Optic Disc Preprocessing for Reliable Glaucoma Detection in Small Datasets

Abstract

Glaucoma detection is an important task, as this disease can affect the optic nerve, and this could lead to blindness. This can be prevented with early diagnosis, periodic controls, and treatment so that it can be stopped and prevent visual loss. Usually, the detection of glaucoma is carried out through various examinations such as tonometry, gonioscopy, pachymetry, etc. In this work, we carry out this detection by using images obtained through retinal cameras, in which we can observe the state of the optic nerve. This work addresses an accurate diagnostic methodology based on Convolutional Neural Networks (CNNs) to classify these optical images. Most works require a large number of images to train their CNN architectures, and most of them use the whole image to perform the classification. We will use a small dataset containing 366 examples to train the proposed CNN architecture and we will only focus on the analysis of the optic disc by extracting it from the full image, as this is the element that provides the most information about glaucoma. We experiment with different RGB channels and their combinations from the optic disc, and additionally, we extract depth information. We obtain accuracy values of 0.945, by using the GB and the full RGB combination, and 0.934 for the grayscale transformation. Depth information did not help, as it limited the best accuracy value to 0.934.

1. Introduction

Glaucoma is an illness that causes blindness in people of any age, but commonly in older adults. This hereditary disease damages the eye’s optic nerve, which usually happens when fluid builds up in the front part of the eye. That extra fluid increases the pressure in the eye (aqueous humor), damaging the optic nerve. Although it is permanent damage and cannot be reversed, medicine and surgery may help to stop further damage. That is why an early diagnosis of glaucoma is very important. The most common way of detecting it is carried out through different analyses that involve the use of tools that are in contact with the patient’s eye, such as tonometry, that consists in applying a small amount of pressure to the eye by using a tonometer or by a warm puff of air, to measure the inner eye pressure; ophthalmoscopy, that is a procedure that consists in dilating the pupil through eye drops, and examining the shape and color of the optic nerve; or gonioscopy, which aims to determine whether the angle where the iris meets the cornea is open and wide or narrow and closed, that defines the type of glaucoma, which is made by a hand-held contact lens placed on the eye (https://www.glaucoma.org/glaucoma/diagnostic-tests.php, last accessed 2 September 2021).

To detect glaucoma, in this work, digital retinal fundus images are required, as the glaucoma is mainly detected in the optic disc, through which the optic nerve transport to the brain the visual information. The image of the optic disc provides the necessary information to detect glaucoma, as shown in Figure 1. In this work, we propose a methodology consisting in preprocessing digital images and extracting their color planes, to compare the internal information this kind of image can contain. We focus our analysis on the optic disc by using some preprocessing techniques, as this element provides all the information about the optic nerve and shows how glaucoma has evolved in the eye. Additionally, estimated depth information will be added as an image to see if it can help the classification methodology, to correctly detect glaucoma by using Convolutional Neural Networks (CNNs).

The content of this work is divided as follows: Section 2 shows some related works; Section 3 describes the methodology employed in this work; Section 4 describes the experiments carried out, and Section 5 presents the conclusions of the paper.

2. Related Work

Glaucoma detection using CNNs is a common task in computer science, as many solutions used so far have shown good results in classification tasks as mentioned by Sultana et al. [2]; other recent works using CNNs to classify digital retinal fundus images with glaucoma are mentioned below. Li et al. [3] use the Inception [4] CNN model trained with their own private dataset; they trained it with the full RGB retinal fundus image and classify them in positive or negative glaucoma. Fu et al. [5] created a four stage CNN named DENet, that consists in first locating the optic disc in the full retinal fundus image, then extract it and classify it; on the other hand, in another stage carries out a classification from the full RGB retinal fundus images. Finally they obtain a final classification using the results of the previous stages; to perform their experiments they used the SCES dataset [6]. Raghavendra et al. [7] proposes a CNN model composed of 18 layers trained with their own dataset, using the full RGB image.

Dos Santos Ferreira et al. [8] use a two stage methodology; the first one consists in using the U-Net [9] model to extract a binary representation of the optic disc and the second stage classifies this representation into two classes. They use the DRISHTI-GS dataset [10] to perform their experiments. Christopher et al. [11] compare the results of different CNN architectures such as VGG16 [12], Inception [4] and ResNet50 [13] with their own dataset. Additionally, they mention the importance of data augmentation and transfer learning as additional information for training the CNNs.

Chai et al. [14] propose a multi-stage CNN trained with their own data, extracting the first stage features from the full digital retinal fundus image; the second stage extracts features from a subimage containing only the optic disc; the final stage joins both features extracted from the previous stages and performs the classification using fully connected layers. Bajwa et al. [15] propose a two-stage framework in which the first one consists in extracting the area that contains the optic disc from the digital retinal fundus image and the second stage consists in a CNN that extracts features and classifies them into two classes. They use the DIARETDB1 dataset [16] to perform their experiments. Liu et al. [17] use a large collection of own fundus images and the CNN model ResNet as classifier and perform previously a statistical analysis on their data.

Finally Barros et al. [1] perform a deep analysis on all the machine learning algorithms and CNNs applied to glaucoma detection using many datasets. In

Table 1. Results obtained from the state of the art (higher is better [1]).

Paper	Accuracy	Precision	Recall	Dataset
Liu et al.  [17]	0.9960	0.9770	0.9620	Private
Bajwa et al. [15]	0.8740	0.8500	0.7117	DIARETDB1 [16]
Chai et al. [14]	0.9151	0.9233	0.9090	Private
Chistopher et al. [11]	0.9700	0.9300	0.9200	Private
Dos Santos Ferreira et al. [8]	1.0000	1.0000	1.000	DRISHTI-GS [10]
Raghavendra et al. [7]	0.9813	0.9830	0.9800	Private
Fu et al. [5]	0.9183	0.8380	0.8380	SCES [6]
Li et al. [3]	0.9200	0.9560	0.9234	Private

we show in an arbitrary order a comparison amongst the results obtained from all previously mentioned works, which were obtained according to the dataset each work used.

Although there are several works solving this task, the main problem is related to the available data, because most of the datasets are private, due to lack of adequate public data; another problem is that the CNN models used in many related works are complex models that require high amounts of data (i.e., above 1000 images). In this work we propose the use of a simple and accurate CNN model that can be trained with a low quantity of data extracted from our own dataset.

3. Proposed Methodology

The methodology used for this work consists of firstly preprocess the digital fundus image with the purpose of obtaining useful information about glaucoma such that one located in the optic disc. Once we have this information, we will use a CNN model to estimate depth information from the extracted information. This depth information consists in a representation of the distance between the user’s point of view and the objects contained in the image; in this work we assume that the further object is the optic disc. Finally, we set out to use all the visual information about the optic disc and depth estimation in order to train a CNN capable to detect glaucoma.

3.1. Image Preprocessing

Glaucoma is a disease that deteriorates the optic nerve that carries visual information to the brain though the optic disc. For that reason we will focus our analysis precisely on the optic disc. To perform its extraction automatically, digital image processing techniques are used. Firstly we extract a Region of Interest (ROI) from the full retinal fundus image containing the optic disc as it will be fed into a CNN to obtain the eventual presence of glaucoma. In order to extract the ROI from the RGB image, we apply image thresholding to the grayscale image in a simple and easy way, to obtain a binary representation of the image; this binary representation contains clear information about the optic disc, as this is always the brightest zone in the RGB image.

After obtaining this binary image showing clearly the optic disc, we compute the centroid of it; then we align the center of the ROI with the previously calculated centroid and extract a subimage from the original color image containing the optic disc, since it is the only retinal element that contains information related to glaucoma in optical retinal images; using only this patch as input of a CNN, we can classify the image if it is a glaucomatous or non-glaucomatous one. In Figure 2 we show a resume of the preprocessing steps taken in this work.

In the next sections we will describe in detail the procedures that we have previously exposed in general. As a first step before the preprocessing, we normalized the size of all the images to 720 × 576 pixels, because the size may vary between all the images in the dataset.

