Interpretation system and interpretation method for detecting fundus abnormalities based on fundus images

Abstract

The disclosure describes an interpretation system and method for detecting fundus abnormalities based on fundus images. The interpretation system uses the first judgment result of the first interpretation module to divide the fundus images obtained by the acquisition module into negative result images and positive result images. The negative result images are sampled and the second judgment result of the sampled negative result images is obtained by using the second interpretation module. The positive result images are first judged by using the second interpretation module to obtain the second judgment result, then the positive result images are sampled and the third judgment result of the sampled positive result images is obtained by using the third interpretation module. The re-read images are uniformly judged by the fourth interpretation module to obtain the fourth judgment result. The total inspection module is used to perform total inspection on the fundus images based on the first judgment result, the second judgment result, the third judgment result or the fourth judgment result of the fundus images. Thus, the accuracy of interpreting fundus abnormalities in fundus images can be improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to an interpretation system and an interpretation method for detecting fundus abnormalities based on fundus images. BACKGROUND

[0002] Medical images often contain many details of body structures or tissues. In modern hospitals, most of the treatment information comes from medical images such as fundus images. In the clinic, by understanding these details in the medical images, doctors can help identify related diseases. Medical images have developed into the main method for identifying diseases in the clinic. However, the traditional identification of disease information based on medical images mainly relies on professional doctors to make judgments based on experience. In this case, developing a system that can assist doctors in identifying related diseases has become a hot direction in the field of medical imaging. With the development of artificial intelligence technology, systems for identifying diseases based on computer vision and artificial intelligence such as machine learning have been developed and applied in medical image identification.

[0003] For example, patent document 1 (CN105513077A) discloses a system for screening diabetic retinopathy, which includes a fundus image acquisition device, an image processing and screening device, and a report output device. The fundus image acquisition device is used to collect or receive the fundus image of the person being examined. The image processing and screening device is used to process the fundus image and detect whether there is a lesion in it, and then transmit the detection result to the report output device. The report output device outputs the corresponding detection report based on the detection result.

[0004] However, in actual clinical applications, due to the diversity of fundus images, the accuracy of different fundus images on the screening system described in patent document 1 is often inconsistent, resulting in a decrease in the accuracy of interpreting fundus abnormalities in fundus images. SUMMARY

[0005] The present disclosure is proposed in view of the above situation, and aims to provide an interpretation system and an interpretation method for detecting fundus abnormalities based on fundus images, which can improve the accuracy of interpreting fundus abnormalities in fundus images.

[0006] To this end, the first aspect of the present disclosure provides an interpretation system for detecting fundus abnormalities based on fundus images, comprising an acquisition module, a first interpretation module, a first sampling module, a second interpretation module, a second sampling module, a third interpretation module, a fourth interpretation module, and a total inspection and release module; the acquisition module is configured to acquire fundus images; the first interpretation module is configured to judge the fundus images based on a machine learning method to obtain a first judgment result and divide the fundus images into negative result images and positive result images based on the first judgment result; the first sampling module is configured to sample the negative result images and extract the negative result images as sampled negative images; the second interpretation module is configured to receive the positive result images from the first interpretation module and / or the sampled negative images from the first sampling module as quality control images, judge the quality control images to obtain a second judgment result of the quality control images or take the quality control images as first images to be re-read; the second sampling module is configured to sample the positive result images and extract the positive result images as sampled positive images; the third interpretation module is configured to judge the sampled positive images to obtain a third judgment result of the sampled positive images or take the sampled positive images as second images to be re-read; the fourth interpretation module is configured to judge the images to be re-read including the first images to be re-read and the second images to be re-read to obtain a fourth judgment result; and the total inspection and release module is configured to perform total inspection on the fundus images based on the first judgment result, the second judgment result, the third judgment result, or the fourth judgment result of the fundus images to release a final judgment result of the fundus images or re-judge the fundus images.

[0007] In the present disclosure, the first judgment result of the first interpretation module is used to divide the fundus images into negative result images and positive result images, the second judgment result of the sampled negative result images is obtained by sampling the negative result images and using the second interpretation module, and the second judgment result is obtained by first judging the positive result images using the second interpretation module, then sampling the positive result images and using the third interpretation module to obtain the third judgment result of the sampled positive result images. The fundus images that meet the major positive condition obtained by the second interpretation module and the third interpretation module are uniformly judged by the fourth interpretation module. The total inspection and release module performs total inspection on the fundus images based on the first judgment result, the second judgment result, the third judgment result, or the fourth judgment result of the fundus images to release a final judgment result of the fundus images or re-judge the fundus images. In this case, the fundus images are divided into negative result images, positive result images, and fundus images that meet the major positive condition. Different interpretation modules are used to judge the three kinds of fundus images and sample the interpretation modules. Therefore, the accuracy of interpreting fundus abnormalities in fundus images can be improved.

[0008] In addition, in the reading system according to the first aspect of the present disclosure, optionally, the total inspection module is configured to obtain the second judgment result of the negative result image in the preset time period and determine whether the first judgment result of the negative result image is qualified, and if so, take the first judgment result of the negative result image as the final judgment result, otherwise, reset the negative result image by using the re-reading module so that the reading system re-judges the negative result image, the total inspection module is further configured to obtain the third judgment result of the positive result image in the preset time period and determine whether the second judgment result of the positive result image is qualified, and if so, take the second judgment result of the positive result image as the final judgment result, otherwise, reset the positive result image by using the re-reading module so that the reading system re-judges the positive result image, and the total inspection module is further configured to take the third judgment result or the fourth judgment result as the final judgment result, and the release module is configured to release the final judgment result.

[0009] In addition, in the reading system according to the first aspect of the present disclosure, optionally, the total inspection module is configured to obtain the second judgment result of the negative result image in the preset time period and calculate a first sampling error rate, and if the first sampling error rate is greater than a first preset error rate, determine that the first judgment result of the negative result image is unqualified, and the total inspection module is further configured to obtain the third judgment result of the positive result image in the preset time period and calculate a second sampling error rate, and if the second sampling error rate is greater than a second preset error rate, determine that the second judgment result of the positive result image is unqualified. Thus, whether the first judgment result is qualified can be determined based on the first sampling error rate, and whether the second judgment result is qualified can be determined based on the second sampling error rate.

[0010] In addition, in the reading system according to the first aspect of the present disclosure, optionally, the first image to be re-read includes a quality control image meeting a major positive condition, and the second image to be re-read includes a sampling positive image meeting the major positive condition, and the major positive condition is a lesion feature that affects vision or threatens vision. Thus, the major positive condition can be determined, and the image to be re-read can be obtained based on the major positive condition.

[0011] In addition, in the reading system according to the first aspect of the present disclosure, optionally, the reading system further comprises a fifth judgment module configured to confirm and judge the first judgment result of the first judgment module, the second judgment result of the second judgment module, the third judgment result of the third judgment module, or the fourth judgment result of the fourth judgment module to obtain a fifth judgment result, and the final judgment result is the fifth judgment result. Thus, the judgment results of the respective judgment modules can be uniformly confirmed.

[0012] In addition, in the interpretation system according to the first aspect of the present disclosure, optionally, the interpretation system further comprises an over-interpretation module configured to convert the first judgment result of the fundus image with the negative result into a positive result to match a preset requirement of a negative prediction rate; and the over-interpretation module is further configured to set the first judgment result of the fundus image meeting a preset condition as a positive result, the preset condition comprising at least one of a history of fundus disease, corrected visual acuity lower than a preset visual acuity, a history of diabetes, a history of diabetes for a period of time greater than or equal to a first preset time, a history of hypertension, a history of hypertension for a period of time greater than or equal to a second preset time, and recent visual acuity decline. In this case, by setting a higher preset negative prediction rate, the accuracy of the interpretation of the fundus image with the negative result can be improved, and by setting the first judgment result of the fundus image meeting the preset condition as a positive result, the subsequent judgment on the fundus image with the positive result can be facilitated.

