Page 19 - Computational Retinal Image Analysis

8 CHAPTER 2 Clinical motivation

(3) facilities for diagnosis and treatment should be available,
(4) there should be a recognizable latent or early symptomatic stage,
(5) there should be a suitable test or examination,
(6) the test should be acceptable to the population,
(7) the natural history of the condition, including development from latent to
declared disease, should be adequately understood,
(8) there should be an agreed policy on whom to treat as patients, and
(9) the cost of case-finding should be economically balanced in relation to
possible expenditure on medical care.
These criteria are all met in diabetic retinopathy, and it has been confirmed that eye
screening in diabetes is cost-effective [8]. Likewise, the introduction of the national
UK diabetic retinopathy screening program led to data demonstrating that DR for the
first time in half a century is no longer the leading cause of blindness in the working-
age population [9]. Understanding the above conditions for successful screening is
quite important when automated image grading system is to be introduced to the real
world setting. Because even when high sensitivity and specificity is achieved with
RIA technically, some diseases are not suitable for screening by their nature.
Second, classification and definitions of retinal signs of diabetic retinopathy
have been clearly standardized through clinical trials and epidemiological studies.
The classification of diabetic retinopathy has been standardized according to the
Early Treatment Diabetic Retinopathy Study (ETDRS) scale based on the modified
Airlie House classification [10]. The scale has stood the test of time, and it has now
been commonly accepted as the gold standard classification. However, even though
the ETDRS scale is well suited for research settings, its practical use is limited by
the high complexity of the scale with a high number of specific and sophisticated
steps. A potential solution for this was presented by the American Academy of
Ophthalmology in 2003, when Wilkinson et al. proposed the International Clinical
Diabetic Retinopathy Disease Severity (IC-DR-DS) scale [11]. This was a five step
scale with two major advantages. Firstly, the numbers of steps are limited and easy
to learn. Secondly, the scale is well-suited to stratify patients according to risk of
diabetic retinopathy progression. Specifically, patients with severe non-proliferative
diabetic retinopathy (NPDR) are important to identify given a high risk of progres-
sion to PDR, which approximates 50% and 71% in 1 and 3 years, respectively. In
most countries, this level of diabetic retinopathy is considered as “referable diabetic
retinopathy” in screening meaning that patients with this level of diabetic retinopathy
should be referred to ophthalmologists for further management. Hence, classification
according to the IC-DR-DS scale would be important in health care programs that
rely on DR-screening with flexible and individualized time intervals.
Even though the introduction of DR-screening has been successful in many coun-
tries, other issues have arisen. The repetitive screening of DR is costly and strenuous,
and specialized health care providers are burdened by the task. These issues may all
be addressed by automated retinal image analysis (ARIA) based on algorithms with
the capability to detect diabetic retinal lesions. However, such programs often intro-
duce a generalization problem, when the algorithms are set out to classify images not