KR20220021171A - Deep Neural Networks(DNN) based eye diseases diagnosis apparatus using fundus images and method thereof - Google Patents

Abstract

The present invention relates to an eye disease diagnosis apparatus using a fundus image and a method thereof. More specifically, the present invention relates to a deep neural network based eye disease diagnosis apparatus using a fundus image and a method thereof. A training dataset is normalized by performing image preprocessing including region-of-interest reduction, isolight plane contrast-limited adaptive histogram equalization, and data augmentation. A deep neural network is applied to diagnose an eye disease including normal, glaucoma (GLC), age-related macular degeneration (AMD), and diabetic retinopathy (DR).

Description

DEEP Neural Networks (DNN) based eye diseases diagnosis apparatus using fundus images and method thereof

The present invention relates to an apparatus and method for diagnosing eye diseases using a fundus image, and more particularly, image preprocessing including region of interest (ROI) reduction, isolight plane contrast-limited adaptive histogram equalization, and data augmentation. to normalize the training dataset, and apply a deep neural network to apply multiple To an apparatus and method for diagnosing eye diseases based on a deep neural network using a fundus image capable of diagnosing eye diseases.

Diabetic Retinopathy (DR), Glaucoma (GLC), and Age related macula degeneration (AMD) are the leading causes of vision loss and blindness worldwide.

Diabetic retinopathy is a vascular complication of the retina that damages the retina, leading to severe vision loss if not treated promptly.

From 2000 to 2030, the global diabetes population growth rate is estimated to be 2.8% to 4.4%, which corresponds to about 171 million to 195 million people suffering from diabetes. There is a risk of blindness in 2% of people with diabetes and loss of vision in about 10% over the next 15 years.

Glaucoma (GLC) also occurs under the influence of differential pressure within the eyeball, which damages the optic nerve head and causes vision loss.

Age-related macular degeneration (AMD) is also a cause of vision loss and blindness in the elderly over the age of 60, and appears to have a significant impact on the elderly over the age of 60.

Many of these patients will struggle in their lives because of sight, one of the five basic human senses.

There are various techniques for diagnosing the above-mentioned eye diseases by experts or doctors, and these techniques determine eye diseases using fundus images taken through optical coherence tomography (OCT) and fundus imaging, which are cross-sectional imaging. do.

In addition, for rapid diagnosis and accurate diagnosis of eye diseases, eye disease diagnosis apparatuses by image classification to which various artificial intelligence models are applied in each of DR, GLC, and AMD diagnosis areas are being researched and developed.

It has been shown that the risk of blindness due to diabetic retinopathy can be reduced by 50% by early detection and appropriate treatment through periodic screening of diabetic retinopathy.

In general, in order to diagnose an ophthalmic disease, a fundus image is obtained by performing optical coherence tomography (OCT) and fundus imaging.

Fundus images are being used to diagnose and screen many eye diseases, such as diabetic retinopathy (DR), glaucoma (Glaucoma), and age-related macular degeneration (AMD). Therefore, the quality of fundus images can affect the clinician's ability to perform the correct examination and diagnosis.

Although a plurality of ophthalmic diseases can be diagnosed from the fundus image as described above, the conventional ophthalmic disease diagnosis apparatus to which artificial intelligence is applied can detect only one specialized eye disease from the fundus image.

That is, there is a problem that the conventional apparatus for diagnosing eye diseases using artificial intelligence detects only one eye disease from the fundus image and cannot detect two or more eye diseases.

Republic of Korea Patent Publication No. 10-2020-0065923 (published on June 9, 2020)

Therefore, an object of the present invention is to normalize the learning dataset by performing image preprocessing including reduction of the region of interest, isolight plane contrast restriction adaptive histogram equalization, and data augmentation, etc., and apply a deep neural network to obtain Normal, Glaucoma: An apparatus and method for diagnosing eye diseases based on a deep neural network using fundus images that can diagnose multiple eye diseases including GLC), age related macular degeneration (AMD), and diabetic retinopathy (DR) is in providing.

