Keywords

1 Introduction

Throughout the world around 314 million people are suffering from Diabetic retinopathy, hypertensive retinopathy, glaucoma. These diseases gradually leads to vision loss of the patient which is a major area of concern in the developing counties [3]. Early identification and treatment can cure more than 85% visual impairments cases. In this field computer aided diagnosis system can make the process faster and assist ophthalmologists to cater more patients in less time.

There are different approaches followed in identification and segmentation of Optic Disc (OD). Welfer et al. [21] identified the OD boundary using morphological operations and watershed transformation technique. Aquino et al. [2] segmented the OD using morphological operations, edge detection method and circular Hough transformation technique. Morales et al. [17] applied inpainting as preprocessing for removing blood vessels and stochastic watershed transformation for determining the OD boundary. Xu et al. [22] applied active contour model (ACM) and proved better segmentation. Lowell et al. [13] applied a direction sensitive gradient-based technique to remove the vessel obstructions and deformable ACM for finding the OD boundary in low resulotion images. Chrastek et al. [6] applied distance map algorithm to remove the blood vessels and then segmented the OD by using sequence of methods like morphological operation, Hough transformation and ACM. The method presented by Joshi et al. [11] improved the robustness of ACM proposed by Chan and Vese [4] by taking care of the variations in the OD region. Morales et al. [17] detected the boundary of optic disc by the principal component analysis.

To overcome all these types of shortcomings deep learning based algorithms are playing an important role because of its ability to learn features during training time. The success of convolutional neural network in object segmentation [5, 7, 18] has motivated us to investigate the performance of fully convolutioal network for optic disc detection and segmentation. The major contributions of the present work are (i) initial segmentation of optic disc using U-Net based fully convolutional network and (ii) removal of false-positives based on anatomy-aware features.

In this paper, Sect. 1 has covered the existing works in this area. Then we have described the proposed segmentation framework in Sect. 2. Experimental results and comparison of the proposed method with the state-of-the-art techniques is provided in Sect. 3. Conclusion and future scope of work is stated in Sect. 4.

2 Segmentation Framework

In our proposed method (Fig. 1) U-net [19] based fully convolutional network has been used for initial segmentation after preprocessing of the input images. Then false positives have been reduced using anatomy-aware features. The U-Net architecture is used for initial segmentation as it provides better segmentation using few number of training images.

Fig. 1.
figure 1

Block diagram of the proposed segmentation framework

Full size image

2.1 Preprocessing

First all the images are resized to \(512\times 512\) pixels. Then red channel image is threshold. Morphological opening, closing and erosion operations [8] with square structuring element are used to create a mask of circular retinal fundus region-of-interest, which allows focusing only on the foreground of retinal images.

2.2 Segmentation Using U-Net

Architecture of U-Net. The U-net [19] is a fully convolutional network which consists of convolution operation for down-sampling, max pooling, ReLU operation, concatenation and de-Convolution operation for up-sampling. The down-sampling path has 5 convolutional blocks and each block has two convolutional layers with a filter size of \(3\times 3\) and stride of 1. Max pooling with stride 2 is applied to the end of every blocks except the last block. The data is propagated through the network along all possible paths and generates the segmentation map at the end of the network. At the end input images of \(512\times 512\) size reduces to \(32\times 32\). The second part of the U-Net is the expansion layer which basically create the high resolution segmentation map. This part consists of a sequence of up-convolutions and concatenation with high-resolution features from contraction path. Therefore, the size of feature maps increases from \(32\times 32\) to \(512\times 512\). High-level information is represented at up-sampling blocks, and low-level features are transferred through skip connection.

Training of U-Net. First data augmentation techniques have been applied on the images of extended MESSIDOR database (MESSIDOR-II) [16] which is then used for training of U-Net from scratch. A stochastic gradient-based optimization [12] (ADAM) is applied to minimize the cross-entropy based cost function. The learning rate for the ADAM optimizer is set to 0.0001 and over-fitting is reduced by using dropout [10]. The weights of background and foreground are maintained as 1:10 and training were performed upto 60, 000 iterations.

