KR20210017429A - Method for optic disc classification - Google Patents

Abstract

The present invention relates to a method for optic disc classification including: (a) a step in which a local system acquires a fundus image for training; (b) a step in which the local system pre-treats the acquired fundus image for training; (c) a step in which the local system divides an optic disc; (d) a step in which the local system extracts features from the divided optic disc image; (e) a step in which the local system transmits the extracted optic disc features to a cloud server; (f) a step of training a classifier model using the optic disc features received at the cloud server; (g) a step in which a mobile device acquires a test fundus image; (h) a step in which the mobile device pre-treats the acquired test fundus image; (i) a step in which the mobile device divides an optic disc; (j) a step in which the mobile device extracts features from the divided optic disc image; (k) a step in which the mobile device transmits the extracted optic disc features to a cloud server; (l) a step in which the cloud server classifies the optic disc using the received optic disc features as a classifier model input; and (m) a step in which the cloud server transmits the result of the classification to the mobile device.

Description

Optic nerve nipple classification method {METHOD FOR OPTIC DISC CLASSIFICATION}

The present invention relates to a method for classifying an optic nerve nipple.

Centralized online services are the core of smart cities that provide diagnostic services to patients. Mobile cloud architecture is commonly used by researchers to share and diagnose relevant information. Mobile computing is used in research to reduce computational complexity in terms of system performance by performing various activities from a smart phone to the cloud. This is an emerging area of research, especially for patients in remote areas. Cloud computing offers a number of benefits, such as cost savings, scalability, focus on core competencies, service resilience, facilitating online collaboration and reusability.

Mobile cloud plays an important role in helping ophthalmologists (ophthalmologists) monitor patients remotely and effectively in smart cities. The latest requirement in this state-of-the-art age is to enable doctors and ophthalmologists to diagnose diabetic retinopathy remotely by giving them 24-hour or any location access to medical facilities in smart cities.

Diabetes disturbs small blood vessels, ie the capillaries of various organs, especially the retina, kidney and nervous tissue. Diabetes, which affects the capillaries of the retina, is known as diabetic retinopathy. Capillary damage can lead to blood leakage and irregular blood vessel formation, leading to vision loss, leading to blindness.

The fundus photograph as shown in FIG. 1 is mainly used for pre-detection of important retinal diseases such as glaucoma and diabetic retinopathy. The optic disc (OD) is the main visible area of the retinal image and blood vessels derived from and around the retina. The inner part of the macula is the fovea responsible for the sharp vision of the center. The abnormal fundus image may consist of a large area (exudate) larger than the area of the optic nerve.

OD localization is useful for detecting other anatomical retinal structures such as fovea, macula and vascular changes. The OD cup to OD disc ratio is a basic step towards the early onset of glaucoma. Automated computer-assisted screening programs for early detection of the disease are essential and a major step in screening for glaucoma in countries with large populations.

In the literature, many researchers have performed OD segmentation and classification work in each application field. In general, edge-based techniques were used for OD segmentation and classification. In this approach, the red band is extracted from the RGB color space, and the closing operator is applied to fill the vessel and then opened to remove the large selection. The Sobel edge operator is used for OD localization. Similarly, a histogram matching based technique was used to locate the center of the OD. This technique used a histogram as a model for each color space in images in the DRIVE data set. The shape-based technique is also used for OD localization in which the segmentation of the foreground and background is performed using a critical technique followed by various morphological operations. OD localization is performed using the main joining section with a round outline. However, the computational complexity of these techniques is high in terms of time and space.

The Hough transform technique is also used to detect circular shapes in image processing. However, the Hough transform technique does not provide the best results because it accepts resolution. Entropy filtering technique is used for OD localization. After applying various pre-processing steps such as smoothing, segmentation and mathematical morphology techniques. The entropy of the smooth region is lower compared to the non-flat region. Principal component analysis (PCA) is also used for OD localization. In this approach, various magnification factors of PCA are applied to the area to find the minimum distance between the input area and the projected image. The center of the OD is placed at the point with the minimum distance of the candidate section among all the scaling elements. It should be noted that PCA-based technology has a limitation of high time complexity.

The blood vessel direction matching filter is used for OD segmentation by matching the direction of blood vessels in a corresponding region. Similarly, segmenting blood vessels using a 2D Gaussian filter is also used to detect OD in the fundus image. The variance between matching filters was re-adjusted to four different sizes. The center of the OD is the region adjacent to each candidate region of the blood vessel. OD localization is performed by a vascular directional pattern. Therefore, the OD center was localized based on the blood vessel direction. A Markov random field (MRF) and a compensation factor method are used for OD partitioning. The first retinal blood vessel is extracted using a graph cutting technique, followed by the blood vessel information used to estimate the OD location.