[0013] In addition, in the interpretation system according to the first aspect of the present disclosure, optionally, the interpretation system further comprises a quality control module and a recall module, for the judgment result of the fundus image corresponding to the final judgment result that has been released, the quality control module is configured to sample the first interpretation module, the second interpretation module, the third interpretation module or the fourth interpretation module based on a preset period and at least one dimension to determine whether the judgment result of the at least one dimension is qualified, and the recall module is configured to perform recall processing on the fundus image of the at least one dimension that is determined by the quality control module to be unqualified. In this way, the accuracy of the interpretation of the fundus abnormalities in the fundus image can be further improved.

[0014] In addition, in the interpretation system according to the first aspect of the present disclosure, optionally, the first interpretation module establishes different sub-judgment models for each fundus abnormality, the sub-judgment model receives the fundus image and obtains a sub-judgment result, and the first interpretation module obtains the first judgment result based on a plurality of sub-judgment results. In this way, the first judgment result can be obtained based on a plurality of sub-judgment models.

[0015] The second aspect of the present disclosure provides a method for interpreting fundus abnormalities based on fundus images, comprising: obtaining a fundus image; judging the fundus image based on a machine learning method to obtain a first judgment result and dividing the fundus image into negative result images and positive result images based on the first judgment result; sampling the negative result images and taking the sampled negative result images as sampled negative images; taking the positive result images and / or the sampled negative images as quality control images, judging the quality control images to obtain a second judgment result of the quality control images or taking the quality control images as first images to be re-read; sampling the positive result images and taking the sampled positive result images as sampled positive images; judging the sampled positive images to obtain a third judgment result of the sampled positive images or taking the sampled positive images as second images to be re-read; judging the images to be re-read including the first images to be re-read and the second images to be re-read to obtain a fourth judgment result; and performing a final judgment on the fundus image based on the first judgment result, the second judgment result, the third judgment result or the fourth judgment result of the fundus image to release the final judgment result of the fundus image or re-judge the fundus image.

[0016] In the present disclosure, the fundus image is divided into negative result images and positive result images by the first judgment result, the second judgment result of the sampled negative result images is obtained by sampling the negative result images, and the second judgment result is obtained by judging the positive result images first, then the positive result images are sampled and the third judgment result of the sampled positive result images is obtained, the fundus images meeting the major positive condition are uniformly judged, the final judgment result of the fundus image is released or the fundus image is re-judged based on the first judgment result, the second judgment result, the third judgment result or the fourth judgment result of the fundus image. In this case, the fundus image is divided into three kinds of fundus images, i.e. negative result images, positive result images and fundus images meeting the major positive condition, and the three kinds of fundus images are judged and sampled accordingly. Thus, the accuracy of interpreting fundus abnormalities in the fundus image can be improved.

[0017] In addition, in the interpretation method according to the second aspect of the present disclosure, in the total inspection, a second judgment result of a negative result image in a preset time period is obtained, and it is determined whether the first judgment result of the negative result image is qualified. If yes, the first judgment result of the negative result image is taken as the final judgment result. Otherwise, the negative result image is reset for re-judgment. In the total inspection, a third judgment result of a positive result image in the preset time period is obtained, and it is determined whether the second judgment result of the positive result image is qualified. If yes, the second judgment result of the positive result image is taken as the final judgment result. Otherwise, the positive result image is reset for re-judgment. In the total inspection, the third judgment result or the fourth judgment result is taken as the final judgment result.

[0018] According to the present disclosure, an interpretation system and an interpretation method for detecting fundus abnormalities based on fundus images can improve the accuracy of interpreting fundus abnormalities in fundus images. BRIEF DESCRIPTION OF DRAWINGS

[0019] The present disclosure will now be explained in further detail by way of example only with reference to the drawings wherein:

[0020] Figure 1 FIG. 1 is a diagram illustrating an application scenario of an interpretation method for detecting fundus abnormalities based on fundus images according to an example of the present disclosure.

[0021] Figure 2 FIG. 2 is a block diagram illustrating an interpretation system for detecting fundus abnormalities based on fundus images according to an example of the present disclosure.

[0022] Figure 3 FIG. 3 is a schematic diagram illustrating a fundus image according to an example of the present disclosure.

[0023] Figure 4 FIG. 4 is a block diagram illustrating a total inspection and release module for detecting fundus abnormalities based on fundus images according to an example of the present disclosure.

[0024] Figure 5 FIG. 5 is a block diagram illustrating an interpretation system for detecting fundus abnormalities based on fundus images according to an example of the present disclosure.

[0025] Figure 6 FIG. 6 is a flowchart illustrating an interpretation method for detecting fundus abnormalities based on fundus images according to an example of the present disclosure.

[0026] Figure 7 FIG. 7 is a flowchart illustrating an interpretation method for detecting fundus abnormalities based on fundus images according to an example of the present disclosure. DETAILED DESCRIPTION

[0027] The preferred embodiments of this disclosure are described in detail below with reference to the accompanying drawings. In the following description, the same reference numerals are used for the same components, and repeated descriptions are omitted. Furthermore, the drawings are merely schematic diagrams, and the proportions of the components or the shapes of the components may differ from actual figures. It should be noted that the terms "comprising" and "having," and any variations thereof, in this disclosure, do not necessarily limit the process, method, system, product, or apparatus to the explicitly listed steps or units, but may include or have other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. All methods described in this disclosure may be performed in any suitable order unless otherwise indicated herein or clearly contradicted by the context.

[0028] This disclosure relates to a fundus image-based system 200 and a method for interpreting fundus abnormalities, which can improve the accuracy of interpreting fundus abnormalities in fundus images. The fundus image-based system 200 is sometimes simply referred to as the interpretation system 200. The fundus image-based method is sometimes simply referred to as the interpretation method.

[0029] Figure 1 This diagram illustrates an application scenario of the interpretation method for detecting fundus abnormalities based on fundus images, as described in this disclosure.

[0030] In some examples, the interpretation method (described later) can be applied to, for example... Figure 1 In application scenario 100, operator 110 can control acquisition device 130 connected to terminal 120 to acquire fundus images of the fundus of human eye 140. After acquisition device 130 completes fundus image acquisition, terminal 120 can submit the fundus images to server 150 via computer network. Server 150 can execute computer program instructions stored on server 150 to implement an interpretation method. This interpretation method can receive fundus images and generate a fundus image result report. Server 150 can return the generated fundus image result report to terminal 120. In some examples, terminal 120 can display the result report. In other examples, the result report can be stored as an intermediate result in the memory of terminal 120 or server 150. In some examples, the result report can be displayed on a user's end for the user to view, such as a patient. In other examples, the fundus images received by the interpretation method can be fundus images stored in terminal 120 or server 150.

[0031] In some examples, the operator 110 can be a professional such as an ophthalmologist. In other examples, the operator 110 can be a layperson trained in interpretation. The interpretation training can include, but is not limited to, operation of the acquisition device 130 and operation of the terminal 120 involving the interpretation method. In some examples, the terminal 120 can include, but is not limited to, a laptop, a tablet, or a desktop computer, etc. In some examples, the acquisition device 130 can be a camera. The camera can be, for example, a handheld fundus camera or a table-top fundus camera. In some examples, the acquisition device 130 can be connected to the terminal 120 through a serial port. In some examples, the acquisition device 130 can be integrated in the terminal 120.

