Automated Retinal Image Analysis to Detect Optic Nerve Hypoplasia

Identification of the optic disc and fovea is crucial for automating the diagnosis and screening of retinal diseases. Based on quantitative calculations


Introduction
An important layer of the eye is the retina, which can show signs of disease early.The majority of visual disorders are caused by structural deterioration of vari-ous structures in the retina layer of the eye (vessels, optic disc, macula, fovea, etc.) [2].Fundus cameras with different settings are used to take retinal images.
There are three different modes for fundus cameras, including Color Fundus Retinal Imaging, Red-Free Imaging, and Fluorescent Angiography Imaging [4].A fundus image provides insight into the anatomy of the eye's crucial structures, including the optic disc, blood vessels, macula and fovea.An essential anatomical structure in retinal fundus images is the optic disc (OD).According to fundus images, the OD appears bright, yellowish, and almost circular.It is where the main blood vessels enter the retina.It is difficult to get an accurate location or outline of OD with segmentation techniques because of the distractions in the images, such as abnormal lesions, imprecise boundaries, and peripapillary atrophy, despite OD being an approximate circle and having high intensity characteristics [11].In Figure 1, OD appears as a bright disc-shaped area at the point where blood vessels gather 3-4 mm from the fovea, the macula's center.An area between superior and inferior temporal retinal arteries and veins that run tangential to OD is known as the macula.Figure 1 shows the macula's four topographic components: perifovea, parafovea, fovea, and foveolar [36].The retina provides detailed central vision thanks to this small, highly sensitive area [35].Located in the macula's center, the fovea is the most sensitive region of vision.This is part of the eye that controls central vision and colour vision.It has a diameter of about 1.5 mm.This region is a region where depression occurs at the center of the macula in the retina.In general, the depth of this region can vary from person to person by 4.0 mm and 0.8 mm, but on average it is 0.25 mm.Further, when evaluating and grading retinal diseases, it is important to know the distance between the OD and the fovea.To determine and locate OD and fovea accurately and automatically, this distance is highly important [19].In this study, a method was developed for the diagnosis of optic nerve hypoplasia, which is associated with detecting regional anatomical structures of the retina, including the OD, the macula and the fovea, where the effects of changes associated with many eye diseases can be observed early.
Optical nerve hypoplasia (ONH) occurs when the retinal ganglion cells and axons constituting the optic nerve do not develop properly.This disease is also referred to as an optic disc that is smaller than normal, and it has been found to cause significant visual dysfunctions, particularly in children [30].Statistics indicate that ONH is among the leading causes of visual disorders in children worldwide [33].The prevalence of ONH increased six-fold between 1970 and 2000, reaching 1.1 per 10,000 children today.Between 1970 and 2000, ONH prevalence increased six fold, reaching the current estimate of 1.1 per 10,000 children [12].Especially in recent years, Mousa et al. reported an increase in children experiencing health problems related to visual disorders.It was stated that approximately 12% of visual impairment cases in children in the US and Europe were caused by ONH [33].Men and women are equally affected by ONH.Although the disease rarely occurs, it is crucial to detect it accurately to prevent blindness from occurring.Every newborn child is therefore examined and calculated to determine whether or not they have this disease [12].As shown in Figure 2, a smaller ratio than 0.3 exists in ONH images between the horizontal disc diameter (DD) and the distance between the fovea center and the temporal edge of the optic disc (DM).In normal retinal images, the ratio is greater than 0.3 [30].
This study aimed to measure these parameters most accurately and quickly so that the diagnostic result had the smallest error margin possible.The diameter of the optic disc and its distance from the macula's center are manually measured in ophthalmology, and a diagnosis is based on the ratios of these measurements.Diagnosing this disease requires the use of a calculator or paper and pencil.
This study aimed to measure these parameters most accurately and quickly so that the diagnostic result had the smallest error margin possible.In ophthalmology, the diameter of the optic disc and the distance between the optic disc edge and the umbo point located at the center of the macula are manually measured, and a diagnosis is made based on the ratios of these parameters.Diagnosing this disease requires the use of a calculator or paper and pencil.This study provides doctors with a useful tool for diagnosing optic nerve hypoplasia, which had not been diagnosed before with deep learning methods.This paper makes the following major contributions: _ On retinal fundus images, boundaries of OD, OD diameter, and the umbo point, also called the macula center, were automatically detected to diagnose this disease._ Calculations necessary for diagnosis were done automatically without the use of paper, pencil, or calculators._ It takes a neural network less than a second to predict in an appropriate setting, while a doctor needs more time to comprehend and make a decision based on the image features.Also, deep neural networks do not get tired even if they perform the same tasks repeatedly.Thus, it helps eliminate disease risks that may have been overlooked during diagnosis.