Advances in the field of modern health care systems require real-time detection and classification of ODs using a mobile cloud to achieve high classification accuracy at low cost in terms of computational complexity (time and space). Systems that meet the requirements for real-time OD detection and classification must address the challenges and risks associated with mobile cloud-based systems such as energy consumption, bandwidth, availability, service stability and regulation.

JP 5348982 B2

The problem to be solved by the present invention not only enables patients to easily access health services remotely, but also makes it easy for an ophthalmologist to diagnose diabetic retinopathy remotely, as well as the accuracy of classification of the optic nerve nipple in the fundus image and the accuracy of the optic nerve nipple. It is to provide a method for classifying optic nerve nipples that can improve the reconstruction rate.

The method for classifying optic nerve nipples according to an embodiment of the present invention for solving the above problems,

(a) obtaining a training fundus image by the local system;

(b) preprocessing, by the local system, the acquired training fundus image;

(c) dividing the optic disc (OD) by the local system;

(d) extracting features from the segmented optic nerve papillary image by the local system;

(e) transmitting, by the local system, the extracted optic nerve nipple features to a cloud server;

(f) training a classifier model using the optic nerve papillary features received from the cloud server;

(g) obtaining, by the mobile device, a test fundus image;

(h) pre-processing the acquired test fundus image by the mobile device;

(i) dividing the optic nerve nipple (OD) by the mobile device;

(j) extracting features from the segmented optic nerve papillary image by the mobile device;

(k) transmitting, by the mobile device, the extracted optic nerve nipple features to a cloud server;

(l) classifying, by the cloud server, the optic nerve nipples by using the received optic nerve nipple features as an input of a classifier model; And

(m) transmitting, by the cloud server, the classification result to the mobile device.

In the optic nerve nipple classification method according to an embodiment of the present invention, the pretreatment steps of steps (b) and (h),

(b-1) converting the input image from an RGB color space to an HSI color space image;

(b-2) smoothing the brightness component of the converted image using an intermediate value filter;

(b-3) improving the contrast of the smoothed image using contrast-limited adaptive histogram equalization;

(b-4) converting an image with improved contrast from an HSI color space to an RGB color space image; And

(b-5) extracting a green channel image from the RGB image.

In addition, in the method for classifying the optic nerve nipple according to an embodiment of the present invention, the features include a shape feature of the optic nerve nipple,

The shape features of the optic nerve nipple may include an area, a circumference, a circle, and a center of the optic nerve nipple.

According to the method for classifying the optic nerve nipple according to an embodiment of the present invention, an ophthalmologist (ophthalmologist) may help diagnose various eye diseases.

In addition, it helps ordinary citizens to efficiently and functionally simplify high-quality medical services related to eye diseases around the clock, through which medical professionals can make time decisions about abnormalities in the target area, thereby making physical use of patients. Patients can be monitored without possibility.

Additionally, cloud-based computing can be used to achieve computational gains for time complexity reduction. Classifiers for OD classification are learned in the cloud, and Internet bandwidth requirements can be minimized by offloading features extracted from images instead of uploading the entire image.

In addition, an up-scaled global segmentation method may be incorporated into the framework to achieve the optic nerve papillary classification accuracy, thereby providing an improved segmentation result compared to the recently published state-of-the-art method.