[0032] In some examples, the server 150 can include one or more processors and one or more memories. The processors can include, but are not limited to, central processing units, graphics processing units, and any other electronic components capable of processing data and capable of executing computer program instructions. The memories can be used to store computer program instructions. In some examples, the server 150 can implement the interpretation method by executing the computer program instructions on the memories. In some examples, the server 150 can also be a cloud server.

[0033] The interpretation system 200 according to the present disclosure is described in detail below in conjunction with the accompanying drawings. The interpretation system 200 according to the present disclosure is used to implement the interpretation method described above. Figure 2 is a block diagram showing an interpretation system 200 for detecting fundus abnormalities based on fundus images according to an example of the present disclosure.

[0034] As shown in FIG. 1, the interpretation system 200 according to the present disclosure can include an acquisition device 130, a terminal 120, and a server 150. Figure 2As shown, in some examples, the interpretation system 200 can include an acquisition module 210, a first interpretation module 220, a first sampling module 230, a second interpretation module 240, a second sampling module 250, a third interpretation module 260, a fourth interpretation module 270, and a total review module 280. The acquisition module 210 can be configured to acquire the fundus image. The first interpretation module 220 can be configured to interpret the fundus image based on a deep learning method to obtain a first interpretation result and divide the fundus image into a negative result image and a positive result image based on the first interpretation result. The first sampling module 230 can be configured to sample the negative result image as a sampled negative image. The second interpretation module 240 can be configured to interpret a quality control image including the sampled negative image and / or the positive result image to obtain a second interpretation result of the quality control image or take the quality control image as a first image to be re-read. The second sampling module 250 can be configured to sample the positive result image as a sampled positive image. The third interpretation module 260 can be configured to interpret the sampled positive image to obtain a third interpretation result of the sampled positive image or take the sampled positive image as a second image to be re-read. The fourth interpretation module 270 can be configured to interpret the fundus image meeting a major positive condition to obtain a fourth interpretation result. The total review module 280 can be configured to review the fundus image based on the interpretation results of the fundus image in each interpretation module to release a final interpretation result of the fundus image or re-interpret the fundus image. In this case, the fundus image is divided into three fundus images of the negative result image, the positive result image, and the fundus image meeting the major positive, the different interpretation modules are used to interpret the three fundus images, and the interpretation modules are sampled. Thus, the accuracy of interpreting the fundus abnormalities in the fundus image can be improved.

[0035] Figure 3 FIG. 1 is a schematic diagram illustrating a fundus image related to examples of the present disclosure.

[0036] In some examples, the acquisition module 210 can be configured to acquire the fundus image. In some examples, the fundus image can include fundus images from the left eye and the right eye of the user. In some examples, the fundus image can be a color fundus image. The color fundus image can clearly present the rich fundus information such as the optic disc, the optic cup, the macula, and blood vessels. In addition, the fundus image can be an image in RGB mode, CMYK mode, Lab mode, or grayscale mode, etc. In some examples, the fundus image can be acquired by the acquisition device 130. In other examples, the fundus image can be a fundus image stored in the terminal 120 or the server 150. As an example of the fundus image, for example Figure 3 FIG. 2 is a fundus image of a human eye 140.

[0037] In some examples, after the acquisition module 210 acquires the fundus image, the fundus image can be preprocessed. In some examples, the preprocessing can include cropping the fundus image. Generally, since the fundus images acquired by the acquisition module 210 can have different image formats or sizes, the fundus images need to be cropped to convert them into images of a fixed standard form. The fixed standard form can mean that the images have the same format and consistent size. For example, in some examples, the size of the fundus image after preprocessing can be unified to 256x256, 374x374, 512x512, 768x768, or 1024x1024 pixels.

[0038] In addition, in some examples, the preprocessing can include normalizing the fundus image. In some examples, the normalization can include operations such as coordinate centering and scaling normalization on the fundus image. In this way, the differences between different fundus images can be overcome, and the performance of the interpretation system 200 can be improved. In addition, in some examples, the preprocessing can include noise reduction, grayscale processing, and other processing of the fundus image. In this way, the features of the fundus image can be highlighted. In some examples, the fundus image can also not be preprocessed, and the fundus image can be directly judged later.

[0039] In some examples, the first interpretation module 220 can be used to judge whether the fundus image from the acquisition module 210 has fundus abnormalities. In some examples, the first interpretation module 220 can output a first judgment result. In some examples, the first judgment result can include a negative result or a positive result. In some examples, the first judgment result can also include the presence of fundus abnormalities.

[0040] In some examples, the fundus abnormalities can include at least one of suspected glaucoma, optic nerve developmental abnormalities, optic nerve diseases, retinal vascular lesions, retinal vascular sclerosis, retinal vascular obstruction, leopard-shaped fundus, macular abnormalities, macular hole, and diabetic retinopathy. In this way, a variety of fundus abnormalities can be evaluated. However, the interpretation system 200 of the present disclosure is not limited to the types of fundus abnormalities and can be easily extended to the evaluation of other fundus abnormalities. In some examples, different fundus abnormalities can correspond to different lesion characteristics.

[0041] In some examples, the first interpretation module 220 can use a machine learning method to judge the fundus image to obtain the first judgment result. In some examples, the machine learning method can be at least one of a traditional machine learning method and a deep learning method.

[0042] In some examples, the deep learning based method can be based on a convolutional neural network (CNN). In this case, since the convolutional neural network (CNN) has the characteristic of high efficiency in identifying image features, the performance of the interpretation system 200 can be effectively improved.

[0043] However, examples of the present disclosure are not limited thereto, and in other examples, the machine learning method of the first interpretation module 220 can be a traditional machine learning method. In some examples, the traditional machine learning method can include, but is not limited to, a linear regression algorithm, a logistic regression algorithm, a decision tree algorithm, a support vector machine algorithm, or a Bayesian algorithm, etc. In this case, the fundus features in the fundus image can be extracted using an image processing algorithm first, and then the fundus features are input into a judgment model established based on the traditional machine learning method to realize judgment on the fundus image.

[0044] In some examples, the first interpretation module 220 can establish a judgment model based on the machine learning method, and the judgment model can include a plurality of sub-judgment models. Each sub-judgment model can be for each fundus abnormality. Each sub-judgment model can receive the fundus image and obtain a sub-judgment result. In some examples, the first interpretation module 220 can obtain the first judgment result based on a plurality of sub-judgment results. Thus, the first judgment result can be obtained based on a plurality of sub-judgment models.

[0045] Specifically, different sub-judgment models can be established and trained for each fundus abnormality to obtain a sub-judgment result (e.g., a negative result or a positive result) of whether it is a certain fundus abnormality and a confidence corresponding to the sub-judgment result, and then the first judgment result can be obtained according to each sub-judgment result. For example, if there is no any fundus abnormality, it is a negative result, otherwise it is a positive result.

[0046] In some examples, the first interpretation module 220 can classify the fundus image with the first judgment result as a negative result as a negative result image, and classify the fundus image with the first judgment result as a positive result as a positive result image.

[0047] In some examples, the interpretation system 200 can include an overfitting module (not shown). In some examples, the first interpretation module 220 can use the overfitting module to convert the first judgment result of the fundus image with the first judgment result as a negative result to a positive result, and classify the fundus image into a negative result image and a positive result image based on the converted first judgment result.

[0048] In some examples, the over-translation module can be configured to convert the first judgment result of the fundus image with the first judgment result as a negative result to a positive result to match the requirement of the preset negative prediction rate. Generally, the higher the preset negative prediction rate, the more sensitive the first judgment module 220 is to the negative result, and the first judgment result of the fundus image with an uncertain negative result can be set as a positive result. In this case, by setting a higher preset negative prediction rate, the accuracy of the judgment of the fundus image with a negative result can be improved. In some examples, the preset negative prediction rate can be 95% to 99%. For example, the preset negative prediction rate can be 95%, 96%, 97%, 98%, or 99%, etc.

