Automatic Diagnosis of Glaucoma Using Digital Fundus Images

Fig. 6.2. A typical fundus image (reprinted from Ref. [14], with kind permission from Journal of Medical Systems).

6.2. Materials and Methods

The fundus images (60 fundus, 30 normal, and 30 glaucoma images) were taken from patients at the Kasturba Medical College, Manipal, India. The subjects were between 20 and 70 years old. The images were graded by physicians from the ophthalmology department of the same hospital. The fundus images were stored in as 720 × 560 pixel bitmap images. Figure 6.2 shows a typical fundus image, Fig. 6.3(a) shows a normal fundus image, and Fig. 6.3(b) shows a glaucomatous fundus image. Blood vessels in the glaucomatous fundus image are often observed to be swollen.

6.2.1. c/d Ratio

The optic cup is enlarged and swollen in a subject suffering from glaucoma. An increase in the c/d ratio is one of the most significant symptoms of glaucoma. Hence, the ratio of the optic cup area to the area of the optic disc was taken as a feature. Figure 6.4 shows the block diagram to compute the c/d ratio.

After analyzing the RGB components of the image, we discovered that it is easier to differentiate the optic disc with the red image component. However, both optic disc and cup could not be easily distinguished in this

211

Rajendra Acharya U. et al.

Fig. 6.3. (a) Fundus image of normal eye and (b) fundus image of glaucomatous eye.

Fig. 6.4. Block diagram of the computation of c/d ratio (reprinted from Ref. [14], with kind permission from Journal of Medical Systems).

212

Automatic Diagnosis of Glaucoma Using Digital Fundus Images

color component, because there was no well-defined border between these two. By trail and error, we decided that the green component yields the best results for distinguishing the optic cup from the remainder of the image. The contrast of the resultant images on these two channels was enhanced to facilitate feature extraction.

For accurate measurements of both optic disc and cup areas, the blood vessels were removed from the images. This removal was done through a series of morphological operations, namely, dilation, erosion, as well as opening and closing operations.15 Morphological erosion is used to wear away foreground pixels in specified regions,7 whereas morphological dilation operations are used to enlarge specific regions of foreground pixels. These techniques remove unwanted bright spots or boundaries in a fundus image.

First, a disc-shaped structuring element of size 15 was created, and a close and open morphological operations were performed both on the red and on the green component images. The structure of image is modified using morphological imaging. Dilation and erosion are widely used on the image. The image grows due to dilation and shrinks because of erosion.7 The close operation filled the gaps and smoothened the outer edges, whereas the open operation removed small bright spots present.

Then, the boundary of the optic disc and cup was shown using thresholding, in which the eight-bit red and green images were converted to binary images. The thresholding value used was 0.805. Finally, the images were eroded and dilated to further smoothen outer boundaries, as illustrated in Fig. 6.5(a) and (b). The white pixels in the corresponding regions were counted to get the ratio of the area of the optic cup and disc.

6.2.2. Measuring the Area of Blood Vessels

A subject with glaucoma will have swollen blood vessels. Hence, the area of blood vessels in a fundus image is an important and useful parameter in the diagnosis of glaucoma. Figure 6.6 shows the flowchart to evaluate the area of blood vessels.

The location of the blood vessel was determined using the green component, as the blood vessels within the optic disc are easier to see than the other components. Image intensity value was complemented, and,

213

Rajendra Acharya U. et al.

Fig. 6.5. (a) Optic disc area and (b) optic cup area.

hence, in the output image, the initial dark areas become bright and vice versa.

A disc-shaped structuring element of size 10 was created and followed by various morphological operation techniques to remove unwanted bright spots, fill the gaps, and smoothen the outer edges. By thresholding with threshold value of 0.1, a binary image was obtained. The blood vessel area

214