1 is a diagram showing a fundus image of the retina and its key anatomical features.
2 is a diagram showing classifier training in a cloud server.
3 is a diagram showing a conceptual diagram of a method for classifying an optic nerve nipple according to an embodiment of the present invention.
4 is a diagram showing a flow chart of a method for classifying an optic nerve nipple according to an embodiment of the present invention.
5 is a diagram showing a detailed flowchart of a pre-processing step.
6 is a view showing a detailed flowchart of the step of dividing the optic nerve nipple.
7(a) is an input image, and FIG. 7(b) is a diagram showing a preprocessed output image.
Fig. 8(a) is a preprocessed optic nerve nipple image, and Fig. 8(b) shows a segmented optic nerve nipple.
9(a) is a diagram showing a divided fundus image, and FIG. 9(b) is a diagram showing a fundus image to which a morphological operation is applied.
Fig. 10 is a diagram showing an optic nerve papilla region deleted using a morphological operation.
11(a) to 11(l) are diagrams showing the detection of the optic nerve nipple boundary for the Al-Shifa and DRIVE data sets.
12(a) to 12(l) are diagrams showing the detection of the optic nerve papillary boundary for the DIARETDB0 and DIARETDB1 data sets.
13 is a diagram comparing the classification of the optic nerve nipples in the method for classifying the optic nerve nipples according to an embodiment of the present invention for the DIARETDB0 and DIARETDB1 data sets and the classification of the optic nerve nipples by an ophthalmologist, and FIG. 13(a) is The result of classification of the optic nerve nipples, FIG. 13(b) is a result of classification of the optic nerve nipples according to the method of classification of the optic nerve nipples according to an embodiment of the present invention, and FIG. Figure showing accuracy.
14 is a view comparing the accuracy of the classification of the optic nerve nipples and the accuracy of other techniques of the method of classifying the optic nerve nipple according to an embodiment of the present invention for DIARETDB0 and DIARETDB1 data sets.

Objects, specific advantages and novel features of the present invention will become more apparent from the following detailed description and preferred embodiments associated with the accompanying drawings.

Prior to this, terms or words used in the present specification and claims should not be interpreted in a conventional and dictionary meaning, and the inventor may appropriately define the concept of the term in order to describe his own invention in the best way. It should be interpreted as a meaning and a concept consistent with the technical idea of the present invention based on the principles that exist.

In adding reference numerals to elements of each drawing in the present specification, it should be noted that, even though they are indicated on different drawings, only the same elements are to have the same number as possible.

In addition, terms such as "first", "second", "one side", and "the other side" are used to distinguish one component from other components, and the component is limited by the terms. It is not.

Hereinafter, in describing the present invention, detailed descriptions of related known technologies that may unnecessarily obscure the subject matter of the present invention will be omitted.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Health care services play an important role in the future reality of the present age, smart cities. Automatic recognition of the optic disc (OD) region is an essential step in automated retinal image analysis. In general, OD should be excluded in order to detect exudate, a major symbol of diabetic retinopathy. Mobile cloud-assisted classification, which classifies OD from color fundus images, can improve classification accuracy and reconstruction speed of OD.

In an optic nerve nipple classification method according to an embodiment of the present invention, a smart phone-based cloud-based cloud-supported resource recognition framework for detecting and classifying OD in a fundus image using a classification procedure trained in the cloud is proposed.

The optic nerve nipple classification method according to an embodiment of the present invention includes shape-based features trained by a Naive Bayes classifier in the cloud for OD and non OD classes. Experiments were conducted on three online available data sets and real data sets developed in local hospitals with different contrasts, illuminances and abnormalities. The optic nerve nipple classification method according to an embodiment of the present invention achieves an accuracy of 98.25% in the classification of OD.

Mobile cloud computing is an important source of using the computational process as a service. It facilitates global access to search, store and perform data analysis in a secure and easy way. Mobile cloud-based OD segmentation and classification is an innovative technology, and system experts and ophthalmologists can easily perform tasks locally and remotely through the system. This can be done locally on the device or in the cloud.

The method for classifying the optic nerve papilla according to an embodiment of the present invention accesses a fundus retinal image data set for image processing, distribution, and analysis. Initially, the fundus image is segmented into ODs, and then feature extraction is performed locally to perform additional classifier training in the cloud used for additional OD classification. Once features are extracted locally, it is offloaded (transmitted) to the cloud server for further analysis. The trained classifier exists in the cloud, allowing experts to obtain a classifier that classifies OD from fundus images on mobile devices. The processed images and trained models can be stored on a server in the cloud for future work. 2 shows a training mechanism of a cloud-based classifier, reference number 200 is a mobile device of an ophthalmologist, reference number 202 is a cloud server, and reference number 204 obtains a training data set and transmits it to the cloud server 202. It is a local system. It is used extensively to compress the training time of a classifier to process large data sets.

Mobile cloud support OD analysis

The optic nerve papillary classification method according to an embodiment of the present invention for OD segmentation and classification in a color fundus retinal image processed through different steps such as image acquisition and preprocessing, OD segmentation and morphological computation It extracts and trains a classifier model in the cloud with these features for classifying two classes, OD class and Non OD (Non OD) class. A conceptual diagram of a method for classifying an optic nerve nipple according to an embodiment of the present invention is shown in FIG. 3.

