, macular degeneration, and suspected glaucoma | |
| Niemeijer et al. | Automated detection and differentiation of drusen, exudates, and cotton-wool spots in digital color fundus photographs for diabetic retinopathy diagnosis | |
| Trucco et al. | Validating retinal fundus image analysis algorithms: issues and a proposal | |
| AU2014271202B2 (en) | A system and method for remote medical diagnosis | |
| CN111481166A (en) | Automatic identification system based on eye ground screening | |
| Hansen et al. | Diabetic retinopathy screening using digital non‐mydriatic fundus photography and automated image analysis | |
| CN110875092B (en) | Health big data service method and system based on remote fundus screening | |
| Abdel-Ghafar et al. | Progress towards automated detection and characterization of the optic disc in glaucoma and diabetic retinopathy | |
| CN109615633A (en) | Crohn disease assistant diagnosis system and method under a kind of colonoscopy based on deep learning | |
| Yauney et al. | Automated process incorporating machine learning segmentation and correlation of oral diseases with systemic health | |
| CN114694236B (en) | Eyeball motion segmentation positioning method based on cyclic residual convolution neural network | |
| CN112446860B (en) | Automatic screening method for diabetic macular edema based on transfer learning | |
| US20230077125A1 (en) | Method for diagnosing age-related macular degeneration and defining location of choroidal neovascularization | |
| CN111179258A (en) | Artificial intelligence method and system for identifying retinal hemorrhage image | |
| Hu et al. | ACCV: automatic classification algorithm of cataract video based on deep learning | |
| CN110432860B (en) | Method and system for identifying lattice-change cracked holes in wide-area fundus map based on deep learning | |
| CN116705326A (en) | End-stage renal disease curative effect assessment and complication risk prediction system | |
| CN111402184B (en) | Method and system for realizing remote fundus screening and health service | |
| KR102208508B1 (en) | Systems and methods for performing complex ophthalmic tratment | |
| EP4258205A1 (en) | Quality control method and quality control system for data annotation on fundus image | |
| KR102595429B1 (en) | Apparatus and method for automatic calculation of bowel preparation | |
| CN110974151A (en) | Artificial intelligence system and method for identifying retinal detachment | |
| Grau et al. | Teleglaucoma using a new smartphone-based tool for visual field assessment | |
| CN111340778A (en) | Glaucoma image processing method and equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |