Automated Microaneurysm Detection in Fluorescein Angiograms for Diabetic Retinopathy

Fig. 9.1. Eyes at various stages of DR: (a) eye with mild nonproliferative, (b) eye with moderate nonproliferative, (c) eye with severe nonproliferative, and (d) eye with proliferative (Images obtained from Kasturba Medical Hospital, Manipal, India).

on fluorescein angiograms.2 These methods usually follow a sequence of computational processing steps. The first steps include image preprocessing for noise removal and contrast enhancement for discrimination between MAs and blood vessels. Next, image registration is performed to reduce the errors in alignment and scaling in the images so that the same eye images obtained at different visits are registered. Vasculature processing assists in the separation of vessel cross-section from the MA cross-section. Finally, classification is performed to discriminate the actual MAs from the blood vessels, based on the abnormality and severity of false detections, which is based on extracted features. This chapter will cover a systematic description of these computational techniques.

9.1.1. Preprocessing

In many cases, a set of preprocessing steps is applied to a set of fluorescein angiographic images before the images are analyzed. The motivations for applying these steps is to reduce the unwanted variation in intensity or contrast that may occur due to the over or under exposure of the image and to separate MAs from the vessels, as they have the same intensity characteristics. It is essential to analyze the size, shape, and energy characteristics of the retinal objects in the image before segmenting for MAs. Further, the fluorescein angiograms are acquired during different visits for the same patient. The lapse of time between obtaining the images may result in discrepancies between the contrasts and provide an ineffective comparison. The common techniques in image processing include, but are not limited to, shade correction, Hough transform, and top-hat transform and edge detection. These techniques are applied to the image as a preprocessing step for the intensity

283

Prerna Sethi and Hilary W. Thompson

and contrast enhancement. We briefly discuss these techniques, highlighting the related research in this domain.

9.1.1.1. Shade correction

Shade correction is a related technique that has been used to preprocess retinal images.11−14 Shade correction is an approach that eliminates the background illumination that tends to be present in fundus or angiographic images. Cree et al. in Ref. [14] explains that the true image, F(x, y) is an ideal image when there is a perfect optical system, the fundus is evenly illuminated by the flash, and the point (x, y) is defined as a position in the two-dimensional (2D) plane where all the imaging takes place. However, in practice, the fundus is often unevenly illuminated, or an optical disturbance eases the signal before it reaches the plane. Hence, the intensity of light I(x, y) is defined as a product of the camera function, C(x, y) and the image F(x, y). This equation is represented below:

The background fluorescence due to the capillary network that extends through a large part of the retina and the choroid has an additive effect to the foreground features of the image. Therefore, the image F can be considered a sum of the background fluorescence, Fb and all the other features Ff of the image. For example,

The illumination variation in the images is primarily caused by the following two factors: (i) the camera function, C, and (ii) the background fluorescence, Fb. Substituting Eq. (9.2) into Eq. (9.1), the image I can be defined as:

Spencer et al.15 have described a two-step procedure for preprocessing fluorescein angiograms. First, they normalized the images using a radiometric correction factor for the images to show the same range of pixel gray-levels and the same average gray-level. They then used a shade correction approach to remove the unnecessary changes in intensity that were incurred due to the photographic process and choroidal fluorescence.

284

Automated Microaneurysm Detection in Fluorescein Angiograms for Diabetic Retinopathy

9.1.1.2. Hough transform

Hough transform was proposed to identify straight lines in an image, but, later, the method has been extended to identify curves such as ellipses and circles. The purpose of Hough transform is to detect the parameterized curves in an image by groupings of edge points by object candidates and performing an explicit voting procedure over a set of parameterized image objects.16 The mathematical representation behind a Hough transform is defined as follows. Consider a straight line in a normal form such as,

This equation represents a line passing through the point (x, y) and is perpendicular to the line passing from the coordinates (0, 0) and (ρ, θ), which is (ρ cos θ, ρ sin θ) in rectangular space. To generate a Hough transform, a particular value of θ is chosen and is iterated through the angle defined by the granularity of θ. This process generates a curve in the rectangular representation of Hough transform (Fig. 9.2.).

As a preprocessing step to identify the MAs and the region of interest in the image Bernadres et al.17 first segmented the image using two parallel lines to extract a circular area. They then applied the generalized and the modified Hough transform to detect the lines and the center and the radius of the circle. Azeem et al.18 developed an approach to detect MAs based on circular Hough transform in fluorescein angiographic images. They were able to detect MAs due to the circular nature of the MAs and objects such

Fig. 9.2. Hough transform.

285