A deep neural network-based eye disease diagnosis apparatus using a fundus image according to the present invention for achieving the above object is: A plurality of fundus images for each of at least two or more eye diseases, and an expert's comment on the corresponding eye disease for each fundus image An fundus image acquisition unit for obtaining and outputting a training dataset comprising; Including a deep neural network artificial intelligence model, the learning dataset input from the fundus image acquisition unit is applied to the deep neural network (DNN) artificial intelligence model to learn and deep learned for diabetic retinopathy, age-related macular degeneration, glaucoma and normal a deep learning unit that outputs a neural network artificial intelligence model; a diagnostic image acquisition unit for acquiring and outputting a fundus image to be diagnosed; and the deep neural network artificial intelligence model is input and driven, and the diagnosis target fundus image is applied to the deep neural network artificial intelligence model to determine whether an eye disease is present in the diagnosis target fundus image, and if an eye disease exists, the corresponding eye disease It characterized in that it comprises an eye disease diagnosis unit for outputting a diagnosis result.

The training dataset includes a plurality of fundus images with diabetic retinopathy and diabetic retinopathy learning data with annotations on the diabetic retinopathy, fundus images with age-related macular degeneration (AMD) and age with annotations on the age-related macular degeneration. It is characterized in that it includes normal learning data including relevant macular degeneration learning data and a fundus image having glaucoma, glaucoma learning data having an annotation for the glaucoma, and a normal fundus image.

The apparatus is characterized by further comprising: an image preprocessing unit that performs image preprocessing for adjusting the size and number of the fundus image of the training dataset and the fundus image to be diagnosed, and outputs the image preprocessing unit.

The image preprocessor may include: a region of interest acquisition unit configured to extract a region of interest from the fundus image of the training dataset and the evaluation target fundus image and adjust the size; and an image enhancer configured to enhance the number of images for the fundus image by adjusting at least one parameter of the fundus image output from the region of interest acquiring unit.

The region of interest obtaining unit may include: a channel selector for extracting fundus images of a red channel and a blue channel of the fundus image; a size adjustment unit for outputting the fundus image including the red channel and the blue channel by adjusting the size including only the region of interest; and a contrast-limited adaptive histogram equalizer that applies and outputs the scaled fundus image contrast-limited adaptive histogram equalization.

The apparatus is: After dividing the training dataset generated by the image preprocessing into a training set and a validation set, K-fold cross-validation is performed by the validation set on the trained deep neural network artificial intelligence model to obtain a validation result It characterized in that it further comprises a verification unit to output.

A deep neural network-based eye disease diagnosis method using a fundus image according to the present invention for achieving the above object is: A fundus image acquisition unit provides a plurality of fundus images for each of at least two or more eye diseases and the corresponding eye disease for each fundus image. Fundus image acquisition process of acquiring and outputting a training dataset including an expert's annotation on the subject; The deep learning unit applies the learning dataset input from the fundus image acquisition unit, including the deep neural network artificial intelligence model, to the deep neural network (DNN) artificial intelligence model to learn diabetic retinopathy, age-related macular degeneration, glaucoma and normal A deep learning process that outputs the trained deep neural network artificial intelligence model; a diagnostic image acquisition process in which a diagnostic image acquisition unit acquires and outputs a fundus image to be diagnosed; and an eye disease diagnosis unit receives and operates the deep neural network artificial intelligence model, and applies the diagnosis target fundus image to the deep neural network artificial intelligence model to determine whether an eye disease exists in the diagnosis target fundus image and whether an eye disease exists It is characterized in that it includes an eye disease diagnosis process for outputting the corresponding eye disease diagnosis result.

The training dataset includes a plurality of fundus images with diabetic retinopathy and diabetic retinopathy learning data with annotations on the diabetic retinopathy, fundus images with age-related macular degeneration (AMD) and age with annotations on the age-related macular degeneration. It is characterized in that it includes normal learning data including relevant macular degeneration learning data and a fundus image having glaucoma, glaucoma learning data having an annotation for the glaucoma, and a normal fundus image.

The method is characterized by further comprising: an image pre-processing step of performing image pre-processing in which an image pre-processing unit adjusts the size and number of fundus images of the training dataset and the fundus images to be diagnosed, and outputting the image pre-processing steps.

The image preprocessing may include: obtaining, by the image preprocessing unit, a region of interest acquiring a region of interest by extracting a region of interest from the fundus image of the training dataset and the fundus image to be evaluated through the region of interest acquiring unit and adjusting the size; and an image augmentation step in which the image preprocessor adjusts at least one parameter of the fundus image output from the region of interest acquiring unit through the image augmentation unit to enhance the number of images for the fundus image.