Related Work
Automated retinal image analysis begins with OD detection.
A general classification of OD segmentation methods is shape-based, template-matching, and deformable and active contour-based.A number of segmentation methods are used in the detection of anatomical structures for the diagnosis of ONH, including gradient information-based methods, thresholding methods, shape detection methods, edge detection methods, and active contour-based methods, as in other biomedical image segmentation tasks [31].One of the most frequently encountered problems with conventional image processing methods is that the method is valid only for a certain target dataset.As these methods have not worked in different environments and conditions, researchers are looking for a general solution with deep learning.Depending on the imaging conditions and assumptions about the acquired images, classical image segmentation techniques can be accurate [9].Thus, automated systems based on these methods may not work for images with different properties.These issues can be solved by deep learning segmentation solutions that learn how to represent certain objects from various annotated images.Deep learning architectures can be improved to produce more reliable models through problem-specific improvement studies.This study showed the effects of such an improvement on segmenting anatomical structures like the optic disc and fovea in retinal images.
Especially in the fields of convolutional neural networks (CNNs) and medical image analysis, deep learning has shown noteworthy segmentation performances in recent years [28,65].Anatomical structures of the retina are increasingly being segmented with deep learning methods [26,29].In Table 1, CNN methods for detecting anatomical structures in retinal images are compared with existing methods.In CNN-related work, we learn how to map features in an image and use that knowledge to map them in more detail.Converting an image to a vector gives good results in classification problems.To segment an image, the vector created after converting a feature map into a vector must be used to reconstruct the image.
Creating an image from a vector is very challenging.It is based on this problem that U-Net architecture emerged.In the same way that an image is converted to a vector, the vector is also converted back into an image using the feature mapping used during conversion.As a result, distortions in the image will be greatly reduced and the structural integrity of the image will be preserved.To detect the fovea and optic disc, this study used the U-net model, a deep learning technique commonly used in medical imaging.