3.1.1. Image Thresholding

Thresholding is the image processing operation that converts a multi-tone (graylevel) image into a bi-tonal (binary) image. This was carried out by the well-known Otsu threshold algorithm [18], $T_{Otsu}$. It was derived from the histogram of the grayscale image intensity values, h, which typically has $L=256$ bins for 8-bit pixel images. Any chosen threshold $0\le T\le L$ partitions the histogram into two segments: the optic disc and the background. The number of pixels w (Equation (1)), weighted mean intensity $\mu$ (Equation (2)), and variance ${\sigma }^2$ of both zones (Equation (3)), respectively, are given by:

$$w_0\left(T\right)=\sum _{i=0}^{T-1}h\left(i\right)w_1\left(T\right)=\sum _{i=T}^{L-1}h\left(i\right)$$

(1)

$${\mu }_0\left(T\right)=\frac{1}{w_0}\sum _{i=0}^{T-1}ih\left(i\right){\mu }_1\left(T\right)=\frac{1}{w_1}\sum _{i=T}^{L-1}ih\left(i\right)$$

(2)

$$\begin{array}{cc}\hfill {\sigma }_0^2\left(T\right)& =\frac{1}{w_0}\sum _{i=0}^{T-1}h\left(i\right){\left(i-{\mu }_0\left(T\right)\right)}^2\hfill \\ \hfill {\sigma }_1^2\left(T\right)& =\frac{1}{w_1}\sum _{i=T}^{L-1}h\left(i\right){\left(i-{\mu }_1\left(T\right)\right)}^2\hfill \end{array}$$

(3)

The threshold of Otsu $T_{Otsu}$ (Equation (4)) was then defined as the threshold that minimizes within-cluster variance:

$$T_{\mathit{Otsu}}=\underset{T}{argmin}{w}_0\left(T\right){\sigma }_0^2\left(T\right)+w_1\left(T\right){\sigma }_1^2\left(T\right)$$

(4)

or equivalently maximizes the between-cluster variance (Equation (5)), which reduces to:

$$T_{\mathit{Otsu}}=\underset{T}{argmax}{w}_0\left(T\right)w_1\left(T\right){\left({\mu }_1\left(T\right)-{\mu }_0\left(T\right)\right)}^2$$

(5)

The search for $T_{Otsu}$ was performed by testing all values of T that minimized Equation (4) or maximize Equation (5). Afterwards, thresholding (Equation (6)) was performed globally:

$$B(i,j)=\left\{\begin{array}{cc}0\hfill & I(i,j)<T_{\mathit{Otsu}}\hfill \\ 255\hfill & I(i,j)\ge T_{\mathit{Otsu}}\hfill \end{array}\right\}$$

(6)

where $(i,j)$ represent the pixel coordinates, I represents the actual grayscale value and B is the resulting threshold value.

Figure 3 presents as an example, the thresholding resulting from the graylevel image obtained from the shown RGB image. Once the binary image was obtained, the centroid of the binary image was calculated according to the procedure explained in Section 3.1.2.

3.1.2. Calculation of Centroids

The centroid of a binary image is given by the arithmetic mean of the position of all pixels that conformed to a shape of an object. Each shape contained in a binary image was composed of white pixels, and the centroid is the average of the coordinates of all the white pixels constituting the shape. On the other hand, an image moment is an average of all the pixel intensities contained in an image. First we find the image moments ${\mu }_{0,0}$, ${\mu }_{1,0}$ and ${\mu }_{0,1}$ of the binary image using Equations (7) and (8), where w is the width and h is the height of the image. In this case, f takes the pixels on the $(x,y)$ coordinates with the value of 1 in the object, as this operation is performed in a binary image:

$${\mu }_{0,0}=\sum _{x=0}^w\sum _{y=0}^{h}f(x,y)$$

(7)

$${\mu }_{1,0}=\frac{su{m}_x}{{\mu }_{0,0}}{\mu }_{0,1}=\frac{su{m}_y}{{\mu }_{0,0}}$$

(8)

To obtain the sum of x and y coordinates of all white pixels, we used Equation (9):

$${\mathit{sum}}_x=\sum \sum xf(x,y){\mathit{sum}}_y=\sum \sum yf(x,y)$$

(9)

Finally the coordinates of the centroid were given by Equation (10):

$$C_x=\frac{{\mu }_{10}}{{\mu }_{00}}{C}_y=\frac{{\mu }_{01}}{{\mu }_{00}}$$

(10)

$C_x$ is the x coordinate and $C_y$ is the y coordinate of the centroid and $\mu$ denotes the Moment (https://docs.opencv.org/3.4/dd/d49/tutorial_py_contour_features.html, last accessed 2 September 2021).