[0049] Generally, the division of the negative result and the positive result by the over-translation module can be based on the confidence, and thus the negative prediction rate can be adjusted by setting the confidence threshold to match the requirement of the preset negative prediction rate. Specifically, the confidence threshold can be inversely solved according to the preset negative prediction rate. Since the sensitivity, specificity, positive prediction rate (the positive prediction rate can be the number of true positives / (the number of true positives + the number of false positives)), and the negative prediction rate (the negative prediction rate can be the number of true negatives / (the number of true negatives + the number of false negatives)) and other performance indicators corresponding to each confidence threshold are determined on the gold standard data. In some examples, all confidence thresholds can be traversed in a preset step size and the related performance indicators can be solved. For example, the preset step size can be 0.01, 0.001, or 0.0001, etc. In this case, a table recording the performance indicators corresponding to each confidence threshold can be obtained, and the confidence threshold corresponding to the preset negative prediction rate in the table is the confidence threshold. In some examples, when calculating the confidence threshold, the correspondence between the confidence threshold and the preset negative prediction rate can be calculated, a curve graph can be formed, and the confidence threshold corresponding to the preset negative prediction rate can be obtained according to the curve graph.

[0050] In some examples, the confidence threshold can be for each sub-judgment model. That is, the sub-judgment result of the fundus image with the sub-judgment result as a negative result is converted to a positive result based on the confidence threshold. For example, assuming that the sub-judgment result of a fundus image is that there is no glaucoma (i.e., a negative result) and the confidence is 90%, if the confidence threshold is 92%, since the confidence of the sub-judgment result is less than 92%, the sub-judgment result of the fundus image can be set as the existence of glaucoma. In this case, the confidence threshold can be applied to the sub-judgment result of each sub-judgment model to affect the first judgment result.

[0051] In some examples, the over-fitting module is further configured to set the first determination result of the fundus image meeting a preset condition as a positive result. In some examples, the preset condition can include at least one of a history of fundus disease, corrected visual acuity lower than a preset visual acuity, a history of diabetes or a history of diabetes for a period of time greater than or equal to a first preset period of time, a history of hypertension or a history of hypertension for a period of time greater than or equal to a second preset period of time, and recent visual acuity decline. In this way, subsequent determination for the fundus image with a positive result can be facilitated. In some examples, the preset visual acuity can be 0.6. In some examples, the first preset period of time can be 5 to 10 years. In some examples, the second preset period of time can be 5 to 10 years. In some examples, the recent visual acuity decline can be a recent visual acuity condition provided by the user, such as whether there has been significant visual blurring recently.

[0052] In some examples, the first sampling module 230 can be configured to sample the negative result images and extract the sampled negative result images as sampled negative images. Specifically, the first sampling module 230 can sample the negative result images according to a preset time (e.g., every hour) or a preset number (e.g., 100 images). In some examples, the extracted negative result images can be labeled to distinguish the extracted negative result images and the extracted negative result images can be used as the sampled negative images. In some examples, the negative result images that are not extracted can be input to the total pass and release module 280 (described later) for total inspection.

[0053] In some examples, the fundus images can be sampled based on a sampling scheme. In some examples, the sampling scheme can include uniformly sampling the fundus images. Specifically, one sampling method can be performed on all fundus images. In some examples, the sampling method can include, but is not limited to, random sampling, stratified sampling, whole sampling, or systematic sampling, etc. In some examples, the sampling method can be based on a national standard for sampling inspection. In some examples, the national standard for sampling inspection can include, but is not limited to, GB / T2828.1-2012, GB / T2828.5-2011, or GB / T8052-2002, etc.

[0054] In some examples, the sampling scheme can include classifying the fundus images according to one or more different dimensions, and sampling separately for each class (e.g. several dimensions can be a class). In some examples, the dimensions can include at least one of a source of the fundus image, a model of a photographing device of the fundus image, a person who photographs the fundus image, a photographing quality of the fundus image, and an interpreter information of the fundus image. In some examples, the judging model established by the machine learning based method can be regarded as an interpreter, and if there are multiple judging models, they can correspond to multiple interpreters. In this case, the fundus images can be classified according to different dimensions and sampled separately for each class. In this way, the reliability of sampling can be improved.

[0055] In some examples, the sampling scheme can be adjusted according to the dimensions or the release time length of the interpretation system 200 in sampling. In some examples, the sampling scheme can have different sampling degrees. In some examples, the sampling degrees can include general, special, more stringent, or more relaxed, etc. In some examples, different sampling degrees can correspond to different sampling methods or different sampling parameters. In some examples, the sampling parameters can include but are not limited to sampling proportion or AQL (Acceptable Quality Limit) value, etc.

[0056] In some examples, a preliminary sampling proportion can be obtained, and the preliminary sampling proportion can be adjusted based on the sampling degree to determine the sampling proportion. For example, if the sampling degree is relaxed, the preliminary sampling proportion can be reduced to determine the sampling proportion.

[0057] Generally, AQL refers to the worst average process quality level that can be allowed when a continuous series of batches is submitted to acceptance sampling. If the quality level is expressed in terms of percentage of nonconforming units, the AQL value should not exceed 10%, and if the quality level is expressed in terms of nonconforming units per hundred units of the fundus images to be sampled, the AQL value should not exceed 10. Different sampling degrees can correspond to different AQL values.

[0058] In some examples, the sampling scheme can be adjusted according to the release time length of the interpretation system 200 in sampling. For example, the sampling degree of the interpretation system 200 just released can be more stringent, and after a period of operation, if it is determined that the interpretation accuracy of the interpretation system 200 meets the requirements, the sampling degree can be adjusted to be relaxed. In some examples, the sampling scheme can be adjusted according to the dimensions in sampling. For example, the sampling degree of the fundus images interpreted by the interpreters with lower seniority can be more stringent.

[0059] In some examples, the second interpretation module 240 can be used to receive a positive result image from the first interpretation module 220 and a negative sample image from the first sampling module 230. That is, the positive result image from the first interpretation module 220 and the negative sample image from the first sampling module 230 can be input into the second interpretation module 240. In some examples, the positive result image from the first interpretation module 220 and the negative sample image from the first sampling module 230 can be used as quality control images.

[0060] In some examples, the second interpretation module 240 can interpret the quality control image to obtain a second interpretation result. The second interpretation result may include a negative result or a positive result. In some examples, the second interpretation result may also include the presence of fundus abnormalities. In some examples, the second interpretation module 240 can interpret the quality control image to use it as the first image to be reread.

[0061] In some examples, the first image to be reread may include a quality control image that meets the major positive criteria. In some examples, the major positive criteria may be the presence of a lesion feature that affects or threatens visual acuity. Thus, it is possible to determine the major positive criteria and obtain the first image to be reread based on the major positive criteria. In some examples, the lesion feature may include at least one of the following: cup-to-disc ratio greater than or equal to a preset cup-to-disc ratio, anterior disc membrane, myelinated nerve fibers, other optic nerve abnormalities, cotton wool spots, hard exudates, retinal vascular sclerosis, retinal vascular occlusion, leopard-spot fundus, macular pigmentary abnormalities, macular hole, drusen, hemorrhage, proliferative membrane, other retinal abnormalities, non-proliferative diabetic retinopathy, and proliferative diabetic retinopathy. In some examples, the preset cup-to-disc ratio may be 0.6. In some examples, the lesion feature that affects or threatens visual acuity may include at least one of macular hole, retinal vascular occlusion, and proliferative diabetic retinopathy. Thus, it is possible to determine the presence of a lesion feature that affects or threatens visual acuity. However, the examples in this disclosure are not limited to these; in other examples, the criteria for meeting the major positive criteria may be defined according to the specific circumstances.