False Positive Removal. Morphological opening is applied to separate false positives from Optic disc. E.g., for Fig. 2(a), after initial segmentation Fig. 2(b) is showing some false positives caused by exudates. Compactness feature is used to eliminate false positive candidates from initial segmentation results which will create two objects. The object having bigger size is considered as optic disc Fig. 2(c).

Fig. 2.
figure 2

(a) Original image (b) false positives caused by exudates (c) optic disc candidate.

Full size image

3 Experimental Results and Discussions

3.1 Database Used for Evaluation of Segmentation Result

MESSIDOR. MESSIDOR [15] database contains 1200 colour retinal images, acquired using non-mydriatic camera, Topcon TRC NW6 with field-of-view set to 45\(^ \circ \). Binary mask of the optic disc of MESSIDOR dataset was provided by the experts of the University of Huelva [1].

3.2 Evaluation

The performance of optic disc detection is evaluated using Success Rate (SR) which represents the percentage of retinal images in a dataset where the centroid of optic disc is successfully localized within the boundary of the ground truth mask of optic disc. The performance of optic disc segmentation is evaluated in terms of Overlap Measure (OM) and Mean Absolute Distance (MAD) [20]. The OM represents the ratio of the intersecting area between the actual optic disc and segmented optic disc. MAD represents the mean of the shortest distances from the boundary of the actual optic disc to the boundary of the segmented optic disc.

3.3 Experimental Results

Quantitative Analysis. The evaluation of the proposed segmentation algorithm is performed on MESSIDOR datasets. During testing, we divided the images into three subsets. Out of three subsets, two are used for fine tuning of pre-trained network and remaining set is used for testing. Thus U-Net learns database specific features through transfer learning. The average value of SR, OM and MAD of the proposed framework and competing techniques are provided in Table 1. The OM of the proposed method is larger as compared to the competing techniques. Such improvement of OM is due to the application of fully convolutional network in initial segmentation.

There is only one failure case for optic disc detection for the MESSIDOR dataset. The MAD value of the proposed method is slightly better or comparable with the competing techniques. The high value of OM depict that the segmented mask of optic disc matches accurately with the ground truth mask. A comparison of segmentation performance in terms of percentage of test images included in various OM distributions, is provided in Table 2. The proposed method outperforms the competing techniques at all the four different OM levels such as \(\ge \) 0.7, 0.75, 0.85 and 0.9.

Table 1. Comparative result of optic disc segmentation
Full size table
Table 2. Percentage of images in various OM levels
Full size table

Qualitative Results. We have analysed the proposed framework for images of healthy subjects [Fig. 3(a)–(c)], the images with the presence of pathologies [Fig. 4(a)–(c)], and low contrast [Fig. 4(b)]. These qualitative results reveal that the proposed algorithm is capable of identifying and segmenting the optic disc in bad quality retinal images. Few images with poor segmentation results is shown in Fig. 5(a)–(c). The poor segmentation is due to non-uniform illumination during image acquisition.

Fig. 3.
figure 3

Segmentation evaluation examples for normal fundus images. The contour of ground truth and segmented optic disc is shown in blue and green color respectively. (Color figure online)

Full size image
Fig. 4.
figure 4

Segmentation evaluation examples for images having pathologies,haemorrhage and low contrast. The contour of ground truth and segmented optic disc is shown in blue and green color respectively. (Color figure online)

Full size image

3.4 Implementation

The U-Net architecture is implemented in Python using the PyTorch library in Linux environment using a 8 GB GPU (NVIDIA GeForce GTX 1070 with 8GB GDDR5 memory) on a system with Core-i7 processor and 32 GB RAM.

Fig. 5.
figure 5

Segmentation evaluation examples for some typical cases. The contour of ground truth and segmented optic disc is shown in blue and green color respectively. (Color figure online)

Full size image

4 Conclusion

In the proposed method, fully convolutional network is trained by feeding thousands of varying grades of fundus images, where it is learns the best features on its own. Therefore, the proposed method outperforms the other competing techniques in most of the metrics measurements. The method is also successful in optic disc localization and segmentation, when tested on both dilated and non-dilated types of fundus images. The performance of this algorithm does not degrade while handling images containing strong distractors like yellowish exudates which prove the effectiveness and robustness of the proposed process. In future more research needs to be accomplished by testing on other independent datasets.