Related Work

The semantic segmentation task is one of the essential tasks in computer vision in recent decades and even more so if we focus on medical images since they have become an aid in detecting diseases. It has been proven that using neural networks to perform this task yields good results. However, when we talk about semantic segmentation, it is prevalent that convolutional networks are used from start to finish.

The development of these networks for medical images has been used in countless applications, whether in the eyes, chest, abdomen, brain, or heart. The database used for this article is from the retina, that is, from the human eye. Therefore, we will contribute with state of the art of retina analysis. One of the challenges is the segmentation of veins or blood vessels. It depends on the technique to obtain an image, such as an eye fundus image or a tomography taken to some specific area. In ^{[3] [4]}, use a residual network based on U-Net using a batch normalization and an automatic encoder named PixelBNN, which segments the veins in fundus images and is tested in various databases to achieve acceptable results; down testing time was their contribution. For Xiao et al. and Son J et al., the contribution to RV focused on reducing the edges of the veins, making them more precise and sharper ^{[5] [6]}. Other ocular applications in the area of glaucoma where the main task is to segment the optic disc, in ^{[7] [8]} use a ResU-Net and propose using a modified FCN; the image takes to center in the optic disk to reduce noise in fundus images.

RV segmentation

OCT-A is a non-invasive technique developed and applied relatively a few years ago, through which the vascular structures and choroids can be visualized in the retina. In turn, the segmentation of retinal vessels, the obtained vasculature through this technique, is an available research opportunity to make a difference from the methods studied for more than 20 years ^{[9] [10] [11]}, such as those based on techniques such as deep learning, filters, and some classifiers for VR segmentation in fundus images. Implemented the method used in image thresholding in ^{[4] [5]} to determine the density of the veins.

However, vector and exponential flow optimization algorithms have been among the best algorithms implemented in OCTA images for RV enhancement and segmentation. The implementation based on Gibbs Markov and its field model in ^{[12] [13] [14]} was applied in the projection maps for the RV segmentation to segment retinal vessels into different projection maps.

The papers of ^{[8] [9] [10]} proposed a modified U-Net for circle segmentation in a maximum intensity projection of OCT-A. Wylęgała et al. present in ^[11] a new partition-based modular array to detect thick and thin vessels separately.

FAZ segmentation

The FAZ is the fovea region where the absence of blood, the vasculature, can be observed. A new field of study is the segmentation of the FAZ through OCT-A images compared to retinal vascular segmentation. The authors of the papers ^{[12] [13] [14]} introduced an active contour model for FAZ detection. Li et al. and Bates et al. present in ^{[15] [16]} a series of morphological factors to identify FAZ candidates in OCTA projections using two types of fields of view. A modified U-Net to segment the FAZ on the OCTA projection map of the superficial retina was introduced in the papers ^{[9] [16]}. Through a CNN, Azzopardi et al. develops a segmentation for the FAZ and the nonischemic capillary region ^[12].

Dataset

The plexuses database deployed in this work is the OCTA-500 ^[15], which consists of images collected from 500 subjects taken under identical conditions (projection maps). The retina image size is 3mm x 3mm, with 304px by 304px collected from 200 subjects, and 6mm x 6mm images, with 400px by 400px for the remaining 300 subjects. Images with the same number and characteristics were collected for OCT and OCT-A.

Figure 1 shows the OCT and OCT-A projection maps from the OCTA-500 database. The corresponding OCT projections and the full OCT-A projection correspond to 3D volumes from averaging in the axial direction. On the other hand, the Inner Limiting Membrane (ILM) - Outer Plexiform Layer (OPL) projections, as well as the OPL - Brunch Membrane (BM), respond to a maximum in the same direction.

Figure 1 Projection maps for each of the techniques (All images are retrieved from the database ^[15]) 

U-Net

U-Net is a network based on a "Fully Convolutional Network" (FCN) which was designed, tested, and presented in 2015 mainly for the segmentation of biomedical images. The value and reliability that U-Net offers, with its 23 convolutional layers, has been shown in ^[17] for the field of images of medical origin. A network that does not entirely connect all its layers and when taking the shape of a "U" reflects almost a symmetry in the expansive and contraction paths.

Due to the nature of the network in this article, it was decided to use the same architecture proposed in ^{[17] [18]}, which can be defined in general terms as a network formed by a series of encoders located in the first half of the network, which allows the data to be contextualized, input is called the contraction path. This first part of the network is built from unpadded convolutions, followed by a rectified linear unit activation layer ReLU ^[19] for each previous convolution.