The obtaining of the region of interest may include: a channel selection step of extracting fundus images of a red channel and a blue channel of the fundus image; a size adjustment step of adjusting and outputting the fundus image including the red channel and the blue channel to a size including only the region of interest; and a contrast-limited adaptive histogram equalization step of applying and outputting the scaled fundus image contrast-limited adaptive histogram equalization.

The method is: After a verification unit divides the training dataset generated by the image preprocessing into a training set and a verification set, the trained deep neural network artificial intelligence model is verified by performing K-fold cross-validation by the verification set. It characterized in that it further comprises a verification process for outputting the result.

The present invention applies a deep neural network (DNN) to learn eye diseases including normal, glaucoma, connection-related macular degeneration, and diabetic retinopathy from the fundus image, so it is possible to classify and diagnose two or more eye diseases from the fundus image. can have an effect.

In addition, the present invention has the effect of overcoming the lowest average absolute error rate from the linear cumulative distribution function (CDF) by applying control-limited adaptive histogram equalization (CLAHE) data preprocessing.

In addition, since the present invention performs K-fold cross-validation, it has the effect of providing improved verification results because verification is performed on a model with a larger number of features.

In addition, the present invention has the effect of increasing the number of real images rather than fake images through data augmentation, thereby improving the diagnosis accuracy of eye diseases.

1 is a diagram showing the configuration of an apparatus for diagnosing eye diseases based on a deep neural network using a fundus image according to the present invention.
2 is a view showing a fundus image on which image pre-processing according to the present invention is performed.
3 is a view showing a fundus image on which image enhancement image pre-processing according to the present invention is performed.
4 is a flowchart illustrating a method for diagnosing eye diseases based on a deep neural network using a fundus image according to the present invention.
5 is a flowchart illustrating an image preprocessing method among the deep neural network-based eye disease diagnosis methods using fundus images according to the present invention.

Hereinafter, the configuration and operation of a deep neural network-based eye disease diagnosis apparatus using a fundus image according to the present invention will be described in detail with reference to the accompanying drawings, and an eye disease diagnosis method using the fundus image in the device will be described.

1 is a diagram showing the configuration of an apparatus and method for diagnosing an eye disease based on a deep neural network using a fundus image according to the present invention, FIG. 2 is a diagram showing a fundus image on which image preprocessing according to the present invention is performed, and FIG. 3 is this It is a view showing a fundus image that has been subjected to image enhancement image pre-processing according to the present invention. Hereinafter, it will be described with reference to FIGS. 1 to 3 .

An apparatus for diagnosing eye diseases based on a deep neural network using a fundus image according to the present invention includes a fundus image acquisition unit 100, a diagnostic image acquisition unit 200, a deep learning unit 400 and an eye disease diagnosis unit 600, According to an embodiment, the image preprocessor 300 and the cross-verification unit 500 may be further included.

The fundus image acquisition unit 100 acquires and outputs a learning dataset including a plurality of fundus images for each of at least two or more eye diseases and an expert's annotation for each eye disease for each fundus image.

In the present invention, a plurality of normal fundus images, a plurality of glaucoma fundus images, a plurality of age-related macular degeneration fundus images, a plurality of diabetic retinopathy fundus images, and each fundus image include annotations corresponding to corresponding eye diseases.

For example, in the present invention, a total of 2,335 fundus images can be acquired and used for learning, and among the 2,335, the normal fundus image is 1195, the glaucoma fundus image is 168, the ligament-related macular degeneration fundus image is 65, and the diabetic retinopathy fundus image is 1195. may contain 907. Each fundus image will include annotation information recorded by an expert.

The fundus image may be collected from databases such as ORIGA(-light)1, IDRID2, Messidor3, ARIA4, STARE5, and the like. Each database stores and manages fundus images for at least one or more specific eye diseases.

The diagnosis image acquisition unit 200 may acquire a fundus image to be diagnosed, which is a fundus image of a target to be diagnosed, from a fundus camera and an OCT camera, or may be obtained by loading a fundus image to be diagnosed from a storage means, and may be obtained through communication means. It may be obtained by receiving the fundus image to be diagnosed.

