Source of Materials

Images for the experiment were selected by a consultant ophthalmologist from the fundus image database of the diabetic retinopathy clinic at Sunderland Eye Infirmary. Retinal photography was performed to all patients according to a standardised protocol as part of their routine clinical care. Mydriasis was induced using tropicamide (1%) eye drops. All digital fundus images were graded by a qualified diabetic retinopathy grader using the developed system of assessment of the EURODIAB study based on the principles of the Modified Airline House classification scheme as described by Aldington and his co-workers.

Image Sets

The high resolution image set (HRIS)

It consists of four images (download images ). They represent different sever grades of diabetic retinopathy. Table I illustrates the formal grading of the images. The use of HRIS supports an evaluation to sub-pixel accuracy. This set consists of four images where 90 segments containing 2368 profiles are manually marked. The images were subsequently down-sampled by a lateral factor of four for submission to the measurement algorithms, so that the vessels widths are known to ±0.25 of a pixel, discounting human error.

TABEL 1: The disease on images in the high resolution Image set.
	Image #	Formal Grading
	1	Level 3: Severe Non-Proliferative Retinopathy
	2	Level 2: Moderate Non-Proliferative Retinopathy
	3	Level 3: Severe Non-Proliferative Retinopathy
	4	Level 1: Minimal Non-Proliferative Retinopathy

The vascular disease image set (VDIS)

It consists of eight digital captured images (download images ). The images were selected randomly from the database of images of patients attending the diabetic retinopathy outpatient clinic at Sunderland Eye Infirmary. Six of the images illustrate different type of diabetic retinopathy, as summarized in Table II. This set consists of eight images where 79 segments containing 2249 profiles are manually marked. This set of images is very noisy and suffers from pathologies, and this manifests in the greater variance of the observers.

TABLE II: The disease on images in the vascular disease Image set.
	Image #	Formal Grading
	1	Arteriosclerotic Arterio-Vascular Changes
	2	Level 5: Proliferative Retinopathy
	3	Level 1: Minimal Non-Proliferative Retinopathy
	4	Level 1: Minimal Non-Proliferative Retinopathy
	5	Level 2: Moderate Non-Proliferative Retinopathy
	6	Normal
	7	Normal
	8	Severe Non-Proliferative Retinopathy

The central light reflex image set (CLRIS)

It consists of two digital images (download images ), containing 21 marked segments with 285 profiles. They were selected to represent early atherosclerotic changes in retinal vessels characterised by the presence of arterial silver-wiring (exaggerated vascular light reflex) secondary to stiffness and changes in the vascular wall.

The kick point image set (KPIS)

It consists of two images (download images ), three segments with 164 gold standard widths. Profiles were identified by choosing profiles directions orthogonal to the centre line direction recovered using the previous and next pixels from the centre line determined by the “tramline”’ algorithm. The three observers then identified the kick-points. By down-sampling the images to a lower resolution, these gold standard widths have higher accuracy than expressed in the measured images so the measurements can be assessed with sub-pixel accuracy. The KPIS subset provides an alternative benchmarking method to the edge-marked subsets.

Retinal Image Computing & Understanding

The high resolution image set (HRIS)

The vascular disease image set (VDIS)

The central light reflex image set (CLRIS)

The kick point image set (KPIS)