The conceptual diagram of the method for classifying an optic nerve nipple according to an embodiment of the present invention shown in FIG. 3 is a step of acquiring a fundus image (step S300), preprocessing a fundus image (step S302), and dividing the optic nerve nipple (OD). (Step S304), extracting features from the divided optic nerve nipples (Step S306), transmitting the extracted features to a cloud server (Step S308), and naive Bayes using the training data set by the cloud server After training the classifier, classifying the optic nerve nipple segmented from the test fundus image into an optic nerve nipple (OD) class or a non-optic nerve nipple (non OD) class (step S310).

4 is a flowchart of a method for classifying an optic nerve nipple according to an embodiment of the present invention.

Referring to the method for classifying the optic nerve nipple according to an embodiment of the present invention shown in FIG. 4, in step S400, the local system (204 in FIG. 2) acquires a training fundus image, and the training fundus image obtained in step S402 Preprocessing, splitting the optic disc (OD) in step S404, extracting features from the split optic nerve nipple image in step S406, and storing the features of the optic nerve nipple extracted in step S408 on a cloud server (202 in FIG. 2) Transfer to.

In step S410, the cloud server (202 in FIG. 2) trains a naive Bayes classifier model using the received optic nerve nipple features.

Meanwhile, in step S412, the ophthalmologist's mobile device (200 in FIG. 2) acquires a test fundus image, pre-processes the test fundus image obtained in step S414, and divides the optic nerve nipple (OD) in step S416, and step S418. Features are extracted from the optic nerve nipple image divided in step S420, and the optic nerve nipple features extracted in step S420 are transmitted to the cloud server (202 in FIG. 2).

In step S422, the cloud server (202 in Fig. 2) classifies the optic nerve nipples using the received optic nerve nipple features as input of the Naive Bayes classifier model, and transmits the classification result to the mobile device (200 in Fig. 2) in step S424. do.

Each step will be described in detail below.

Data collection and preprocessing

Fundus retinal images are collected using a special fundus camera that photographs the core surfaces of the eye such as the retina, optic nerve papilla, macula and posterior pole. The field of view of the image varies between 30° and 50°. Two types of fundus cameras are used. One is based on pupil dilation, called pupil dilation, and the other is a non-pupil dilation fundus camera that does not require pupil dilation. In the optic nerve nipple classification method according to an embodiment of the present invention, all acquired images are in RGB color space using a non-pupillary dilated fundus camera.

5 is a detailed flowchart of the preprocessing steps of steps S402 and S414. Referring to FIG. 5, since the intensity component is separately processed in the retinal image obtained in step S500 without interfering with the other two components, hue, saturation, and intensity in the RGB color space. )(HSI) is converted to an image in color space. The HSI color model separates color and gray scale information and closely matches the color interpretation of the human eye. Color is a characteristic defining clean color, saturation is the expansion of pure color by white light, and brightness is an achromatic factor for the brightness of an object.

In step S502, the brightness component is smoothed to remove noise present in the image. Noise in the fundus image occurs during image acquisition due to the blinking of the patient's eyelashes or eyelids, which can image the eyelashes, and the patient's head movement turns black due to out-of-focus images, images with irises, and fatally illuminated retinas. Can result in images. Thus, these factors generate noise, for example random fluctuations in the brightness of the image formed by a scanner or digital camera within the image. Different noise removal filtering techniques were evaluated for noise removal such as average filtering, intermediate filtering, and the like. Since the average filter generates a blurred image compared to the median filter), the median filter is used to smooth the input image G(x, y) in the proposed method as follows.

Here, w denotes a neighboring pixel centered on the position of the image. F(x, y) is the output image, and G(x, y) is the input image.

Image contrast enhancement is a process of improving the contrast of an image, so the resulting image is visually appealing. In step S504, in order to improve the contrast of an image, a contrast limited adaptive histogram equalization (CLAHE) technique is applied as shown in the following equation.

Here, F(x, y) is the input image, H(x, y) is the output image, and T is the transform function.

In this technique, the entire image is evenly divided into regions, and histogram equalization is applied to each region. This distributes the gray value assignment of the image evenly and makes dark features more visible. Then the results of the relevant sections are combined with interpolation to reproduce the entire image. Interpolation efficiently combines the related areas into the entire image so that the junctions of the connected areas are not visible.

7(a) is a diagram showing an input image, and FIG. 7(b) is a diagram showing a preprocessed output image.

After pre-processing in step S506, the image in the HSI color space is converted back into an image in the RGB color space, and in step S508, since OD is better visible in the green channel, a green channel image is extracted from the image in the RGB color space.

