Eye is an exclusive organ for sense of sight in human. Retina is a light sensitive layer that lines at the back of the eye are affected due to vascular occlusion. Ophthalmologic diseases such as diabetes retinopathy, hypertension and arteriosclerosis are the main cause of vascular occlusion. These ophthalmologic diseases are related to morphological changes in vascular diameter and branching pattern leading to blindness. Segmentation of retinal vessel are done to analyze these morphological changes in retinal vessel. However, due to presence of illumination, multiplex distribution of blood vessel, low contrast between target and background the segmentation of retinal blood vessel is a challenging task. In this paper we propose a method to segment retinal blood vessels based on fully convolutional neural network and pixel classification with cross entropy function to avoid class imbalance problem. Our method provides an automatic segmentation of retinal blood vessel.

Introduction

There are lot of people in the world who suffer from blindness due to retinal diseases which leads to morphological changes in retinal blood vessels. The important requirement for diagnosis of retinal diseases involves segmentation of retinal blood vessel. Therefore accurate segmentation of retinal blood vessel is of greater importance for dignosis and treatment of retinal diseases. Retinal images such as two dimensional (2-D) colour fundus images and three dimensional colour fundus images are used for diagnosis of opthalamologic diseases. Experts or specialist segment the retinal vessel manually from these fundus images. This leads to time consuming and improper segmentation of vessel. So an automatic segmentation of retinal vessel is proposed for reliable and robust segmentation. Segmentation of retinal blood vessels have been proposed using various methods with the development of computer aided system. These methods can provide robust segmentation of retinal blood vessels without the requirement of experts. segmentation can be done either supervised or unsupervised. Salem et al. proposed an unsupervised method of retinal vessel segmentation by distance based principle using RAdius based Clustering ALgorithm (RACAL). Wang et al. proposed a segmentation method based on pixel classification in which CNN is used as an feature extractor and employed random forests as a trainable classifier. Marin et al. proposed a method which involves feature extraction for every pixel such as grey level features and momentum features and then these features are applied to neural network for classification. In general retinal vessel segmentation are based on vessel tracking Yin et al. or region growing technique Fraz et al.

Conventional supervised method involves feature extraction and classification. Feature extraction is a trivial task, because selecting best feature is important for segmentation of retinal blood vessels. Choice of feature can affect the segmentation result. Many recent approaches or works use convolution neural network for segmentation of retinal blood vessels. Convolutional neural network extract best feature by performing convolution, batch normalization and pooling operations.

Liskowski et al. , author proposed patch wise segmentation of retinal vessel using CNN, which requires more memory and is a time consuming process. Orlando et al. proposed a method to reduce luminous and contrast variation in retinal images for better segmentation of retinal vessel using local normalization and proposed a CNN architecture which takes input of arbitrary size and produce output with efficient and inference learning. Zhenkiang et al. , author used a pre-trained network alexnet for segmentation of retinal blood vessel. Zhexin et al. , author cosidered green channel as it provide better contrast between vessel and background and developed an architecture of neural network with four convolution and one fully connected layer. Jose et al. , proposed a CNN architecture where class balance problem is not taken into consideration which lead to segmentation of majority of non vascular region. Kai et al. , proposed a method for segmentation of retinal blood vessels using multiscale CNN to get a probability map and then proposed a class balance cross entropy loss function to improve the performance of segmentation.

Proposed method

Overview: In the proposed work, the input image is preprocessed to eliminate illumination and low contrast between vessels and background. The image is then enhanced to get detail information of vessel edges. Features are then extracted using multiscale CNN and are classified based on pixel clasification. To avoid missclassification of vessel as non vessel, as vessels has less number of pixel count compared to non vessels pixel count, a class balanced cross entropy loss function is sperformed. The green channel of the fundus image is taken to reudce low contrast differnce between vessels and background. The image is then local normalized to reduce illumination effect which is further enhanced by Contrast Limited Adaptive Histogram Equalization (CLAHE).

Green Channel Extraction

The green channel provides better contrast between vessels and background, while red channel is saturated and blue channel is the darkest channel and so it does not contain any information. So inorder to obtain better contrast between vessels and background green channel of image is taken for robust segmentation of blood vessels.