3.1.3. Patch Extraction

As we mentioned before, glaucoma manifests in the optic disc; sub-images are required to be square for simplicity, given the circularity of optic discs. It is required to have an odd number of pixels for the sub-image to have a center both horizontally and vertically. On the other hand, the dimensions only depend on the size of the images and that of the optic disc they contain. That is why we propose empirically a square ROI of 173 × 173 pixels. Once we obtained $C_x$ and $C_y$, we located them in the original image, we aligned the center of the ROI $C_x^{\prime }$ and $C_y^{\prime }$ with $C_x$ and $C_y$ and extracted a subimage or patch of the full image that contained the optic disc. Figure 4 depicts this operation.

3.2. Depth Estimation

Depth estimation was used to calculate the distance between the user’s point of view and the object in the image; in the case of glaucoma classification, this cue was important as it could show the status of the optic disc. As can be seen in Figure 5, depth information can show a different perspective of the cup height inside the optic disc. The pixels with a value near to 0 or black pixels represent the optic cup.

In this work we used this cue as an additional input channel in order to add more features to the RGB data. We obtained it using the proposal of Shankaranarayana et al. [19], which consists of a CNN model capable of estimating depth from RGB images of the optic disc. We implemented their model and trained it with the INSPIRE-stereo dataset [20] to obtain the depth estimation. Once the network is trained, we obtained results from our own dataset (see Section 4). Some examples are shown in Figure 6.

3.3. CNN Model

The CNN model used in this work was based on the original AlexNet [21], given that this model has shown good results in classification tasks [22,23,24]. This CNN consisted of six convolutional layers with Rectified Linear Unit (ReLU) [25] as activation function combined with Max-pooling layers used as feature extractors. After the convolutional layers, as classification layer, we used two fully connected layers with 1024 neurons each one with ReLU as activation function. The output of the model was obtained from two neurons that classified if the retina presented glaucoma or not. The implementation of this CNN model is shown in Figure 7.

4. Experiments and Results

In order to train the proposed CNN, we collected a private collection of retinal RGB images from real patients, 257 images labeled as normal and 109 images with glaucoma certified by two ophthalmologist specialists (Glaucoma Dataset, Centro de Investigación en Computación, Instituto Politécnico Nacional, available at http://cscog.cic.ipn.mx/glaucoma, last accessed 2 September 2021). This database (366 images) gathers images of the retina of both eyes provided by two cooperating private ophthalmologists as specialists, in a project on analysis of retinal images carried out a decade ago at the Computer Research Center of the National Polytechnic Institute. These images were of specific patients of both ophthalmologists, who, motivated by the due professional secrecy to which they are due by their profession, were provided to us without details of the patients to whom they belonged (name, sex, ages, systemic diseases they suffered, etc.). All the images were of Mexican natives who came to them as patients, in order to be consulted to learn about the disease that afflicted them when they noticed deficiencies in their vision systems, namely, glaucoma, diabetic retinopathy, hypertensive retinopathy, retinitis pigmentosa, among other. For us, knowing details of the images did not play any role in order to later carry out statistical or other analyzes. The manual classification of the images was done by the two ophthalmologists, which served for countless scientific publications as a result of the automatic analysis of the system that was being developed at that time.

We used data augmentation to randomly increase the amount of training data by adding a modified version of the existing data; in our case we used the mirroring operation done by reversing the pixels horizontally. In a horizontal mirroring, the pixel positions located at coordinates $(x,y)$ were situated at coordinates $(image\_width-x-1,y)$ in the new image. In order to train the CNN model, we randomly divided the full dataset into a training set (275 images) and testing set (91 images); we expanded the training set from 275 images to 550 images using the horizontal mirroring method.