Optic nerve nipple (OD) segmentation

Segmentation is the process of dividing an image into individual segments. The threshold value is an important part of image segmentation where the foreground object is separated from the background by assigning an intensity value t (threshold value) for each pixel. The OD threshold is important to separate the foreground object from the background for further OD classification.

6 is a diagram showing a detailed flowchart of the step of dividing the optic nerve nipple. The image preprocessed in step S600 uses global Otsu's threshold technology (Otsu 1979) to separate the foreground object from the background area. This technique is based on a threshold that minimizes variance within a weighted class. This is equivalent to maximizing the variance between classes given in Equation 3. Otsu systematically searches for a threshold value that reduces the intraclass variance, which is divided by the difference between the sum of the foreground and background weighted variances given in the following equation.

과

는 이 두 클래스의 분산이다. 총 분산

는 클래스 내 분산(가중치)과 클래스 간 분산의 합계이다. 이는 클래스 평균과 전체 평균(grand mean) 사이의 가중된 제곱 거리의 합이다.Where q ₁ (t) and q ₂ (t) are the weighted probabilities of the two classes, t is the threshold and

and

Is the variance of these two classes. Total variance

Is the sum of the intra-class variance (weight) and the inter-class variance. This is the sum of the weighted squared distances between the class mean and the grand mean.

Fig. 8(a) is a preprocessed optic nerve nipple image, and Fig. 8(b) is a view showing a segmented optic nerve nipple.

In step S602, a morphological operation is applied to the foreground object to detect the optic nerve papillary region.

During the segmentation process, some unwanted objects, such as blood vessels or exudates, are also detected and morphological operations to exclude these objects are performed to separate these objects according to their properties. A threshold value for removal other than the OD region is set as given by the following equation.

는 다른 오브젝트를 제거하기 위해 설정된 임계 값이며, J(x, y)는 출력 이미지이다. 다른 임계 값

는 원하지 않는 오브젝트를 제거하는 데 사용될 수 있다. OD에 경계 픽셀을 포함하기 위해 오브젝트를 확대하고 오브젝트의 틈을 복구할 때 형태학적 확장을 사용한다. 구조화 요소

및 입력 이미지 J(x, y)는 다음 수학식을 사용하여 확장된다:Where R(x, y) is the input image given in Equation 3,

Is the threshold set to remove other objects, and J(x, y) is the output image. Different threshold

Can be used to remove unwanted objects. Enlarging the object to include the boundary pixels in the OD and using morphological expansion to repair gaps in the object. Structured element

And the input image J(x, y) is expanded using the following equation:

The presence of background pixels in the foreground is commonly known as a hole. To fully detect OD, this hole must be filled inside the OD area. For this, a morphological operation is used to completely detect the OD region as shown in the following equation.

Where X ₀ is the starting point within the boundary, B is the structural element, and A ^c is A's complement.

9(a) is a diagram showing a divided fundus image, and FIG. 9(b) is a diagram showing a fundus image to which a morphological operation is applied.

Feature extraction from segmented OD

Classification of OD in fundus images requires strong feature extraction. Objects can be described by various parameters such as shape, texture and color. Feature extraction from the optic nerve papilla is a subsequent step to detect the OD center after morphological calculation. For OD localization, the size and shape features of the optic nerve papilla are extracted. These features are fundamental to OD localization.

(a) Area: The area represents the total number of pixels in the OD area. Since a binary fundus image has a black background and a white foreground, that is, an OD area, the number of white pixels (foreground) represents the OD area. The area K(x, y) of the optic nerve is calculated by the following equation.

은 영역 R의 행과 열이고

는 수학식 5에 주어진 결과 이미지이다.here

Is the row and column of area R

Is the result image given in Equation 5.

(b) Perimeter: Generally, perimeter is the distance around the boundary of the OD region. Mathematically, it can be written as

의 경계 픽셀이다.Where B is the optic nerve nipple

Is the border pixel of.

(c) Circularity: Because the OD is circular, the circular feature is important in the optic nerve papilla. Segmentation of the fundus image may detect one or more candidate regions. Therefore, it is excellent to search for the largest circular shape object using the following equation.

Density is the area of the detected object divided by the square of the perimeter of the detected object and multiplied by 4π.

(d) Centroid: The center is defined as the center of the optic nerve and is also called the center of gravity. The shape center is basically the center point of the shape and is calculated as follows.

Where X _c is the average of all x coordinates of the component pixels and Y _c is the average of all y coordinates of the component pixels.

