fidoptSTARE, retinal vessels drawn by hand
A colour photograph of the retina, 2D, with its blood vessels drawn by hand: one of the twenty images of the STARE project's vessel segmentation set (Hoover, Kouznetsova and Goldbaum 2000), nine normal, eleven diseased.
Read before comparing (5)
- Images. 700 x 605 pixels, in colour as stored, one vector of red, green and blue a pixel. No pixel size is given, so distances are in pixels and the surface tolerance is one pixel. No mask of the field of view is given or made.
- Truth. The page gives two drawings of each image; the truth is Adam Hoover's, the first it lists. The two disagree a good deal: Hoover marks 7.6% of the pixels as vessel (4.3 to 10.2), Valentina Kouznetsova 10.9%, drawing many more of the thin vessels. A method that finds the thin vessels Hoover left out is scored down for them.
- Strictness. On the ten test images Kouznetsova's drawing scores a Dice of 0.75 (0.68 to 0.80) and a surface Dice of 0.70 (0.59 to 0.78) against Hoover's; the truth grown by one pixel scores 0.76 and 0.75; the dark pixels of the green channel (Otsu) a Dice of 0.06. A method near 0.75 is as close to the truth as the second person.
- Strata. The project's diagnosis list calls nine images normal and gives the other eleven a disease: diabetic retinopathy, artery or vein occlusion, emboli, choroidal neovascularisation, drusen, retinitis, vasculitis or hypertensive retinopathy. im0077, 'Normal AND Hypertensive Retinopathy', counts as diseased.
- Linked images. The project names no patient, but some images are linked: im0235 and im0236 show one eye; im0239 and im0240, and im0081 and im0082, look like one person's two eyes; im0162 and im0163 may be one eye in another field. Their test members' partners are kept out of development. im0001 to im0005 may also come from one person.
| Accuracy | Cost | |||||||
|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Surface Dice | HD95 px | Time s | RAM GB | GPU GB | Images |
| 1 | Little W-NetCPUapp | 0.786 ±0.027 | 0.795 ±0.041 | 12.44 ±4.69 | 3 | 0.5 | ||
The order averages each method's ranks on Dice, Surface Dice, HD95 px, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 cases. Two neighbours are tied when a paired Wilcoxon test on the shared cases does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.
Dice: how much the organ contours overlap, 1 is perfect. Surface Dice: the share of the contours' surfaces within a tolerance of each other, 1 is perfect. HD95: distance between the contours' surfaces.
Case by case
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 10 test images drawn with seed 2026 among the 20, normal and diseased retinas in proportion, and 3 development images. The partners of four linked test images are kept out of development, and seven images are unused. One image a case.
- Truth
- contours by experts the vessel network of the photograph drawn by hand by the first of the two observers the page names, Adam Hoover, one of the paper's authors; the second observer's drawing (Valentina Kouznetsova) is not used
- Source
- cecas.clemson.edu/~ahoover/stare/probing/index.html
cecas.clemson.edu/~ahoover/stare
cecas.clemson.edu/~ahoover/stare/diagnoses/all-mg-codes.txt
doi.org/10.1109/42.845178 - Licence
- none stated
3 cases for tuning, never tested
im0044, im0003, im0324