Proposed Methodology
The purpose of this study was to develop a method to precisely segment the optic disc and fovea, which are both necessary to detect optic nerve hypoplasia.Aside from providing a flexible and extensible structure, the original U-Net is also effective for medical segmentation [43,34].By adapting to new tasks, the improved model will be able to meet a variety of needs [61,63].
In semantic segmentation processes, pixel-based classification can be done with little training data.Because the classification process is pixel-based, im-ages to be used for education do not have to be split into fovea, disc, and background.All three classes can coexist in the training image.Semantic segmentation architectures can train on a pixel-by-pixel basis because the label image shows what class each pixel belongs to.In this study, at the training stage, the Messidor database was used.There are 1200 colour fundus images in this database taken with a The Topcon TRC NW6 monocular retinography camera can be used for non-mydriatic retinography.It has a field of view of 45°.We divided the processed dataset into three subgroups, 900 for training, 180 for internal validation, and 120 for independent testing, based on a ratio of 0.75:0.15:0.10[58].Segmenting retinal images requires annotating input images for segmentation training.In Ground truth for Messidor's database, fovea was not provided.The ground truth for the annotation of the optic disc and fovea was obtained from a manual label made by an ophthalmologist, the third author of this paper.LabelMe [61], a free and open-source annotation tool developed by MIT, was used for marking optic disc and fovea boundaries and binding boxes for the segmentation of ground truth.LabelMe lets you label images and create annotated masks on the web.LabelMe has been used to map retinal regions and then generate the image mask from the original image to annotate our retinal images.
As a pre-processing step, grayscale conversion and resizing processes were applied to the images before training began.U-Net is a convolutional neural network segmentation architecture that is fully connected.(https://github.com/seva100/optic-nerve-cnn).Due to its shape, it was named after the letter U. Encoders and decoders are the two parts of this architecture [51].In the encoder part, discriminative features are extracted from the image via convolutional layers.In the convolution layer, there are three 3*3 filters, whereas the maximum pooling layer has two 2*2 filters.Encoder part is an iterative convolution operation where the ReLU and max pooling (2x2 max pooling) operations are applied, followed by each convolution.The bottommost layer serves as a mediator between the encoder layer and the decoder layer.It consists of CNN layers followed by an upper convolution layer.Feature information is increased while spatial information is decreased, thus enabling the architecture to learn complex structures more effectively during contracting.Expansion paths, also known as decoders, are symmetrical paths in which precise localization is achieved by transposed convolutions.Therefore, it is a completely convolutional network from end to end, that is, it does not contain any dense layers, but only consists of convolutional layers.Therefore, it can be applied to any image.CNN layers receive input from the expansion layer, which is then passed on to the upsampling layer.When the input is added (depth concatenation), the contraction layer's attribute map corresponds to the current expansion layer.This action will ensure that the attributes learned when shrinking the image are used to reconstruct it.Additionally, the convolutional layer uses half as many feature maps after each block to maintain symmetry.There is the same number of expansion blocks as contraction blocks.After passing through these blocks, the resulting mapping is passed through a 3x3 CNN layer with as many cluster feature maps as there are clusters to be reached.Therefore, while solving the segmentation problem, multiclass classification is also performed at the same time, so that each pixel is classified accordingly [46].
Our U-Net architecture is shown in Figure 3 of this study.There are four blocks in the encoder phase of a proposed U-Net model, each containing two 3x3 convolutional layers.The output of each block is connected to a max pool layer.Like the encoder phase, the decoder phase uses the same blocks, except that the max-pooling layers are replaced by upsampling layers.Batch Normalization (BN) and ReLU layers are included in both the encoder and decoder stages.Using batch normalization, a convolutional neural network can be made more regular.In other words, it normalizes the input layer by scaling the convolutional neural network, making the model more stable and faster.The ReLU function is nonlinear.For negative inputs, ReLU returns 0, while for positive inputs, x returns x.In the encoder part of the pooling layer, the weight parameters are transmitted to the next layer with the maximum pooling method.The unspooling layer can be expressed as performing this process in the decoder phase.A maximum value from each set is selected in the max-pooling layer to reduce the resolution of the image.Data size reduction was the main purpose of this layer.
The number of first feature maps was taken as 16 and model training was carried out to increase to 32, 64,128,256 throughout the layer.Lastly, the output layer calculated values between 0-1 using the sig-   2. In this study, the Adam method was used due to its ease of application, efficiency in calculation, low memory requirement, invariance to diagonal rescaling, and suitability for large dataset problems [25].This method can also be used for non-stationary objectives and noise-or sparse-gradient problems.The learning rate in Table 2 is user-defined values that typically range from 0 to 1. Training and testing the model take longer if the learning rate is too slow.The learning rate, however, may not provide an optimal value at a certain point if it is too high.We tested the learning rate at 0.1, 0.01, 0.001 and 0.0001 for this study, and the best result was 0.001.