Classification

A class is a collection of objects with some important common attributes such as shape characteristics, namely area, perimeter, circle, and center, and the class to which the object belongs is indicated by the class label. Classification is the process of assigning labels (OD class and Non OD (Non OD) class) to images according to some representation of the attributes of the image. Various classification techniques have been used in the literature. Neuve Bayesian classifier (Friedman et al. 1997), artificial neural network (ANN) (Gardner et al. 1996), Nearest Neighbor (KNN) and Support Vector Machine (SVM) (Cortes and Vapnik 1995). In the method for classifying optic nerve nipples according to an embodiment of the present invention, a Naive Bayes classifier is used for classifying optic nerve nipples (OD) in a fundus image.

The Naive Bayes classifier is a supervised machine learning technique based on Bayes' theorem. It supports conditional independence. That is, the resulting value of the feature (X _i ) for a given class (C _i ) is independent of the value of other features. It is suitable for large data sets and is considered the best classifier when the input dimension is high. The following is an equation of the Naive Bayes classifier that provides the probability of a class.

는 특징들(크기, 형상, 원형 및 조밀도)이 주어지는 경우 클래스 (OD, 비 OD(Non OD))의 사후 확률,

는 (OD, 비 OD(Non OD)) 클래스의 사전 확률,

는 특징이 주어진 클래스 (OD, 비 OD(non OD))의 확률이며,

는 주어진 특징들의 사전 확률이다.

Is the posterior probability of class (OD, Non OD (Non OD)) given features (size, shape, circularity and density),

Is the prior probability of the class (OD, Non OD),

Is the probability of a given class (OD, non OD),

Is the prior probabilities of the given features.

으로 표시된다. 클래스 라벨이 없는 X로 알려지지 않은 데이터 샘플을 제공함으로써 분류기는 X가 가장 높은 사후 확률을 갖는 클래스 C_i에 속한다는 것을 추측한다.For testing, the extracted shape features are feature vectors showing n measurements performed on the sample from n attributes.

Is indicated by By providing an unknown data sample as X without a class label, the classifier guesses that X belongs to class C _i with the highest posterior probability.

Experiment results and discussion

In this section, a data set of retinal images used in the method for classifying an optic nerve papilla according to an embodiment of the present invention and experimental results of the proposed technique will be described in detail. We also discussed and evaluated the accuracy of this technique through consultations with ophthalmologists and related research recently published.

data set

As described in Table 1, four data sets were tested to measure the performance of the optic nerve nipple classification method according to an embodiment of the present invention. Three of the data sets are DIARETDB0 (Kauppi et al. 2006), DIARETDB1 (Kauppi et al. 2007) and DRIVE (Staal et al. 2004), which are available online, and the fourth data set is from local Al-Shifa (Al-Shifa 2017). , Evaluated for OD classification in this study. The actual data set images are collected by a Cannon CF-1digital retina camera. One of the contributions of the method for classifying optic nerve nipples according to an embodiment of the present invention uses an actual data set provided by Kohat (Al-Shifa 2017), a Khyber Pakhtunkhwa Eye Hospital in Pakistan. All images in the data set are in .JPEG format.

The experiment was performed on a total of 309 retinal images selected from four data sets. The Al-Shifa data set consists of 50 retinal images, 10 of which do not have a DR mark, and the remaining 40 images have some marks for DR. The DIARETDB0 data set consists of 130 retinal images, 20 images without DR markers and 110 images with DR markers. The DIARETDB1 data set consists of 89 retinal images, 5 images have no DR mark, the remaining 84 images have signs of anomalies, the DRIVE data set consists of 40 retinal images, and 33 images have a DR mark. There is no indication of abnormality in the remaining 7 images. The resolution of the image varies from 565×584 to 1500×1152, and the fundus image is accustomed to 256×256 pixels depending on the simulation requirements. The field of view of the fundus image varies from 35° to 50° and the image is in .JPEG format as indicated in Table 1.

debate

In this subsection, step-by-step results are evaluated for all four data sets. The image collection and pre-processing results are shown in FIG. 7 showing the retinal input and subsequent pre-processed output images after applying various pre-processing steps.

Segmentation of the pre-processed image is performed using the Otsu technique. Otsu's partitioning value is less than necessary. Since the target OD is small in size and is the brightest globally, the threshold value is adjusted by a factor of another value and finally by a coefficient of α=3.20 as given in Equation 13. 8 shows a preprocessed input and result binary image to which the improved Otsu algorithm is applied.

