Page 238 - Computational Retinal Image Analysis

5 Lesion localization 235

4.7 Performance comparison
Table 1 presents a comparison of representative RL and BL detection methods. It
is important to note that the fair performance comparison of the aforementioned
methods is impossible because they were validated using different datasets and
metrics. As such, we only report their performance against specific datasets. It can be
observed that there is no single superior approach across all available methodologies.
From the clinical point of view, finding a DR lesion is more important than
segmenting it, especially for screening, diagnosis, and monitoring purposes.
Thus, the table provides a comparison of detection methods only. The reader
is referred to Chudzik et al. [20, 50] for the in-depth discussion of evaluation
metrics.

5 Lesion localization
Screening programs are common in western countries for the detection of early DR.
The rule-based assessment of intervals must be conservative to reduce the risk of
vision-threatening conditions occurring between examinations. Consequently, the
frequency of screening visits can be overestimated. Superfluous examinations could
be avoided if the frequency could be reduced in patients without progression. A
number of known risk factors for the development and progression of DR are not
included in currently employed rule-based decision models: metabolic regulation of
blood glucose (HbA1c), arterial hypertension, and gender. In recent years, several
models have been presented for individualized determination of the screening
interval in DR [65–67]. The model presented by Mehlsen et al. included the number
of MAs and HMs, HbA1c, blood pressure, gender, diabetes duration, and diabetes
type, and with this information the screening interval could on average be prolonged
with a factor of 2.88 in type 1 diabetic (T1D) patients and with a factor of 1.24 in
type 2 diabetic (T2D) patients. However, since the screening interval defined by
the grader was itself a risk factor it was concluded that the recommendation by
the grader had depended on observations during grading of the FP that were not
described by the number of MA/HM. This was supported by findings that early
progression of DR lesions temporal in the macular area and peripheral from the
larger temporal vascular arcades can predict the development of retinopathy to a
vision-threatening stage [64]. Recently, a study by Ometto et al. [68] investigated
the role of regional distribution of the earliest retinopathy lesions showing in FPs.
This work identified the need for a spatial system that is repeatable to compare the
distribution of lesions showing across photographs from different eyes. A coordinate
system based on the main retinal landmarks, fovea, optic disc, and main vascular
arcades, and suitable for automatic identification was proposed as in reference [63].
The x-axis of a retinal coordinate system was identified by the line connecting the
fovea and optic disc centers, and their distance was used to define one unit of length.
The unit of length in the normal y-axis was defined by the average of the distances