In CNN models, millions of weights have to be learned in both the convolutional and fully connected layers.
The training dataset needs to be big with a lot of images so that the weights can be optimized.It is hard to find such a dataset for problems like surface defect detection, though.Typically, this problem is solved with pre-trained network architectures or by increasing the number of samples in the dataset with data augmentation methods.The use of data augmentation is particularly useful when there are relatively few training samples.Large-scale models can be trained more robustly this way.In this way, the data augmentation process prevents problems caused by low numbers of data units and the model's memory behavior [67].This study uses various augmentation methods to increase the number of data points, including flipping, mirroring and cropping.Although data augmentation methods add more samples to a data set, they can be insufficient, especially when there is a lot of similarity between classes.Furthermore, when small defects are small and very sim-ilar to the background, the disappearance of negative samples or blurring may not distinguish the surface error in data augmentation methods.When viewing abnormal fundus images, it can be challenging to segment the optic disc and fovea due to various distractions, such as variations in illumination, blurry boundaries, and occlusion of retinal vessels.In these situations, it will be more advantageous to use pre-trained network architecture [14].ResNet won first place for classification, detection, and localization in an ImageNet competition [14].By using residual convolutions, ResNet improves feature use and network performance.Furthermore, the residual mechanism maintains high performance while adding depth to the network.As networks get deeper, this mechanism has also reduced gradient loss.A residual skip-connection has been proposed as a robust alternative to the U-Net skip-connection, which concatenates encoder features and decoder features.The ResNet-18 model is adapted to U-Net by convolution of the previous layer's output by 2x2 and the encoder part by 1x1 to obtain the decoding blocks.Before going into the next decoder block, the combined tensor is batch-normalized.This final layer is also convolution-transposed with the same plane number as the target classes and with the same size as the output image.
Xiuqin and friends, proposed the ResUnet method, which combines U-Net with a residual learning strategy, to segment retinal blood vessels accurately.As a result of the proposed algorithm, the complexity of the network decreased as well as the accuracy of segmentation increased [66].Based on retinal images, Baid and friends developed a Convolutional Neural Network (CNN)-based system that predicts Pathological Myopia.Their novel Residual UNet architecture has also been used to segment the optic disc from the retinal images [6].As reported by Puchaicela-Lozano et al., the researchers proposed a hybrid approach for glaucoma fundus image localization utilizing pre-trained region-based convolutional neural networks (R-CNNs) ResNet-50 and cup-to-disk segmentation [42].The U-Net architecture was modified in order to develop a robust segmentation method for optic disc and fovea segmentation, combining widely used pre-trained segmentation algorithms.The ResNet-34 model consists of encoding layers and classic U-Net decoding layers [68].Compared to ground truth values, their correlation agreement is over 80% and their mean absolute error is less than 0.08.The Weighted Res-UNet was proposed by Xiao and friends to address the challenging retinal vessel segmentation problem by incorporating weighted attention and skip-connecting strategies.
A baseline model is implemented with no attention and without skip connections [64].Our method differs from studies conducted with traditional U-net in the following ways: _ With the addition of ResNet's residual learning module to the U-Net architecture, we propose Res-UNet, which increases learning ability and training efficiency.The network learns the optimal representational features through the filters, so no handcrafted features are required._ Using the trained classifier, a segmentation map is obtained for each retinal image.A bright blob around the retinal image edge or lesions could be misinterpreted as an optic disc.Finally, post-processing is performed on the segmentation map to eliminate disturbance pixels._ Fovea boundaries become nonsharp due to poor or overexposed illumination, including light reflections from the light source of the camera.With the proposed method, segmentation of the real fovea region is achieved.
In addition, DropOut and batch normalization were applied with 0.3 probability to prevent overfitting.Moreover, using the early stopping function during training prevented the network from memorizing, allowing the training to stop at the optimal point with this function, training stops if there are no differences in the loss function after 10 iterations.A model between actual and predicted values was analyzed using the dice_coef_loss function, which is defined in Equation 1.
segmentation map is obtained for each retinal image.A bright blob around the retinal image edge or lesions could be misinterpreted as an optic disc.Finally, post-processing is performed on the segmentation map to eliminate disturbance pixels.
 Fovea boundaries become nonsharp due to poor or overexposed illumination, including light reflections from the light source of the camera.With the proposed method, segmentation of the real fovea region is achieved.
In addition, DropOut and batch normalization were applied with 0.3 probability to prevent overfitting.Moreover, using the early stopping function during training prevented the network from memorizing, allowing the training to stop at the optimal point with this function, training stops if there are no differences in the loss function after 10 iterations.A model between actual and predicted values was analyzed using the dice_coef_loss function, which is defined in Equation 1. .