[0062] Specifically, the second interpretation module 240 can acquire and display basic information about the quality control image. This basic information may include, but is not limited to, eye identification (e.g., left or right eye), creation time, affiliated medical institution, camera model, and the current interpretation module (e.g., the second interpretation module 240). After the interpreter assesses the quality control image based on the basic information and generates assessment information, the second interpretation module 240 can receive this assessment information. If the quality control image does not meet the criteria for a major positive result, the second interpretation module 240 can generate a second assessment result for the quality control image based on the assessment information; otherwise, it can mark the quality control image as the first image to be re-read. In some examples, the basic information may also include medical history, visual acuity information, gender, and age.

[0063] In some examples, the judgment information can include, but is not limited to, the existing fundus abnormalities, the image quality level of the fundus image, the existing lesion characteristics, etc. In some examples, the image quality level can be five levels of very good, good, general, poor and very poor.

[0064] In some examples, the second sampling module 250 can be configured to sample the positive result images and extract the sampled positive result images as the sampled positive images. Specifically, the second sampling module 250 can sample the positive result images according to a preset time (e.g., every hour) or a preset number (e.g., 100 images). In some examples, the extracted positive result images can be marked to distinguish the extracted positive result images and the extracted positive result images can be regarded as the sampled positive images. In some examples, the positive result images that are not extracted can be input into the total review module 280 (to be described later) for total review.

[0065] In some examples, the fundus images can be sampled based on the sampling scheme. For details, please refer to the relevant description of the sampling scheme in the first sampling module 230. In some examples, the second sampling module 250 can be configured to sample the positive result images with the second judgment result and extract the sampled positive result images as the sampled positive images (i.e., the positive result images meeting the major positive condition are not sampled).

[0066] In some examples, the third reading module 260 can be configured to receive the sampled positive images from the second sampling module 250. In some examples, the sampled positive images can be judged to obtain the third judgment result of the sampled positive images. The third judgment result can include a negative result or a positive result. In some examples, the third judgment result can also include the existing fundus abnormalities. In some examples, the sampled positive images can be judged to regard the sampled positive images as the second images to be re-read. In some examples, the second images to be re-read can include the sampled positive images meeting the major positive condition. In some examples, the seniority of the reading personnel of the third reading module 260 can be higher than that of the reading personnel of the second reading module 240.

[0067] Specifically, the third reading module 260 can obtain the corresponding above-mentioned basic information of the sampled positive images, wherein the current reading module in the basic information can be the third reading module 260. After the reading personnel judges the sampled positive images based on the basic information and generates the judgment information, the third reading module 260 can receive the judgment information of the reading personnel. If the sampled positive images do not meet the major positive condition, the third reading module 260 can generate the third judgment result of the sampled positive images based on the judgment information, otherwise the sampled positive images can be marked as the second images to be re-read.

[0068] In some examples, the fourth reading module 270 can be configured to read the fundus images that meet the significant positive condition to obtain a fourth reading result. In some examples, the fundus images that meet the significant positive condition can be the first images to be re-read from the second reading module 240 and the second images to be re-read from the third reading module 260. In some examples, the fundus images that meet the significant positive condition can be read as images to be re-read and unified by the fourth reading module 270. In some examples, the fourth reading module 270 can read the images to be re-read to obtain the fourth reading result. In some examples, the fourth reading result can include a negative result or a positive result. In some examples, the fourth reading result can also include the existing fundus abnormalities. In some examples, the seniority of the readers of the fourth reading module 270 can be higher than the seniority of the readers of the third reading module 260.

[0069] Specifically, the fourth reading module 270 can obtain the corresponding basic information of the images to be re-read, wherein the current reading module in the basic information can be the fourth reading module 270. After the readers read the images to be re-read based on the basic information and generate reading information, the fourth reading module 270 can receive the reading information of the readers and generate the fourth reading result of the images to be re-read based on the reading information.

[0070] In some examples, the total inspection and release module 280 can perform total inspection on the fundus images based on the reading results of the fundus images in each reading module to release the final reading result of the fundus images. In some examples, the total inspection and release module 280 can perform total inspection on the fundus images based on the reading results of the fundus images in each reading module to re-read the fundus images. In some examples, the reading results of each reading module can include the first reading result of the first reading module 220, the second reading result of the second reading module 240, the third reading result of the third reading module 260, or the fourth reading result of the fourth reading module 270.

[0071] Figure 4 is a block diagram illustrating the total inspection and release module 280 for detecting fundus abnormalities based on fundus images according to examples of the present disclosure.

[0072] As shown in Figure 4 In some examples, the total inspection and release module 280 can include a total inspection module 281. The total inspection module 281 can be configured to determine whether the reading result of the fundus images is qualified.

[0073] In some examples, the total detection module 281 can be configured to determine whether the first determination result of the negative result image is qualified. As described above, the second determination module 240 can determine the quality control image from the positive result image of the first determination module 220 and the sampled negative image of the first sampling module 230, and determine the second determination result of the quality control image. Therefore, the sampled negative image has the second determination result (i.e., the extracted negative result image has the second determination result). Specifically, the total detection module 281 can obtain the second determination result of the negative result image in a preset time period (e.g., the same day, the morning or the afternoon), and determine whether the first determination result of the negative result image in at least one dimension (e.g., the information of the interpreter) is qualified according to the at least one dimension. If the first determination result of the negative result image in the at least one dimension is qualified, the first determination result of the negative result image in the at least one dimension is taken as the final determination result. Otherwise, the negative result image in the at least one dimension is reset so that the reading system 200 re-determines the negative result image in the at least one dimension. In some examples, the negative result image in the at least one dimension can be reset by the re-reading module 282 (to be described later).

[0074] In some examples, the determination of whether the first determination result of the negative result image in the at least one dimension is qualified can be based on the first sampling error rate. Specifically, the total detection module 281 can obtain the second determination result of the negative result image in a preset time period, and calculate the first sampling error rate according to the at least one dimension. If the first sampling error rate is greater than the first preset error rate, the first determination result of the negative result image in the at least one dimension is determined to be unqualified. Thus, the first determination result can be determined to be qualified or unqualified based on the first sampling error rate. In some examples, the first preset error rate can be set according to actual conditions.

[0075] In some examples, if the first determination result of the negative result image is inconsistent with the second determination result, the first determination result of the negative result image is an erroneous determination result.

[0076] In some examples, the total detection module 281 can be configured to determine whether the second determination result of the positive result image is qualified. As described above, the sampled positive image can be determined to obtain the third determination result of the sampled positive image. Thus, the sampled positive image has the third determination result (i.e., the extracted positive result image has the third determination result). Specifically, the total detection module 281 can obtain the third determination result of the positive result image in a preset time period (e.g., a day, a morning or an afternoon), and determine whether the second determination result of the positive result image in at least one dimension (e.g., the interpreter information) is qualified according to the at least one dimension. If the second determination result of the positive result image in the at least one dimension is qualified, the second determination result of the positive result image in the at least one dimension is taken as the final determination result. Otherwise, the positive result image in the at least one dimension is reset so that the reading system 200 re-determines the positive result image in the at least one dimension. In some examples, the positive result image in the at least one dimension can be reset by using the re-reading module 282 (to be described later).