The image pre-processing unit 300 pre-processes the fundus image of the training dataset input from the fundus image acquisition unit 100 and outputs it to the deep learning unit 400, or a diagnosis input from the diagnostic image acquisition unit 200 The target fundus image is pre-processed and output to the eye disease diagnosis unit 600 .

The image preprocessor 300 includes a region of interest acquisition unit 310 and an image augmentation unit 320 , and the region of interest acquisition unit 310 includes a channel selector 311 , a size adjuster 312 , and a collation unit. and a constraint adaptive histogram equalizer 313 .

The region of interest acquisition unit 310 extracts and outputs only the region of interest from the fundus image.

Specifically, as shown in Equation 1 below, the channel selector 311 extracts the red channel and the green channel based on the threshold values of 25 and 13 to standardize and output the entire fundus image. The green channel is a complementary layer.

where i _max is the image width and j _max is the image height.

The size adjustment unit 312 adjusts the fundus image of the extracted channel to a size of 384*384.

The contrast-limited adaptive histogram equalizer 313 performs histogram equalization of the fundus image whose size is adjusted by the size adjuster 312 in the luminance plane. The contrast-limited adaptive histogram equalizer 313 can overcome the lowest average absolute error rate from the linear cumulative distribution function (CDF) by performing histogram equalization on the luminance plane.

The region of interest acquisition unit 310 extracts only the region of interest (ROI) from the original fundus image on the left side of FIG. 2 as shown in the center of FIG. To improve the quality of the fundus image from which only the ROI is extracted.

The image augmentation unit 320 adjusts at least one parameter of the fundus image output from the region of interest obtaining unit 310 to augment and output the number of images for the fundus image.

The image augmentation increases the number of fundus images by enlarging the size (parameter) of the fundus image within 20%, rotating it at a certain angle (parameter), or changing the brightness (parameter) within 25%, without creating a fake image. augment

For example, as shown in Figure 3, the left is the original fundus image, the left center is the fundus image with brightness adjustment of +25% and 80 degrees counterclockwise rotation of the original fundus image, and the right center is the original fundus image + Image with 25% brightness adjustment, 20% zoom out (reduce), horizontal flip, 40 degree counterclockwise rotation, etc. The far right is +25% brightness adjustment, 20% zoom in (zoom in), horizontal flip. It is an augmented image that has been performed.

Image enhancement avoids the performance differences that can be created by AMD's very small dataset sizes and NR and DR's very large datasets.

However, since all functions, that is, the optic nerve disc, the macula, and blood vessels, must be included when the fundus image is augmented, the movement of the parameters is not applied.

The deep learning unit 400 configures the shape of the input fundus image to an optimal shape of 384 * 384 * 384 * 383, and a deep neural network (DNN) artificial intelligence model to which the fourth-class connected softmax prediction probability is applied is configured, The fundus image of the image preprocessed learning dataset is applied to the deep neural network artificial intelligence model to learn, and the learned deep neural network artificial intelligence model is output to the cross-validation unit 500 and the eye disease diagnosis unit 600 .

In the deep learning unit 400 of the present invention, a neural network optimizer is applied, the neural network optimizer plays an important role in forming weights and performing learning, and the loss function guides the optimizer to move in the right direction. In the present invention, an adaptive gradient extension optimizer called Adadelta is applied to prevent learning robustness and learning rate change. The learning rate of the Adadelta is within 0.001, and a categorical cross-entropy loss function is used.

Only 80% of the dataset output from the image preprocessor 300 is applied to learning, 10% is used for testing, and 10% is used for verification.

The cross-validation unit 500 receives the 10% verification data set to the deep neural network artificial intelligence model learned in the deep learning unit 400 and performs K-fold cross-validation, particularly 10-fold cross-validation. If the verification fails, the deep learning unit 400 will preferably be configured to perform re-learning.

The eye disease diagnosis unit 600 classifies the eye disease by applying the image pre-processed diagnosis target fundus image to the learned deep neural network artificial intelligence model that has been successfully verified through the cross verification unit 500, and outputs the classified eye disease. do.

4 is a flowchart illustrating a method for diagnosing eye diseases based on a deep neural network using a fundus image according to the present invention.

Referring to FIG. 4 , the fundus image acquisition unit 100 obtains a fundus image learning dataset including an expert's comment on the fundus image and each fundus image required according to the present invention from the above-described plurality of databases to pre-process the image. It outputs to the unit 300 (S111).

