Page 20 - Computational Retinal Image Analysis

1 Introduction 9

equivalent to those used for training. Consequently, the performances of most ARIA
models have been limited by moderate specificities despite very high sensitivities
[12, 13]. Deep-learning models may provide a solution to this with results approach-
ing human performance with high levels of sensitivity as well as specificity [14, 15].
Now with an advanced boost with deep learning technology, the FDA approved the
first fully automated diagnostic system to screen people with diabetes—if the indi-
vidual has a certain level of diabetic retinopathy they are seen by an ophthalmologist
otherwise they continue annual screening by the system if it is in a milder stage.

1.2 Assessing severity and classifying clinical eye diseases
Ophthalmologists inspect retinal images to capture signs of clinical disease, and
then signs to determine the severity of the disease. Although retinal images provide
both qualitative and quantitative information, ophthalmologists or image graders
mainly focus on the qualitative aspects to determine the severity of the diseases.
Traditionally, qualitative information such as presence, shapes or location of the
signs specific to each eye disease are used as evidence for decision making in a
clinical or screening setting. CAD can enhance this process by adding quantitative
retinal imaging analysis and providing more precise, reproducible, measurable, and
comparable assessment for clinical eye diseases.
Glaucoma is one of the leading blinding eye diseases in developed countries re-
flecting an aging society [16]. Imaging diagnosis and severity assessment of glau-
coma is based on optic disc appearance such as an enlarged cup-to-disc ratio, optic
disc rim thinning, and retinal nerve fiber layer defect. These signs shape the char-
acteristics of glaucomatous retinal changes. However, the very initial pathological
changes of glaucoma develop far before those signs are clinically visible. In fact,
optic disc changes are detectable after ganglion cells are lost at a certain level.
Optical coherence tomography (OCT) can capture the early changes of ganglion
cell layer complex thinning and following thinning of the retinal nerve fiber layer.
Quantification of the specific retinal layer contributes to assessing severity, progres-
sion and treatment response of eyes with glaucoma, and provides invaluable informa-
tion to manage the disease.
Another good example of RIA in clinical eye disease classification and manage-
ment is the assessment of macular edema. OCT also provides a quantitative assess-
ment to the macular edema caused by diabetic retinopathy. Classification of diabetic
macular edema has traditionally been performed according to the clinically signifi-
cant diabetic macular edema (CSME) in the ETDRS [17]. However, the scale was
developed at a time where OCT had not been introduced and at present an evidence-
based update on the scale is needed to account for the different abilities to measure
macular edema by OCT as compared to stereo fundus evaluation. In terms of this,
a considerable inconsistency in detection of diabetic macular edema and CSME
exhibits when fundus photography and OCT were compared [18]. In fact, differences
went both ways, but in general there was a higher prevalence of diabetic macular
edema and CSME by fundus photography (61.4% and 48.5%) as compared to OCT
(21.1% and 21.3%). Thus it is now standard to use OCT for the evaluation of diabetic