[0077] In some examples, determining whether the second determination result of the positive result image in the at least one dimension is qualified can be based on the second sampling error rate. Specifically, the total detection module 281 can obtain the third determination result of the positive result image in a preset time period and calculate the second sampling error rate according to the at least one dimension. If the second sampling error rate is greater than the second preset error rate, it is determined that the second determination result of the positive result image in the at least one dimension is unqualified. Thus, it can be determined whether the second determination result is qualified based on the second sampling error rate. In some examples, the second preset error rate can be set according to actual conditions.

[0078] In some examples, if the second determination result of the positive result image is inconsistent with the third determination result, the second determination result of the positive result image is an erroneous determination result.

[0079] In some examples, the first sampling error rate or the second sampling error rate described above can be the sampling error number divided by the sampling total number. As described above, the sampling scheme can include classifying the fundus images according to one or more different dimensions, and sampling each category separately. In this case, the total detection module 281 can obtain the total number of fundus images, the sampling total number and the sampling error number of each category. The sampling total number can be the number of the extracted fundus images in each category of fundus images, and the sampling error number can be the number of the fundus images corresponding to the erroneous determination result in each category of fundus images.

[0080] In some examples, the sampling scheme can be adjusted based on the sampling error rate. For example, if the sampling error rate corresponding to a certain dimension is low, the sampling degree of the sampling scheme can be adjusted to be relaxed.

[0081] As described above, the total check module 281 can determine whether the first determination result or the second determination result is qualified and obtain the final determination result. In some examples, the total check module 281 can also take the third determination result as the final determination result. In some examples, the total check module 281 can also take the fourth determination result as the final determination result.

[0082] In other examples, the total check module 281 can also not determine whether the first determination result of the negative result image or the second determination result of the positive result image is qualified according to at least one dimension. Specifically, the total check module 281 can obtain the second determination result of the negative result image in a preset time period and determine whether the first determination result of the negative result image is qualified, if qualified, take the first determination result of the negative result image as the final determination result, otherwise reset the negative result image by using the re-reading module 282 so that the reading system 200 re-determines the negative result image. The total check module 281 can also obtain the third determination result of the positive result image in a preset time period and determine whether the second determination result of the positive result image is qualified, if qualified, take the second determination result of the positive result image as the final determination result, otherwise reset the positive result image by using the re-reading module 282 so that the reading system 200 re-determines the positive result image. In this case, the first sampling error rate and the second sampling error rate described above can also not be calculated according to at least one dimension.

[0083] In some examples, the total check and release module 280 can also include a re-reading module 282 (see Figure 4 ). The re-reading module 282 can be used to reset the fundus image corresponding to the unqualified determination result so that the reading system 200 re-determines the fundus image corresponding to the unqualified determination result. In some examples, resetting the fundus image corresponding to the unqualified determination result can be resetting the relevant information of the unqualified fundus image (here, the fundus image corresponding to the determination result of the unqualified fundus image) so that the acquisition module 210 can acquire the fundus image. For example, if the acquisition module 210 acquires the fundus image with a processing state to be determined, the processing state of the unqualified fundus image can be reset to be determined. For another example, if the acquisition module 210 acquires the fundus image from a specific data set, the unqualified fundus image can be re-inserted into the data set. For another example, if the acquisition module 210 acquires the fundus image by message, the message can be re-sent to notify the acquisition module 210 to acquire the unqualified fundus image.

[0084] In some examples, the total check and release module 280 can also include a release module 283 (see Figure 4). In some examples, the release module 283 can be configured to release the final determination result. In some examples, the release module 283 can be configured to display the final determination result and release the result report.

[0085] Specifically, the release module 283 can acquire the above-mentioned basic information of the fundus image, the final determination result, the interpretation personnel information corresponding to the final determination result, the result report and the interpretation details, and receive a release command of a release personnel to release the result report. In some examples, the interpretation details can include the determination results of the respective interpretation modules.

[0086] In some examples, the result report can include the final determination result. In some examples, the released result report can be displayed on a user terminal for a user to view. In some examples, the result report can include user information, medical history, fundus images, image quality levels of the fundus images, comprehensive assessment results, comprehensive suggestions, review personnel information and details of fundus abnormalities based on the final determination result. In this way, the result report can be output. In some examples, the user information can include, but is not limited to, name, gender, age, etc. In addition, in some examples, the medical history can include, but is not limited to, diabetes history, hypertension history and eye disease history, etc. In addition, in some examples, the fundus images in the result report can include fundus images of the left eye and fundus images of the right eye. In addition, in some examples, the comprehensive assessment results can include assessments of the left eye and assessments of the right eye. As an example of the comprehensive assessment results, for example, the comprehensive assessment results can be that the left eye fundus has no obvious abnormalities and the right eye fundus has abnormalities. In some examples, the details of the fundus abnormalities based on the final determination result can include at least one of medical explanation, common cause and guidance suggestion.

[0087] In some examples, the interpretation system 200 further includes a fifth interpretation module (not shown). The fifth interpretation module can be configured to confirm and determine the respective determination results to obtain a fifth determination result. In some examples, the respective determination results can include the first determination result of the first interpretation module 220, the second determination result of the second interpretation module 240, the third determination result of the third interpretation module 260 or the fourth determination result of the fourth interpretation module 270. Specifically, the fifth interpretation module can be configured to check the respective determination results and confirm whether to retain the determination result of the fundus image in the last interpretation module (for example, the last interpretation module of the unextracted negative result image is the first interpretation module 220) as the fifth determination result, and if not, to re-determine to obtain the fifth determination result. In this way, the determination results of the respective interpretation modules can be uniformly confirmed. In some examples, the fifth determination result can include a negative result or a positive result. In some examples, the fifth determination result can also include the existing fundus abnormalities.

[0088] In some examples, the final judgment result can be the fifth judgment result. Specifically, if the final inspection module 281 determines that the first judgment result or the second judgment result of the fundus image is qualified, the final inspection module 281 can use the fifth judgment result corresponding to the qualified fundus image as the final judgment result. In some examples, the final inspection module 281 can directly use the fifth judgment result as the final judgment result.

[0089] Figure 5 This is a block diagram illustrating an interpretation system 200 for detecting fundus abnormalities based on fundus images, as described in this disclosure example.

[0090] like Figure 5 As shown, in some examples, the interpretation system 200 may further include a quality control module 290 and a recall module 300. In some examples, for the judgment results of fundus images corresponding to the final judgment results released by the release module 283, the quality control module 290 may sample each interpretation module (e.g., the first interpretation module 220, the second interpretation module 240, the third interpretation module 260, or the fourth interpretation module 270) based on a preset period (e.g., each month or each quarter) and at least one dimension to determine whether the judgment results of at least one dimension are qualified. The recall module 300 may be used to perform recall processing on fundus images of at least one dimension that have been determined to be unqualified by the quality control module 290. In some examples, the recalled fundus images may undergo full inspection, the sampling ratio of the dimensions corresponding to the recalled fundus images may be increased, and at least one activity in the hardware and software of the interpretation system 200 may be inspected. This can further improve the accuracy of interpreting fundus abnormalities in fundus images.

[0091] The following, combined with Figure 6 This disclosure describes in detail the method for interpreting fundus abnormalities detected based on fundus images. The method for interpreting fundus abnormalities detected based on fundus images disclosed herein may sometimes be simply referred to as the interpretation method. The interpretation method is applied in the aforementioned interpretation system 200. Figure 6 This is a flowchart illustrating the interpretation method for detecting fundus abnormalities based on fundus images, as described in this disclosure.