�1�
The Dice coefficient, which is shown as DC in the equation represents the overlap between a probabilistic map and the ground truth that is used to determine the difference between the two.
As shown in the equation,  � represents the ground truth class, and  � represents the probabilistic class.There is a better effect on measuring small objectives [21].The modelʹs accuracy and loss curves are shown in Figure 4.
The Dice coefficient, which is shown as DC in the equation represents the overlap between a probabilistic map and the ground truth that is used to determine the difference between the two.As shown in the equation, t i represents the ground truth class, and p i represents the probabilistic class.There is a better effect on measuring small objectives [21].The model's accuracy and loss curves are shown in Figure 4.

Experimental Results for U-Net Model
In the literature, there are several open-access databases on which disease analysis can be made from fundus images.In this study, the test stage took place on images in the Messidor, IDRID, STARE, DRIVE, DI-ARETDB0 and DIARETDB1, HRF, and APTOS databases shown in Table 3.In addition, a special database called ONH-NET was created based on 189 retinal images obtained from Düzce University, Department of Ophthalmology.In order to use these images in this study, an ethics committee report was obtained from Duzce University Department of Ophthalmology.
By using our method, each segmentation is quantified in terms of Sensitivity, Recall, Precision, F-score (Dice Coefficient), and IOU (Jaccard Index).Table 4 defines these performance metrics.A True Positive (TP) indicates how many pixels are correctly predicted to be OD pixels; a True Negative (TN) indicates how many pixels are correctly detected to be non-OD pixels; a False Positive (FP) indicates pixels that are incorrectly identified as OD pixels; and a False Negative (FN) indicates pixels that are incorrectly identified as non-OD pixels.All accuracy criteria listed above were calculated for the OD measurement of colour retinal images found in different databases.
For evaluating the model's performance, accuracy criteria were calculated, after first training the proposed model and then testing it a few times.Using the images from MESSIDOR, IDRID, DIARETDB1, HRF, DRIVE, and APTOS databases, as well as images obtained from the Düzce University Research Hospital's Department of Ophthalmology, Table 5 shows the model's accuracy metric criteria.
In Table 6, the study's model is compared to other methods according to the metrics shown in Table 5.A review of the literature summarizes the results of OD segmentation using different methods on the same dataset in Table 6.A comparison is made based on Specificity, Recall, Precision, and F-score (Dice Coefficient), IOU (Jaccard Index).Comparing our method with previous approaches, we find that we outperform most previous approaches on most metrics and have comparable quality with other approaches.
In this study, applications about the segmentation of the fovea which is known as the center of the macula, which is difficult to detect with the naked eye, were also carried out.Figure 5 shows the segmented images of OD and fovea.