Local Normalization

Local normalization tends to normalize the local mean and variance of image. This is used to normalize the illumination of fundus image so that the vessels can be well distinguished from background. Local mean and variance are estimated by local spatial smoothing using gaussian filters. Mean of green channel image is subracted from the green channel image and is divided by variance of green channel fundus image, so as to normalize each features. Smoothing windows sigma1 and sigma2 determine the estimation of local mean and variance.

Contrast Limited Adaptive Histogram Equalization

The green level transform image does not have enough contrast between vessels and background which leads to improper segmentation of retinal blood vessels. In this proposed work contrast limited adaptive histogram equalization (CLAHE) is applied to enhance the contrast of image. Most medical image use this enhancement technique to enable the principle part more visible. Convolutional neural network Convolutional neural network consists of muliple layer. These layer extract features and classify image. The convolution neural network is composed of an input layer, an output layer and multiple hidden layer. The hidden layer involves convolution layer, pooling layer, batch normalization layer.

Convolution layer in each stage contains sixty four filters of filter size three, padding in all four sides and striding 1. Batch normalization layer is done between each convolution layer and Relu layer to speed up training of convolutional neural networks and reduce sensitivity to network initialization. Rectified linear unit is an activation layer, do not vanish in extremes and so allow effective training of networks with dozens of layers. Pooling operation performed in our proposec method is 2×2 maxpooling, that is maximum value of 2X2 size is choosen. Information provided by each stage are indispensable to our work. Output of each stage is upsampled to size of input image. The feature map of each stage are merged by convolution operation.

Softmax is performed before pixel classification. The network uses a pixelclassification layer to predict the categorical label for every pixel in an input image.

The pixel for vessel and non-vessel are imabalanced. Only 10% of pixel are labeled for retinal vessels in one fundus image and majority of pixels are labeled for non-vessel which results in segmentation of dominant class. To overcome this problem class weighting is done to balance the classes. Class frequency are obtained by dividing each class weight by total class weight. Class frequency F is presented by equation (3) i=1, . . , N (3) Inverse class weight is obtained by taking inverse of class frequency. Inverse class weight is presented by equation (4). i=1, . . . , N (4) Pixel classification is done by final set of layers, softmax and pixel classification layer. These two layers combine to predict the categorical label for each image pixel. Pixel classification layer is updated with class weight. Better segmentation result are obtained by using class weighting to balance the classes.

Experiment evaluation

Dataset Experiment is conducted on DRIVE dataset for vessel segmentation of reinal colour fundus images. Drive contains 40 images in which twenty images are training images and remaining twenty are test images. Each tarining have one groundtruth segmented by specialist. Each test image has two groundtruth in which one is a groundtruth and other is a gold standard for the groundtruth. Evaluation metric Evaluation metric are used to evaluate the performance of the result. Binary segmentation result conatin four cases: true positive (TP), false negative (FN), true negative (TN), and false positive (FP). If the vessel is predited correctly as vessel it is defined as TP and those are wrongly classified as non vessel pixels are counted as FN. Non vessel pixel correctly predicted as non vessel is defined as TN, and the non vessel pixel wrongly predicted as vessel are defined as FP. Performance is evaluated by four indicators: Sensitivity (Se), specificity (Sp), Accuracy (Acc).

Result

The result obtained at each stage are important for vascular segmentation. Result obatin at low level stage is used for detail detection and high level stage information are used for learning vascular structure. Therfore combination of stages provide better segmentation result. It is also observed that validation accuracy increases with number of iteration. The network is trained for hundred iteration. Epoch initialized for the training phase is hundred. The number of epoch is a parameter that defines the number of times the training algorithm learn the entire training dataset. It is obsorved from the graph that the training accuracy increased with number of iteration initially and it remains constant after certain range of iterations.

In the proposed deep learning framework, the improvements for Se, Sp and Acc mainly are due to cross entropy loss function model to learn better discriminative features to classify non-vessel and vessel pixels.

Conclusion

In our proposed method the input image is resized which leads to loss of tiny vessels. Segmentation of retinal vessels using CNN with inclusion of batch normalization and with the improved cross entropy loss function have optimized the segmentation of retinal blood vessel to some extend.