[0092] In some examples, such as Figure 6As shown, the interpretation method can include obtaining fundus images (step S110), judging the fundus images based on a deep learning method to obtain a first judgment result and dividing the fundus images into negative result images and positive result images based on the first judgment result (step S120), extracting the negative result images as sampled negative images (step S130), judging the quality control images including the sampled negative images and / or the positive result images to obtain a second judgment result of the quality control images or taking the quality control images as first images to be re-read (step S140), extracting the positive result images as sampled positive images (step S150), judging the sampled positive images to obtain a third judgment result of the sampled positive images or taking the sampled positive images as second images to be re-read (step S160), judging the fundus images meeting the major positive condition to obtain a fourth judgment result (step S170), and performing a general review of the fundus images based on the respective judgment results of the fundus images to obtain a final judgment result of the fundus images or re-judging the fundus images (step S180). In this case, the fundus images are divided into three kinds of fundus images, i.e. the negative result images, the positive result images and the fundus images meeting the major positive condition, and the three kinds of fundus images are judged and sampled accordingly. Thus, the accuracy of interpreting the fundus abnormalities in the fundus images can be improved.

[0093] In some examples, in step S110, the fundus images can be obtained. For specific description, refer to the related description of the obtaining module 210.

[0094] In some examples, in step S120, the deep learning method can be used to judge whether the fundus images have fundus abnormalities and output the first judgment result. In some examples, the first judgment result can include a negative result or a positive result. In some examples, the fundus images with the first judgment result as the negative result can be taken as the negative result images, and the fundus images with the first judgment result as the positive result can be taken as the positive result images. For specific description, refer to the related description of the first interpretation module 220.

[0095] In some examples, the first determination result can be over-converted. In some examples, in the over-conversion, the first determination result of the fundus image with the negative result can be converted to a positive result, and the fundus image can be divided into the negative result image and the positive result image based on the converted first determination result. In some examples, the first determination result of the fundus image with the negative result can be converted to a positive result to match the requirement of the preset negative prediction rate. In this case, by setting a higher preset negative prediction rate, the interpretation accuracy of the fundus image with the negative result can be improved. In some examples, in the over-conversion, the first determination result of the fundus image meeting the preset condition can be set to a positive result. In some examples, the preset condition can include at least one of the following conditions: a history of fundus disease, corrected visual acuity lower than a preset visual acuity, a history of diabetes, or a history of diabetes for more than or equal to a first preset period of time, a history of hypertension, or a history of hypertension for more than or equal to a second preset period of time, and recent visual acuity decline. In this way, subsequent determination on the fundus image with the positive result can be facilitated. For specific description, refer to the related description of the over-conversion module.

[0096] In some examples, in step S130, the negative result images can be sampled, and the sampled negative result images can be used as the sampled negative images. In some examples, the negative result images that are not sampled can be transferred to step S180 (described later) for processing. In some examples, in the sampling, the sampling scheme can be adjusted according to the dimensions or the release time length of the interpretation system 200. In this way, the reliability of the sampling can be improved. In some examples, the dimensions can include at least one of the following: the source of the fundus image, the model of the imaging device of the fundus image, the imaging personnel of the fundus image, the imaging quality of the fundus image, and the interpretation personnel information of the fundus image. In this case, the fundus images can be classified based on different dimensions, and sampling can be performed separately for each category. In this way, the reliability of the sampling can be improved. For specific description, refer to the related description of the first sampling module 230.

[0097] In some examples, in step S140, the positive result images and the sampled negative images can be used as the quality control images. In some examples, the quality control images can be judged to obtain a second judgment result of the quality control images or used as the first images to be re-read. In some examples, the first images to be re-read can include the quality control images meeting a major positive condition. In some examples, the major positive condition can be a lesion feature affecting vision or threatening vision. Thus, the major positive condition can be determined and the images to be re-read can be obtained based on the major positive condition. In some examples, the lesion feature can include at least one of a cup-to-disc ratio greater than or equal to a preset cup-to-disc ratio, an optic disc drusen, a medullary nerve fiber, other abnormalities of the optic nerve, cotton wool spots, hard exudates, retinal vascular sclerosis, retinal vascular obstruction, leopard-shaped fundus, abnormal pigmentation in the macular area, macular hole, drusen, hemorrhage, proliferative membrane, other abnormalities of the retina, non-proliferative diabetic retinopathy, and proliferative diabetic retinopathy. In some examples, the lesion feature affecting vision or threatening vision can include at least one of the macular hole, the retinal vascular obstruction, and the proliferative diabetic retinopathy. Thus, the lesion feature affecting vision or threatening vision can be determined. For specific descriptions, refer to the related descriptions of the second interpretation module 240.

[0098] In some examples, in step S150, the positive result images can be sampled and the sampled positive result images can be used as the sampled positive images. In some examples, the positive result images not sampled can be transferred to step S180 (to be described later) for processing. For specific descriptions, refer to the related descriptions of the second sampling module 250.

[0099] In some examples, in step S160, the sampled positive images can be judged to obtain a third judgment result of the sampled positive images or used as the second images to be re-read. In some examples, the second images to be re-read can include the sampled positive images meeting the major positive condition. For specific descriptions, refer to the related descriptions of the third interpretation module 260.

[0100] In some examples, in step S170, the first images to be re-read and the second images to be re-read can be used as the images to be re-read. In some examples, the images to be re-read can be judged to obtain a fourth judgment result. For specific descriptions, refer to the related descriptions of the fourth interpretation module 270.

[0101] In some examples, in step S180, the final judgment result of the fundus images can be obtained based on the respective judgment results of the fundus images. In some examples, the fundus images can be re-judged based on the respective judgment results of the fundus images. In some examples, the respective judgment results can include the first judgment result, the second judgment result, the third judgment result, or the fourth judgment result.

[0102] In some examples, in the general inspection, the second judgment result of the negative result image in the preset time period can be acquired, and it is determined whether the first judgment result of the negative result image in the at least one dimension is qualified according to the at least one dimension. If yes, the first judgment result of the negative result image in the at least one dimension is taken as the final judgment result, otherwise, the negative result image in the at least one dimension is reset to rejudge the negative result image in the at least one dimension. In some examples, whether the first judgment result of the negative result image in the at least one dimension is qualified can be determined based on the first sampling error rate. Specifically, the second judgment result of the negative result image in the preset time period can be acquired, and the first sampling error rate is calculated according to the at least one dimension. If the first sampling error rate is greater than the first preset error rate, it is determined that the first judgment result of the negative result image in the at least one dimension is not qualified. Thus, whether the first judgment result is qualified can be determined based on the first sampling error rate. For specific description, please refer to the related description of the general inspection release module 280.

[0103] In some examples, in the general inspection, the third judgment result of the positive result image in the preset time period can also be acquired, and it is determined whether the second judgment result of the positive result image in the at least one dimension is qualified according to the at least one dimension. If yes, the second judgment result of the positive result image in the at least one dimension is taken as the final judgment result, otherwise, the positive result image in the at least one dimension is reset to rejudge the positive result image in the at least one dimension. In some examples, in the general inspection, the third judgment result or the fourth judgment result can also be taken as the final judgment result. In some examples, whether the second judgment result of the positive result image in the at least one dimension is qualified can be determined based on the second sampling error rate. Specifically, the third judgment result of the positive result image in the preset time period can be acquired, and the second sampling error rate is calculated according to the at least one dimension. If the second sampling error rate is greater than the second preset error rate, it is determined that the second judgment result of the positive result image in the at least one dimension is not qualified. Thus, whether the second judgment result is qualified can be determined based on the second sampling error rate. For specific description, please refer to the related description of the general inspection release module 280.

[0104] Figure 7 is a flow chart showing the interpretation method for detecting fundus abnormalities based on fundus images according to examples of the present disclosure.