Post-Processing for Diagnosing Optic Nerve Hypoplasia
Parameters such as the diameter of OD, boundaries of OD and the distance between OD and the center of the fovea are parameters that are required for the detection of ONH.Two separate models obtained with a U-Net-based semantic segmentation operation were loaded onto the system as shown in Figure 7. Image moments were used to determine different features of contours like the area, circumference, center and bounding box belonging to OD and the fovea.As used in image processing, computer vision, and other related fields, an image moment is a weighted average of the intensity of pixels in the image or a function of such moments.Such moments are usually selected because they have some attractive property or interpretation.f(x, y), which is a continuous 2-dimensional function has the moment of the order (p + q) is defined in equation 2 below [52].boundaries of OD and the distance between OD and the center of the fovea are parameters that are required for the detection of ONH.Two separate models obtained with a U-Net-based semantic segmentation operation were loaded onto the system as shown in Figure 7. Image moments were used to determine different features of contours like the area, circumference, center and bounding box belonging to OD and the fovea.As used in image processing, computer vision, and other related fields, an image moment is a weighted average of the intensity of pixels in the image or a function of such moments.Such moments are usually selected because they have some attractive property or interpretation.f(x, y), which is a continuous 2-dimensional function has the moment of the order (p + q) is defined in equation 2 below [52].( This way, the central coordinates and radius information of OD and the fovea were obtained. ( for p,q = 0,1,2,... Adapting this to a scalar (grayscale) image with pixel intensities I(x,y), the raw image moments Mij are calculated by methods on the same dataset in Table 6.A comparison is made based on Specificity, Recall, Precision, and F-score (Dice Coefficient), IOU (Jaccard Index).Comparing our method with previous approaches, we find that we outperform most previous approaches on most metrics and have comparable quality with other approaches.
In this study, applications about the segmentation of the fovea which is known as the center of the macula, which is difficult to detect with the naked eye, were also carried out.Figure 5 shows the segmented images of OD and fovea.

Post-Processing for Diagnosing Optic Nerve Hypoplasia
Parameters such as the diameter of OD, segmentation operation were loaded onto the system as shown in Figure 7. Image moments were used to determine different features of contours like the area, circumference, center and bounding box belonging to OD and the fovea.As used in image processing, computer vision, and other related fields, an image moment is a weighted average of the intensity of pixels in the image or a function of such moments.Such moments are usually selected because they have some attractive property or interpretation.f(x, y), which is a continuous 2-dimensional function has the moment of the order (p + q) is defined in equation 2 below [52].( In some cases, it may be possible to calculate this by considering the image as a probability density function, e.g., by dividing the above by Area of binary images = M00 . ( This way, the central coordinates and radius information of OD and the fovea were obtained. ( In some cases, it may be possible to calculate this by considering the image as a probability density function, e.g., by dividing the above by o other Table es  boundaries of OD and the distance between OD and the center of the fovea are parameters that are required for the detection of ONH.Two separate models obtained with a U-Net-based semantic segmentation operation were loaded onto the system as shown in Figure 7. Image moments were used to determine different features of contours like the area, circumference, center and bounding box belonging to OD and the fovea.As used in image processing, computer vision, and other related fields, an image moment is a weighted average of the intensity of pixels in the image or a function of such moments.Such moments are usually selected because they have some attractive property or interpretation.f(x, y), which is a continuous 2-dimensional function has the moment of the order (p + q) is defined in equation 2 below [52].( This way, the central coordinates and radius information of OD and the fovea were obtained.