4.1. Color Plane Extraction

To help the CNN model classify color images for glaucoma detection, we decided to extract and combine color planes obtained from the original RGB image and see if this information may be useful for the task. In Figure 8 we show the images obtained from the color planes and some of their combinations. Of them, we discarded the use of the red, blue, and red + blue planes, due to the lack of information and low contrast for training. This step was not contained in the preprocessing stage, because its main objective was the extraction of the optic disc; we created a different dataset for each one of the extracted planes from the optic disc image, i.e., we created a dataset for the red plane and the blue plane separately. We also obtained the grayscale image by using the weighted method [26]: $Grayscale=0.299R+0.587G+0.114B$. In the CNN model, the input image had the follow dimensions: $173\times 173\times c\times n$, where $c=1$ when the input is the grayscale image or the red, green or blue plane, $c=2$ when the input was the combination of two color planes, $c=3$ when the input was the RGB image and $c=4$ when the input was the RGB image plus depth information.

According to what we explained in previous section, we proposed the following 14 experiments: RGB and RGB-DA: training and testing with the original RGB image with and without Data Augmentation (DA). RGB+Depth and RGB+Depth-DA: training and testing with the original RGB image with depth estimation as additional information, with and without DA. RGB+INV-Depth and RGB+INV-Depth-DA: training and testing with the original RGB image with depth estimation as additional information but with inverted gray levels, with and without DA. G and G-DA: training and testing with the Green (G) plane of the image, with and without DA. GR and GR-DA: training and testing with the (Green + Red) (GR) planes of the image, with and without DA. GB and GB-DA: training and testing with the (Green + Blue) (GB) planes of the image, with and without DA. Grayscale and Grayscale-DA: training and testing with the grayscale image, with and without DA. In all the experiments the testing set contained 91 images; the training set without DA contained 275 images, and 550 with DA.

4.2. Experimental Setup

The CNN training and implementation was carried out in a free GPU environment using Google Colaboratory [27] with Tensorflow [23] and Keras (https://keras.io, last accessed 2 September 2021) frameworks. Firstly, we trained the depth estimation architecture; it took approximately 3 h for training and less than a second for testing a single image. The training time for the experiments without data augmentation took approximately 1 h and less than a second to test a single image. The training time for the experiments with data augmentation took approximately 3 h and less than a second to test a single image. The experiments that include depth information took similar time for training and testing, with and without data augmentation respectively (code available at https://github.com/EduardoValdezRdz/Glaucoma-classification, last accessed 2 September 2021).

4.3. Discussion

To evaluate our methodology we used state of the art metrics depicted in the following equations:

$$\mathrm{Accuracy}=\frac{\mathrm{True}\mathrm{Positive}+\mathrm{True}\mathrm{Negative}}{\mathrm{True}\mathrm{Positive}+\mathrm{False}\mathrm{Positive}+\mathrm{False}\mathrm{Negative}+\mathrm{True}\mathrm{Negative}}$$

(11)

$$\mathrm{Precision}=\frac{\mathrm{True}\mathrm{Positive}}{\mathrm{True}\mathrm{Positive}+\mathrm{False}\mathrm{Positive}}$$

(12)

$$\mathrm{Recall}=\frac{\mathrm{True}\mathrm{Positive}}{\mathrm{True}\mathrm{Positive}+\mathrm{False}\mathrm{Negative}}$$

(13)

$$\mathrm{F}1=2·\frac{\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$

(14)

Table 2. Quantitative results (higher is better).

Experiment	Precision	Recall	Accuracy	F1
RGB	0.7000	0.7777	0.8352	0.7368
RGB-DA	0.9230	0.8888	0.9450	0.9056
RGB+Depth	0.7500	0.7058	0.8351	0.6666
RGB+Depth-DA	0.9565	0.8148	0.9340	0.8800
RGB+INV-Depth	0.7500	0.7058	0.8351	0.6666
RGB+INV-Depth-DA	0.9565	0.8148	0.9340	0.8800
G	0.7727	0.6296	0.8352	0.6939
G-DA	0.8800	0.8148	0.9120	0.8561
GR	0.6111	0.8148	0.7912	0.6984
GR-DA	0.9565	0.8148	0.9340	0.8800
GB	0.9375	0.5555	0.8571	0.6977
GB-DA	0.9583	0.8518	0.9450	0.9019
Grayscale	0.7895	0.5555	0.8241	0.6522
Grayscale-DA	1	0.7777	0.9340	0.8750

shows results for all performed experiments. We can see that, in general, data augmentation was helpful and led into a better classification. RGB+Depth, RGB+INV-Depth, G, GR led into a similar classification and we can say that these experiments could be discarded and that depth information was not useful to classify glaucoma. We obtained the best results using the original RGB image and the combination of the Green and Blue planes, both with the augmented dataset. Using the grayscale image led to a good classification of healthy cases.