[0105] As Figure 7As shown, in some examples, the interpretation method can further include step S190. In step S190, the fundus image can be confirmed and judged based on the respective judgment results of the fundus image to obtain a fifth judgment result. In some examples, the respective judgment results can include the first judgment result, the second judgment result, the third judgment result, or the fourth judgment result. In some examples, the final judgment result can be the fifth judgment result. For specific description, please refer to the related description of the fifth interpretation module.

[0106] In some examples, for the judgment result of the fundus image corresponding to the final judgment result that has been released, the interpretation method can further sample the respective judgment results (e.g., the third judgment result or the fourth judgment result) based on a preset period (e.g., every month or every quarter) and at least one dimension to determine whether the judgment result of the at least one dimension is qualified. In some examples, the interpretation method can further perform recall processing on the fundus image of the at least one dimension that is determined to be unqualified. In some examples, the recalled fundus image can be subjected to at least one of full inspection, increasing the sampling proportion of the dimension corresponding to the recalled fundus image, and maintenance of the hardware and software of the interpretation system 200. In this way, the accuracy of interpreting the fundus abnormalities in the fundus image can be further improved. For specific description, please refer to the related description of the quality control module 290 and the recall module 300.

[0107] Although the present disclosure has been specifically described above with reference to the drawings and examples, it should be understood that the above description does not limit the present disclosure in any form. Those skilled in the art can make modifications and changes to the present disclosure as needed without departing from the essential spirit and scope of the present disclosure, and these modifications and changes all fall within the scope of the present disclosure.

Claims

1. A system for interpreting fundus abnormalities based on fundus images, characterized in that, It includes an acquisition module, a first interpretation module, a first sampling module, a second interpretation module, a second sampling module, a third interpretation module, a fourth interpretation module, and a final inspection and release module; The acquisition module is used to acquire fundus images; The first interpretation module uses machine learning to judge the fundus image to obtain a first judgment result and divides the fundus image into negative result images and positive result images based on the first judgment result. The first sampling module is used to sample the negative result image and use the extracted negative result image as the sampled negative image; The second interpretation module is used to receive a positive result image from the first interpretation module and / or a negative sample image from the first sampling module as a quality control image, and to make a judgment on the quality control image to obtain a second judgment result of the quality control image or to use the quality control image as a first image to be reread. The second sampling module is used to sample the positive result image and use the extracted positive result image as the sampled positive image; The third interpretation module is used to interpret the sampled positive image to obtain a third interpretation result of the sampled positive image or to use the sampled positive image as a second image to be reread. The fourth judgment module is used to judge the image to be reread, including the first image to be reread and the second image to be reread, to obtain a fourth judgment result; as well as The final inspection and release module performs a final inspection on the fundus image based on the first, second, third, or fourth judgment result of the fundus image to release the fundus image based on the final judgment result of the fundus image or to re-judge the fundus image. The interpretation system also includes a quality control module and a recall module. For the judgment results of the fundus images corresponding to the final judgment results that have been released, the quality control module samples the first interpretation module, the second interpretation module, the third interpretation module, or the fourth interpretation module based on a preset period and at least one dimension to determine whether the judgment results of the at least one dimension are qualified. The recall module is used to perform recall processing on the fundus images of the at least one dimension that are determined to be unqualified by the quality control module.

2. The interpretation system according to claim 1, characterized in that: The final inspection and release module includes a final inspection module, a rereading module, and a release module. The final inspection module acquires a second judgment result of a negative result image within a preset time period and determines whether the first judgment result of the negative result image is qualified. If qualified, the first judgment result of the negative result image is used as the final judgment result; otherwise, the rereading module resets the negative result image so that the interpretation system can re-judge the negative result image. The final inspection module also acquires a third judgment result of a positive result image within the preset time period and determines whether the second judgment result of the positive result image is qualified. If qualified, the second judgment result of the positive result image is used as the final judgment result; otherwise, the rereading module resets the positive result image so that the interpretation system can re-judge the positive result image. The final inspection module also uses the third judgment result or the fourth judgment result as the final judgment result. The release module is used to release the final judgment result.

3. The interpretation system according to claim 2, characterized in that: The final inspection module obtains the second judgment result of the negative result image within the preset time period and calculates the first sampling error rate. If the first sampling error rate is greater than the first preset error rate, the first judgment result of the negative result image is determined to be unqualified. The final inspection module obtains the third judgment result of the positive result image within the preset time period and calculates the second sampling error rate. If the second sampling error rate is greater than the second preset error rate, the second judgment result of the positive result image is determined to be unqualified.

4. The interpretation system according to claim 1, characterized in that: The first image to be reread includes a quality control image that meets the major positive criteria, and the second image to be reread includes a sampled positive image that meets the major positive criteria, wherein the major positive criteria are the presence of lesion features that affect or threaten vision.

5. The interpretation system according to claim 1, characterized in that: The interpretation system further includes a fifth interpretation module, which is used to confirm and interpret the first interpretation result of the first interpretation module, the second interpretation result of the second interpretation module, the third interpretation result of the third interpretation module, or the fourth interpretation result of the fourth interpretation module to obtain a fifth interpretation result, and the final interpretation result is the fifth interpretation result.

6. The interpretation system according to claim 1, characterized in that: The interpretation system further includes an over-conversion module, which is used to convert the first judgment result of the fundus image that is negative to a positive result to match the requirement of the preset negative prediction rate; the over-conversion module is also used to set the first judgment result of the fundus image that meets the preset conditions as a positive result, the preset conditions including at least one of the following: history of fundus disease, corrected visual acuity lower than preset visual acuity, history of diabetes or duration of diabetes greater than or equal to a first preset number of years, history of hypertension or duration of hypertension greater than or equal to a second preset number of years, and recent visual decline.

7. The interpretation system according to claim 1, characterized in that: The first interpretation module establishes different sub-judgment models for each fundus abnormality. The sub-judgment model receives the fundus image and obtains the sub-judgment result. The first interpretation module obtains the first judgment result based on multiple sub-judgment results.

8. A method for interpreting fundus abnormalities based on fundus images, characterized in that, include: Acquire fundus images; The fundus images are judged based on machine learning methods to obtain a first judgment result, and the fundus images are divided into negative result images and positive result images based on the first judgment result; The negative result images are sampled and the extracted negative result images are used as sampled negative images; The positive result image and / or the sampled negative image are used as quality control images. The quality control images are judged to obtain a second judgment result of the quality control images or the quality control images are used as the first images to be reread. The positive result images are sampled and the extracted positive result images are used as sampled positive images; The sampled positive image is evaluated to obtain a third evaluation result for the sampled positive image or the sampled positive image is used as a second image to be reread. A fourth judgment result is obtained by judging the image to be reread, including the first image to be reread and the second image to be reread. Based on the first, second, third, or fourth judgment result of the fundus image, a final judgment result of the fundus image is performed to approve the fundus image or to re-judge the fundus image. and For the fundus images corresponding to the final judgment results that have been approved, the first judgment result, the second judgment result, the third judgment result, or the fourth judgment result are sampled based on a preset period and at least one dimension to determine whether the judgment result of the at least one dimension is qualified, and the fundus images of the at least one dimension that are determined to be unqualified are subject to recall processing.

9. The interpretation method according to claim 8, characterized in that: In the final inspection, a second judgment result of a negative result image within a preset time period is obtained, and the first judgment result of the negative result image is determined to be qualified. If qualified, the first judgment result of the negative result image is used as the final judgment result; otherwise, the negative result image is reset for re-judgment. In the final inspection, a third judgment result of a positive result image within the preset time period is also obtained, and the second judgment result of the positive result image is determined to be qualified. If qualified, the second judgment result of the positive result image is used as the final judgment result; otherwise, the positive result image is reset for re-judgment. In the final inspection, either the third judgment result or the fourth judgment result is also used as the final judgment result.

Images