. Performance metric measurements for disc and cup segmentation.boundaries of OD and the distance between OD and the center of the fovea are parameters that are required for the detection of ONH.Two separate models obtained with a U-Net-based semantic segmentation operation were loaded onto the system as shown in Figure 7. Image moments were used to determine different features of contours like the area, circumference, center and bounding box belonging to OD and the fovea.As used in image processing, computer vision, and other related fields, an image moment is a weighted average of the intensity of pixels in the image or a function of such moments.Such moments are usually selected because they have some attractive property or interpretation.f(x, y), which is a continuous 2-dimensional function has the moment of the order (p + q) is defined in equation 2 below [52].( This way, the central coordinates and radius information of OD and the fovea were obtained. ( This way, the central coordinates and radius information of OD and the fovea were obtained.After these operations, by drawing circles for which we knew the center points and radii as shown in figure 6, the boundaries of OD and the fovea were obtained.To be able to draw the macula region in the image, the distance between the central coordinates of OD and the central coordinates of the fovea is calculated using the formula After these operations, by drawing circles for which we knew the center points and radii as shown in figure 6, the boundaries of OD and the fovea were obtained.To be able to draw the macula region in the image, the distance between the central coordinates of OD and the central coordinates of the fovea is calculated using the formula �� � � � � � � � � � � � � � � .In this formula,  � is the central y-coordinate of OD, � is the central x-coordinate of OD, Fy is the central y-coordinate of the fovea, and Fx is the central x-coordinate of the fovea.From the difference between the distance found here and the diameter of OD, the radius of the macula is found. .In this formula, OD y is the central y-coordinate of OD, OD x is the central x-coordinate of OD, F y is the central y-coordinate of the fovea, and F x is the central x-coordinate of the fovea.From the difference between the distance found here and the diameter of OD, the radius of the macula is found.
With this operation, as shown in Figure 6, the boundaries of the macula were also obtained.Hence, all information required for the diagnosis of the disease was acquired.In light of these data, the diagnosis of the disease was decided based on the ratio of the OD diameter to the macula radius.If this ratio was smaller than 0.3, it was concluded that the image showed ONH.Therefore, the images where ONH was present were determined by looking at the diameter of OD and the distance of the fovea center to the boundary of OD.
Hence, all information required for the diagnosis of the disease was acquired.In light of these data, the diagnosis of the disease was decided based on the ratio of the OD diameter to the macula radius.If this ratio was smaller than 0.3, it was concluded that the image showed ONH.Therefore, the images where ONH was present were determined by looking at the diameter of OD and the distance of the fovea center to the boundary of OD.  7. According to the results, the res-net architecture modified with u-net resulted in more successful results for both the foveal area and the optic disc.Figure 8 shows the images obtained with the ONH algorithm that we developed in this study.In the literature, several methods that are used for automatically extracting various features from retinal image.Most of these methods focus on detecting or segmenting only one feature.In our study, the boundary of OD, the center of OD, the boundary of the fovea, the center of the fovea and the distance between the centres of OD and the fovea were detected.Table 8 presents a summary of anatomical structures that were found in studies previously conducted on retinal images.
When the studies on this subject are examined, methods have been developed to find only the optic disc boundaries in some studies and only the foveal boundaries in some studies.In our study, the optic disc boundaries, center, radius value, foveal boundaries, central coordinate, macular radius and boundaries in the retina images were obtained.