Table 3. Comparison between our best results vs. the state of the art (sorted by precision).

Paper	Year	Accuracy	Precision	Recall	Model	Images
Dos Santos Ferreira et al. [8]	2018	1.0000	1.0000	1.0000	U-Net	101
Grayscale-DA (ours)	2021	0.9340	1.0000	0.7777	AlexNet	724
Raghavendra et al. [7]	2018	0.9813	0.9830	0.9800	18-layer CNN	1426
Liu et al. [17]	2019	0.9960	0.9770	0.9620	ResNet	274,413
GB-DA (ours)	2021	0.9450	0.9583	0.8518	AlexNet	724
Li et al. [3]	2018	0.9200	0.9560	0.9234	Inception-v3	70,000
Chistopher et al. [11]	2018	0.9700	0.9300	0.9200	ResNet50	4363
Chai et al. [14]	2018	0.9151	0.9233	0.9090	MB-NN	2554
RGB-DA (ours)	2021	0.9450	0.9230	0.8888	AlexNet	724
Bajwa et al. [15]	2019	0.8740	0.8500	0.7117	7-layer CNN	780
Fu et al. [5]	2018	0.9183	0.8380	0.8380	DENet	1676

shows a relative comparison of results with similar works in the state of the art, ordered by precision; however, it is important to note that this could not be a direct comparison, as most datasets were private, and, although glaucoma-detection oriented, they were not available to conduct tests directly. In terms of content, the state-of-the-art datasets were similar since they all had fundus retinal images, the change between each one of them was the number of images and their resolution. In general, the architectures of the previous works were complex models that required several training stages since in intermediate stages they extract the optic disc and together with the complete image perform their classification, except for Bajwa’s work in which is similar to ours, first he uses a preprocessing to extract an image containing the optic disc and then his architecture classifies images in which only the optic disc is presented, however our architecture showed better results when classifying these images.

Finally, additional to these metrics, we were interested on examining the particular number of false classifications our method found. True positives meant that a healthy optic disc was classified as healthy; the true negatives mean that an optic disc with glaucoma was classified as non-healthy; false positives meant that an optic disc with glaucoma was classified as healthy; finally false negative meant that a healthy optic disc was classified as non-healthy. In this work the worst case was the false positives, because a patient with glaucoma could not be classified as healthy. In Figure 9 we show the values obtained in our experiments. Another important point is that we identified some images that were not classified correctly, because the images in question did not have the appropriate contrast to be classified as depicted in Figure 10, since if they had high or low contrast the optic cup was lost completely in the optic disc.

5. Conclusions and Future Work

In this work, we presented a simple CNN model capable of classifying glaucoma in digital retinal fundus color images, under low data conditions. This is achieved by first extracting the optic disc from a full digital color image and other preprocessing methods that we can apply to this type of image to perform a correct classification. Although the best results in terms of accuracy were obtained using the original RGB image, the combination of the Green and Blue planes also showed good results, due to the contrast of optic discs that provide both images. Grayscale images allowed us to obtain a precision of 100%, although with a corresponding decrease in recall. We found that adding depth information was not helpful in the detection of glaucoma. The novelty of our work is based on the comparison of different combinations of planes that can be obtained from an RGB image, and although we show that the best results are obtained using the original image, the green plane and its transformation to grayscale, we conclude that we can use a simple architecture and still be able to adequately classify glaucoma. On the other hand, we identified the type of images that can affect the performance of the classifiers. As future work, we propose a further exploration of preprocessing methods to increase contrast and find the extent to which the classification relies on. Although our method was successful with a small number of images, as future work we plan to test the influence of extending the cases to be tested in this task.