Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Офтальмология / Английские материалы / Computational Analysis of the Human Eye with Applications_Dua, Acharya, Ng_2011.pdf
Скачиваний:
0
Добавлен:
28.03.2026
Размер:
20.45 Mб
Скачать

Chapter 2

Computational Methods for Feature Detection in Optical Images

Michael Dessauer and Sumeet Dua

2.1.Introduction to Computational Methods for Feature Detection

Digital imaging provides medical researchers and computer scientists the opportunity to collaborate in solving medical diagnostic challenges with automated computational methods. The medical researchers in the pathological domain of interest are necessary to define clearly where, how, and why a subject is symptomatic of a specific disease or condition. For retinal images, these domain experts play a significant role in creating automated decision support by grading the severity of pathologies such as diabetic retinopathy and macular degeneration. Computational scientists, leveraging domain expert knowledge, work to solve the pattern recognition task of differentiating retinal images into specific classes of disease severity to provide clinical decision support.

The creation of an automated algorithm that takes retinal fundus images as input and classifies the image into levels of disease severity is a multistaged process, in which each step provides input into the next. Although retinal-imaging hardware preprocesses the acquired image, more low-level processing steps are necessary to condition an image into a form acceptable for higher-level algorithms. These preprocessing steps include image

Department of Computer Science, College of Engineering and Science Louisiana Tech University, Ruston, LA, U.S.A.

39

Michael Dessauer and Sumeet Dua

brightness normalization, non-uniformity correction, noise reduction, and noise artifacts elimination. In longitudinal studies, image registration is an important preprocessing step that allows the retina axis to remain oriented for comparison. After a retinal image has been properly preprocessed, the algorithm typically requires specific anatomical constituents to be identified to a level of specificity somewhere between finding its general area (detection), its central location (localization), and its boundary (segmentation). These retinal regions include the blood vessels, the optic disk, or pathological phenomena such as microaneurysms, hemorrhages, and cotton wool spots. After the identification stage is completed, features are extracted from the different regions to provide the discriminate characteristics for accurate pathological classification. In supervised classification schemes, classifiers are trained with features taken from annotated training sets, creating a data-dependent decision model. A priori methods use the algorithm model parameters to make decisions for classification.

Before discussing the computational methods for accomplishing these algorithm steps, we will describe the digital retinal image representation as a discrete function, f(x, y), of pixel size M × N. The RGB color images (typical retinal fundus image format) will give

f(x, y) = {r, g, b}

(2.1)

for each (x, y) location (note in matrix form, x refers to rows and y refers to columns). These intensities compose three two-dimensional (2D) matrices, with ranges for eight-bit images between [0, 255], or, if normalized, between [0, 1]. We can also think of these intensities as magnitudes at a particular location (Fig. 2.1), to visualize many of the computational methods discussed in this chapter more easily. As shown in Fig. 2.1, the green channel (G) will contain most of the interesting contrast in retinal images; thus, it will be used frequently in the following sections as the representation of the image, f .

For clarity, this chapter is divided into three sections: in Sec. 2.2, we describe preprocessing methods; in Sec. 2.3, we explain segmentation/ localization, and in Sec. 2.4, we explore feature extraction techniques. Methods will overlap in several sections. As you will see in the discussions below, computational methods can both enhance and extract descriptive features that provide the necessary specificity for successful retinal pathology classification. The process of selecting which methods to implement can be a

40