Where I(x, y) is the Otsu result image.

Since an object other than a part of the OD area was also detected after the division, Equations 4 to 6 are applied to separate the object area based on the attribute as shown in FIG. 9 in order to exclude the erroneously identified area.

In order to include the bounding pixel in the OD area, the expansion operation is applied because it is used to expand the object. Any background pixels present in the foreground area, i.e. the OD area, must be removed by using morphological operations to track the segment enclosed by the closed shape as shown in FIG. 10.

The final OD boundary for the Al-Shifa Trust and DRIVE data sets and the results of the central classification of OD are presented in FIG. 11. In the actual data set provided by the Al-Shifa and DRIVE data set images, the segmentation and classification accuracy of the optic nerve nipple classification method according to an embodiment of the present invention is 98% and 100%.

12 shows the result of the OD boundary and the central classification of OD for the DIARETDB0 and DIARETDB1 data sets. In the DIARETDB0 and DIARETDB1 data sets, the segmentation and classification accuracy of the method for classifying the optic nerve papilla according to an embodiment of the present invention is 97% and 98%.

Classifier performance

The extracted feature vectors are passed to the cloud's Naive Bayes (NB) classifier for training and storage for further classification by assigning a class label to a defined test data set image. The performance measurement of the optic nerve nipple classification method according to an embodiment of the present invention is based on a confusion matrix (CM) for evaluation of an NB classifier. In CM, true positive (TP) means when the ground truth about positive is correctly classified as positive, and false positive (FP) means when a negative class is incorrectly classified as positive. Similarly, true negative (TN) means when the ground true negative class is correctly classified as negative, and false negative (FN) means when the negative class is classified as positive. The performance of the NB classifier was measured through CM, sensitivity, specificity and accuracy. Details are as follows.

-Sensitivity: Like a patient's tumor, it is a measure that determines the probability of an outcome, that is, true positive.

-Specificity: A measure that concludes the probability of an outcome, that is, true negative, as if the patient did not have a tumor.

-Accuracy: A measure that determines the probability of how many results are classified correctly. The equation is as follows:

Table 2 compares the classification of OD by the method of classifying the optic nerve nipple according to an embodiment of the present invention and the manual classification by four ophthalmologists of Al-Shifa. Classification by the optic nerve nipple classification method according to an embodiment of the present invention achieved an accuracy of 98.25% because OD was correctly classified in 303 out of 309 images identical to the result of an ophthalmologist.

In the Al-Shifa, DIARETDB0, DIARETDB1 and DRIVE data sets, the classification accuracy of the method for classifying the optic nerve papilla according to an embodiment of the present invention is 98%, 97%, 100%, and 98%, respectively. In the Al-Shifa, DIARETDB0, DIARETDB1 and DRIVE data sets, 49 of 50 images, 127 of 130 images, 87 of 89 images and 40 of 40 images were each correctly classified. Compared with the results of 303 ophthalmologists, the optic nerve nipple classification method according to an embodiment of the present invention also achieved an overall classification accuracy of 98.25% as given in the table. In other words, 303 out of 309 images were correctly classified into the OD class.

13 is a view comparing the classification of the optic nerve nipples in the method for classifying the optic nerve nipples according to an embodiment of the present invention for the DIARETDB0 and DIARETDB1 data sets and the classification of the optic nerve nipples by an ophthalmologist at Al-Shifa Trust Hospital, FIG. 13 (a ) Is an optic nerve nipple classification result by an ophthalmologist, FIG. 13 (b) is an optic nerve nipple classification result according to an optic nerve nipple classification method according to an embodiment of the present invention, and FIG. 13 (c) is according to an embodiment of the present invention. It is a diagram showing the accuracy of the method for classifying the optic nerve nipples.

As described in Table 3, the method for classifying the optic nerve papilla according to an embodiment of the present invention is (Kovacs et al., 2010; Salazar-Gonzalez et al., 2014; Sopharak et al. 2008; Ullah et al. 2013; Yulah et al. 2013; Yulah et al. 2013; Yulah et al. 2013; et al. 2015). The total classification accuracy of the proposed technology for the four data sets Al-Shifa, DIARETDB0, DIARETDB1 and DRIVE is 98%, 97%, 98% and 100% respectively, and the classification accuracy is 97%, 95%, 96% respectively. , 95% and 98% (Kovacs et al. 2010; Salazar-Gonzalez et al. 2014; Sopharak et al. 2008; Ullah et al. 2015).