Finally, the resolution reduction applies a maximum grouping operation. In the second half, the expansion path, a series of decoders perform the exact location of the network by determining it from an upsampling of the function map, followed by an up-convolution, cropping, and two convolutions that, like in the first half it is followed by a ReLU layer for each of them.

Deep ResU-Net

ResU-Net is a semantic segmentation neural network that merges the strengths of the residual neural networks and the U-Net, thus obtaining a Deep ResU-Net. This combination provides two advantages: 1) the residue facilitates the formation of meshes; 2) it facilitates the lossless transmission of information between the low and high levels of the network and within a residual unit with the elimination of connections, which allows for designing a neural network with fewer parameters ^{[20] [21]}. It can achieve similar performance and is consistently better than semantic segmentation.

The architecture of the ResU-Net network comprises 56 convolutional layers; the network proposal by ^[20] solves the degradation of the model in the deeper layers.

A model builds to the base of the pair of convolutional layers corresponding to the expansive path in the U-Net; a residual learning block replaces these. A residual learning block comprises three convolutional layers, the first two with a ReLU activation and three batch normalization layers ^[21]. Hence the ResU-Net on the expansive path includes an input unit, a central unit, and a residual unit; the contraction path comprises four concatenated blocks, adding a block and an output unit.

Fully Convolutional Network

A Fully Convolutional Network (FCN) can be used for semantic segmentation, which means it uses convolution and deconvolution layers. For the development of this research, a FCN network was developed and tested by ^[22] based on a VVG-16 as the authors mention higher accuracy compared to other FCN based on either AlexNet or GoogLeNet. Three variations are known for the FCN network, FCN8s, FCN16s, and FCN32s; in our case, a FCN8s network is used. The discrepancy between these three networks lies in the resolution reduction because the semantic information requires a detailed level of recovery in the connection of the intermediate layers.

Evaluation Metrics

It is essential to mention that due to the nature of the problem in the segmentation of both the FAZ and the RV, the background of the image has a more significant presence than each of the parts to be segmented. Therefore, the metrics ^[23] chosen to evaluate the segmentation performance of each network quantitatively are established in Equations (1) - (5):

• Dice Coefficient (DICE):

• Jaccard index (JAC):

• Balance-precision (BACC):

• Precision (PRE):

• Recovery (REC):

These metrics allow comparing our models with the current state-of-the-art. TPR refers to the true positive rate, TNR negative rate, TP to True Positive, and FN to False Negative.

Manual Segmentation

The ground truth (GT) used to validate the results of the following training sessions is one for each feature.

The ground truth (GT) used to validate the results differs from the given to the database. A manual modification is made to the image provided where the features are isolated, and the original GT has both features FAZ and RV. Figure 2 shows the segmentation of the FAZ; the GT is unique for this feature, doing away with the RV and only placing a black background in the rest of the image. This same process is carried out for the GT of the RV.

Figure 2 a) GT OCTA-500; b) GT of FAZ; c) GT of RV. 

Process

Model training is performed using 80% of the images from the full projection, which correspond to 400 photos from the OCTA-500 database, and the remaining 20% was used for verification. Resizing 304 px by 304 px is applied for the test and training set to keep data in standard dimensions. This procedure is performed under the same conditions and characteristics of retinal imaging.

It uses the U-Net network while preserving the structure and hyperparameters of the original model using the Stochastic Gradient Optimizer (SGD). The training is performed for 50 epochs for both databases. Figure 3 shows the evolution of the proposed methodology in detail.

Figure 3 Metodology. 

Additional training is carried out with a technique that performs a better segmentation of the validation image and compares it with the ground truth, provided and validated by the same database used (OCTA500). OCT-A technology provides the best measurements; however, the goal was to segment the FAZ, and in the results, blood vessels can be observed in the full segmented view of both techniques. In addition, manual segmentation was performed in the target image to enhance the above segmentation, leaving only the visible FAZ, ignoring the vascular system. It is important to note that this training is conducted under conditions similar to those previously applied.

Two projection maps from the database were mainly used for RV segmentation. One of them is the maximum projection map of the retina and the full projection map. These two maps can be found for the two imaging techniques, that is, OCT and OCT-A, which allows a more significant number of examples for training.

Technical Considerations

All experimentation is done in TensorFlow using an NVIDIA GeForce GTX 1080Ti GPU. The standard normal initialization method is used to initialize the network with a difference of 0.02 batch size of 3 as a cross-entropy loss function, and Adam as the optimizer was implemented.

Overtraining is avoided by evaluating the similarity between the validation and training sets throughout the network training process. Further choosing the training model with the best performance and trying to have the best value of this similarity by calculating the dice coefficient.