Figure 8
Retinal structure obtained with the ONH algorithm

Conclusion
To automate diagnosis, it is highly important to find and identify retinal anatomical structures such as the optic disc, macula, and fovea at the retina, which allows early observation of changes related to many ophthalmological diseases.The method proposed in this study not only dealt with the subject of detecting structures like the optic disc, macula and fovea but also presented a decision support system to help doctors by allowing the automated operation of procedures required for the diagnosis of the disease known as optic nerve hypoplasia that can be diagnosed with quantitative calculations.The algorithm we call ONH-NET, developed for optic nerve hypoplasia disease, which has never been diagnosed using deep learning methods before, makes the study completely original.Diagnosing optic nerve hypoplasia requires measurements such as the diameter of the optic disc, its boundaries, and the distance between the centers of the optic disc and the macula.In this study, it was aimed to measure these parameters most accurately and shortly and obtain diagnostic results with the smallest error margins.Existing techniques of optical disc and fovea segmentation have several problems, including variations in illumination, blurry boundaries, occlusion of retinal vessels, large bright lesions that obscure the fovea segmentation, and incorrect segmentation of pathological information.With the improved deep learning U-Net model, we developed a method for segmenting optic discs and fovea.An improved U-Net algorithm combined the encoding layers of the pre-trained ResNet-18 model with the decoding layers of the classical U-Net algorithm was presented in this study as a robust segmentation method for optic discs and foveas.In order to solve the problem of the traditional deep learning U-Net model requiring more depth, we added a residual module.As a result of the improved deep learning U-Net model, low-level information sharing can be prevented and performance degradation can be solved in deep convolutional neural networks under extreme depth conditions by connecting the outputs of the convolutional layer with the outputs of the deconvolution layer.Combining pre-trained ResNet with U-Net is the advantage of the proposed method.Using this method, the network is not trained from scratch, so fewer iterations are required, which prevents overfitting.After a semantic segmentation operation based on U-Net, the diameter and central coordinate information of the detected optic disc and fovea regions was reached.In line with these data, the diagnosis of the disease was decided by looking at the ratio of the optic disc diameter to the distance between the optic disc and the center of the macula.In the case that this ratio was smaller than 0.3, it was concluded that optic nerve hypoplasia was present.In the study, colour retinal images in the Messidor, Diaretdb1, DRIVE, HRF, APTOS and IDRID databases were used as the dataset.Additionally, the application was also tested on retinal images obtained from the Department of Ophthalmology at Düzce University Research Hospital.The performance values of all operations were tested using similarity indices such as the Dice and Jaccard indices, as well as sensitivity, specificity and accuracy performance criteria.An assistive tool was provided for doctors by automatically making the diagnosis of optic nerve hypoplasia, which had not been made using deep learning methods before.Hence, the calculations required for the boundaries of the optic disc, the diameter of the optic disc, the center of the fovea and the distance between the diameter of the optic disc and the radius of the macula were made in an automated manner without needing paper, pencils or calculators.Thus, ophthalmology, deep learning, and image processing skills and expertise have been successfully combined in this study for retinal image analysis and developing disease diagnosis methods.Despite the performance of the proposed method, some limitations remain.Res-UNet's probability map is sometimes discontinuous due to blood, lesions, and nerve fiber layers surrounding the OD, while some incorrect pixels may be included.It is possible to reduce the real OD region when interferences are removed by the proposed post processing.For a more reliable segmentation result, graph cut algorithms may be incorporated into post-processing, including filling OD region gaps and smoothing edges.
The retinal image analysis methods that were developed in this study were not only limited to the detection of certain anatomical structures, but they also presented a decision support system to help doctors by conducting the automated implementation of the procedures required for the diagnosis of optic nerve hypoplasia.The proposed deep learning-based approach will also be a valuable tool that could also be used for other retinal diseases.

Declarations Conflict of interests
There are no commercial affiliations or sources of support that might pose a conflict of interest for the authors.

Figure 1
Figure 1 Anatomical Structures of the Retina

Figure 2
Figure 2 Parameters Used for Morphometric Techniques in the Diagnosis of ONH

Figure 3
Figure 3 Proposed U-Net Architecture used to find Optic Disc and Fovea

Figure 4 .
Figure 4. Training and validation loss curve and accuracy curve

Figure 4
Figure 4Training and validation loss curve and accuracy curve

Figure 6
Figure 6 Structures detected in ONH disease diagnosis We tested the proposed architecture on retinal images in Messidor, Diaretdb1, DRIVE, HRF, APTOS, and IDRID databases.Additionally, 189 retinal images obtained from Düzce University's Department of Ophthalmology were used to create a special database called ONH-NET.
_ This method avoids overfitting by combining pretrained ResNet and U-Net, resulting in faster network training with fewer epochs._

Table 1
Comparison of the CNN methodologies for the detection of Retinal Anatomical Structures with the existing methods in the literature Structures that are simple and flexible. _

Table 2
List of hyper-parameters and their ranges/values

Table 3
Most frequently used databases for the detection of anatomical structures

Table 4
Performance Measure Formulas

Table 5
Performance metric measurements for disc and cup segmentation

Table 6
Comparative performance of optic disc boundary segmentation to previous methods

Table 6 .
Comparative performance of optic disc boundary segmentation to previous methods

Table 7
Results for Optic Disc and Fovea SegmentationFigure 7 Flow Chart Diagram for the Diagnosis of Optic Nerve Hypoplasia

Table 8
A comparison between the proposed methodology for optic disc (OD) and fovea (F) detection and previously used methods