The image preprocessing unit 300 image preprocesses the fundus images of the input training dataset and provides the image preprocessed training dataset to the deep learning unit 400 ( S113 ).

The deep learning unit 400 applies the fundus images of the image preprocessed learning dataset input from the image preprocessor 300 to the deep neural network artificial intelligence model to learn and generate an eye disease diagnosis model and cross-validation unit 500 and provided to the eye disease diagnosis unit 600 (S115).

The cross-validation unit 500 applies the verification dataset to the generated eye disease diagnosis model to perform K-fold cross-validation to determine success (S117).

If successful, the eye disease diagnosis unit 600 drives the eye disease diagnosis model, which is a deep neural network artificial intelligence model input from the deep learning unit 400 , and is acquired from the diagnostic image acquisition unit 200 and uses the image preprocessing unit 300 . It monitors whether a fundus image to be diagnosed is input through the system (S119), and when the fundus image to be diagnosed is input, it is applied to an eye disease diagnosis model (S123) and a diagnosis result is output (S125).

5 is a flowchart illustrating an image preprocessing method among the deep neural network-based eye disease diagnosis methods using fundus images according to the present invention.

Referring to FIG. 5 , first, the region of interest acquisition unit 310 of the image preprocessor 300 extracts only the red channel and the blue channel of each fundus image of the training dataset input through the channel selector 311 and adjusts the size. It outputs to the unit 312 (S211).

The size adjustment unit 312 adjusts the size of the fundus image of the red channel to 384 pixels, and then outputs it to the contrast restriction adaptive histogram equalizer 313 ( S213 ).

The contrast-limited adaptive histogram equalizer 313 outputs the scaled fundus image to the image enhancer 320 by performing histogram equalization on the luminance plane (S215).

The image augmentation unit 320 generates an augmented image for the fundus image while changing the parameter for each input fundus image, and provides the training dataset from which the augmented image is generated to the deep learning unit 400 (S220). ).

On the other hand, it is common knowledge in the art that the present invention is not limited to the typical preferred embodiments described above, but can be improved, changed, replaced, or added in various ways without departing from the spirit of the present invention. Those who have will be able to understand it easily. If implementation by such improvement, change, substitution, or addition falls within the scope of the appended claims below, the technical idea should also be regarded as belonging to the present invention.

100: fundus image acquisition unit 200: diagnostic image acquisition unit
300: image preprocessor 310: region of interest acquisition unit
311: channel selection unit 312: size adjustment unit
313: Contrast restriction adaptive histogram equalizer
320: image augmentation unit 400: deep learning unit
500: cross validation unit 600: eye disease diagnosis unit

Claims (12)