It should be noted that the overall average classification accuracy of the optic nerve nipple classification method according to an embodiment of the present invention is 98.25%. 14 is a method for classifying optic nerve nipples and other methods according to an embodiment of the present invention (Kovacs et al. 2010; Salazar-Gonzalez et al. 2014; Sopharak et al. 2008; Ullah et al. 2013; Yu et al. 2015) shows a bar graph for comparison.

Energy consumption and computational complexity analysis

The computational complexity and energy consumption analysis of the method for classifying optic nerve nipples according to an embodiment of the present invention were integrated from the results of (Sriram and Khajeh-Hosseini 2010). 88.48J of energy is required for 8x8 packet data transmission. Table 4 shows the energy consumption for three identical scenarios: smartphone and mobile cloud-based OD segmentation and classification. Complete offloading, i.e. transmission of a complete OD image, consumes a lot of energy and high communication cost compared to local processing that only offloads extracted features. Processing locally in the system reduces transmission costs, but complex tasks such as pre-processing, segmentation and feature extraction and classification are performed on the local system. Nevertheless, smartphones are not suitable for this energy waste task. Therefore, cloud servers improve the environment that surpasses smartphones for computationally intensive tasks.

conclusion

In the method for classifying the optic nerve papilla according to an embodiment of the present invention, an efficient mobile cloud-based classification method for detecting and classifying OD in retinal images is proposed. The goal is to reduce the computational complexity of existing systems by performing key functions in the cloud. In the method for classifying optic nerve nipples according to an embodiment of the present invention, intensity bands (hue, saturation, and intensity) are extracted from an RGB color space divided into foreground and background regions using Otsu's binarization method. Shape features such as area, perimeter, circle and center are extracted and combined into a feature set for OD localization. Finally, these extracted features are used to train Cloud's Naive Bayes classifier for classifying optic nerve nipples. The optic nerve nipple classification method according to an embodiment of the present invention provides an efficient classification technique for OD segmentation with low computational complexity. The performance of the optic nerve nipple classification method according to an embodiment of the present invention is evaluated on four data sets including the actual data set of Al-Shifa Trust Hospital. It was observed that the classification result obtained by using the optic nerve nipple classification method according to an embodiment of the present invention provides an accuracy of 98.25%. The qualitative and quantitative results are encouraging, and the optic nerve nipple classification method according to an embodiment of the present invention can achieve accurate results by saving energy consumption and calculation time.

Although the present invention has been described in detail through specific embodiments, this is for describing the present invention in detail, and the present invention is not limited thereto, and those of ordinary skill in the art within the technical scope of the present invention It will be said that it is clear that the transformation or improvement is possible.

All simple modifications to changes of the present invention belong to the scope of the present invention, and the specific scope of protection of the present invention will be made clear by the appended claims.

200: mobile device
202: Cloud Server
204: local system

Claims (3)

(a) obtaining a training fundus image by the local system;
(b) preprocessing, by the local system, the acquired training fundus image;
(c) dividing the optic disc (OD) by the local system;
(d) extracting features from the segmented optic nerve papillary image by the local system;
(e) transmitting, by the local system, the extracted optic nerve nipple features to a cloud server;
(f) training a classifier model using the optic nerve papillary features received from the cloud server;
(g) obtaining, by the mobile device, a test fundus image;
(h) pre-processing the acquired test fundus image by the mobile device;
(i) dividing the optic nerve nipple (OD) by the mobile device;
(j) extracting features from the segmented optic nerve papillary image by the mobile device;
(k) transmitting, by the mobile device, the extracted optic nerve nipple features to a cloud server;
(l) classifying, by the cloud server, the optic nerve nipples by using the received optic nerve nipple features as input to a classifier model; And
(m) The method of classifying an optic nerve nipple comprising the step of transmitting, by the cloud server, the classification result to the mobile device. The method according to claim 1,
The pretreatment steps of steps (b) and (h),
(b-1) converting the input image from an RGB color space to an HSI color space image;
(b-2) smoothing the brightness component of the converted image using an intermediate value filter;
(b-3) improving the contrast of the smoothed image using contrast-limited adaptive histogram equalization;
(b-4) converting an image with improved contrast from an HSI color space to an RGB color space image; And
(b-5) An optic nerve nipple classification method comprising the step of extracting a green channel image from the RGB image. The method according to claim 1,
The features include the shape features of the optic nerve papilla,
The shape characteristic of the optic nerve nipple comprises an area, circumference, circle and center of the optic nerve nipple.

Images