a fundus image acquisition unit for acquiring and outputting a learning dataset including a plurality of fundus images for each of at least two or more eye diseases and an expert's comment on the corresponding eye disease for each fundus image;
Including a deep neural network artificial intelligence model, the learning dataset input from the fundus image acquisition unit is applied to the deep neural network (DNN) artificial intelligence model and trained to learn about diabetic retinopathy, age-related macular degeneration, glaucoma and normal a deep learning unit that outputs a neural network artificial intelligence model;
a diagnostic image acquisition unit for acquiring and outputting a fundus image to be diagnosed; and
The deep neural network artificial intelligence model is received and driven, and the diagnosis target fundus image is applied to the deep neural network artificial intelligence model to determine whether an eye disease exists in the diagnosis target fundus image, and if an eye disease exists, the corresponding eye disease diagnosis An apparatus for diagnosing eye diseases based on a deep neural network using a fundus image, characterized in that it includes an eye disease diagnosis unit that outputs a result.
According to claim 1,
The training dataset is
Multiple fundus images with diabetic retinopathy and diabetic retinopathy learning data with annotations for the diabetic retinopathy, fundus images with age-related macular degeneration (AMD) and age-related macular degeneration learning data with annotations for the age-related macular degeneration, and An apparatus for diagnosing eye diseases based on a deep neural network using a fundus image, comprising: a fundus image having glaucoma, glaucoma learning data having an annotation on the glaucoma, and normal learning data including a normal fundus image.
3. The method of claim 2,
An apparatus for diagnosing eye diseases based on a deep neural network using a fundus image, characterized in that it further comprises an image preprocessing unit that performs image preprocessing for adjusting the size and number of fundus images of the learning dataset and the fundus images to be diagnosed.
4. The method of claim 3,
The image pre-processing unit,
a region of interest obtaining unit extracting a region of interest from the fundus image of the training dataset and the evaluation target fundus image and adjusting the size; and
A deep neural network-based eye disease diagnosis apparatus using a fundus image, characterized in that it comprises an image enhancer for increasing the number of images for the fundus image by adjusting at least one parameter of the fundus image output from the region of interest acquiring unit.
5. The method of claim 4,
The region of interest acquisition unit,
a channel selection unit for extracting fundus images of a red channel and a blue channel of the fundus image;
a size adjustment unit for outputting the fundus image including the red channel and the blue channel by adjusting the size including only the region of interest; and
A deep neural network-based eye disease diagnosis apparatus using a fundus image, characterized in that it comprises a contrast-limited adaptive histogram equalizer outputting the size-adjusted fundus image contrast-limited adaptive histogram equalization.
6. The method of claim 5,
After dividing the training dataset generated by the image preprocessing into a training set and a verification set, a verification unit that outputs a verification result by performing K-fold cross-validation by the verification set on the learned deep neural network artificial intelligence model A deep neural network-based eye disease diagnosis apparatus using a fundus image, characterized in that it further comprises.
A fundus image acquisition process in which the fundus image acquisition unit acquires and outputs a learning dataset including a plurality of fundus images for each of at least two or more eye diseases and an expert's annotation for each fundus image;
The deep learning unit applies the learning dataset input from the fundus image acquisition unit, including the deep neural network artificial intelligence model, to the deep neural network (DNN) artificial intelligence model to learn diabetic retinopathy, age-related macular degeneration, glaucoma and normal A deep learning process that outputs the trained deep neural network artificial intelligence model;
a diagnostic image acquisition process in which a diagnostic image acquisition unit acquires and outputs a fundus image to be diagnosed; and
The eye disease diagnosis unit receives and operates the deep neural network artificial intelligence model, and the diagnosis target fundus image is applied to the deep neural network artificial intelligence model to determine whether an eye disease exists in the diagnosis target fundus image and when an eye disease exists A deep neural network-based eye disease diagnosis method using a fundus image, characterized in that it includes an eye disease diagnosis process that outputs the corresponding eye disease diagnosis result.
8. The method of claim 7,
The training dataset is
Multiple fundus images with diabetic retinopathy and diabetic retinopathy learning data with annotations for the diabetic retinopathy, fundus images with age-related macular degeneration (AMD) and age-related macular degeneration learning data with annotations for the age-related macular degeneration, and A deep neural network-based eye disease diagnosis method using a fundus image, comprising: a fundus image having glaucoma, glaucoma learning data having an annotation on the glaucoma, and normal learning data including a normal fundus image.
9. The method of claim 8,
Deep neural network-based eye disease using a fundus image, characterized in that the image preprocessing unit further comprises an image preprocessing process for performing image preprocessing for adjusting the size and number of fundus images of the learning dataset and the fundus images to be diagnosed diagnostic method.
10. The method of claim 9,
The image preprocessing process is
an ROI obtaining step in which the image preprocessor extracts a region of interest from the fundus image of the training dataset and the evaluation target fundus image through the region of interest acquirer and adjusts the size; and
Using a fundus image, characterized in that the image preprocessing unit adjusts at least one parameter of the fundus image output from the region of interest acquisition unit through the image augmentation unit to enhance the number of images for the fundus image. A deep neural network-based ocular disease diagnosis method.
11. The method of claim 10,
The step of acquiring the region of interest is
a channel selection step of extracting a fundus image of a red channel and a blue channel of the fundus image;
a size adjustment step of adjusting and outputting the fundus image including the red channel and the blue channel to a size including only the region of interest; and
A method for diagnosing eye diseases based on a deep neural network using a fundus image, characterized in that it includes a contrast-limited adaptive histogram equalization step of applying and outputting the scaled fundus image contrast-limited adaptive histogram equalization.
12. The method of claim 11,
After the verification unit divides the training dataset generated by the image preprocessing into a training set and a verification set, K-fold cross-validation is performed by the verification set on the trained deep neural network artificial intelligence model to output a verification result A deep neural network-based eye disease diagnosis method using a fundus image, characterized in that it further comprises a verification process.

Images