Chaksu

Optic disc and optic cup segmentation in colour fundus photographs, for glaucoma assessment. Mirror of the Chákṣu database for the MedOtter benchmark suite.

Modality	Colour fundus photography (2D RGB)
Task	Optic disc + optic cup segmentation (also carries a glaucoma decision label)
Images	1,345 — train 1,009 / test 336
Cameras	Remidio FoP 2448×3264 (1,074) · Bosch 1920×1440 (145) · Forus 3Nethra 2048×1536 (126)
Ground truth	STAPLE fusion of 5 expert annotators
License	CC BY 4.0

Ground truth

Five ophthalmologists annotated every image. The source release ships four fusions of those annotations; the paper designates STAPLE as the gold standard, having the highest agreement with the individual experts:

Fusion	OD Dice	Role here
STAPLE	0.9764	primary → mask, disc_mask, cup_mask
Majority (≥3 of 5)	0.9732	secondary → *_mask_majority
Median	0.9706	secondary → *_mask_median
Mean (all 5 agree)	0.9503	not mirrored (lowest agreement)

mask is a disjoint label map at native image resolution:

value	class
0	background
1	neuroretinal rim (disc minus cup)
2	optic cup

so disc = mask > 0 and cup = mask == 2.

STAPLE masks are stored anti-aliased upstream; they are binarised at >127. In a small number of images the cup exceeded the disc by a handful of edge pixels, so disc_mask is stored as disc ∪ cup to guarantee cup ⊆ disc.

Columns

column	notes
image	original fundus photograph, native resolution, unmodified bytes
mask	disjoint 0/1/2 label map (STAPLE)
disc_mask, cup_mask	binary 0/255 (STAPLE)
disc_mask_majority, cup_mask_majority	binary 0/255 (Majority fusion)
disc_mask_median, cup_mask_median	binary 0/255 (Median fusion)
image_id, camera, split, height, width	identifiers / geometry
glaucoma_decision	majority vote of 5 experts: NORMAL / GLAUCOMA SUSPECT
decision_expert_1..5	per-expert decision (whitespace/spelling normalised)
disc_area, cup_area, rim_area, cup_height, cup_width, disc_height, disc_width	pixel measurements from the source CSVs
acdr, vcdr, hcdr	area / vertical / horizontal cup-to-disc ratio

Scope of this mirror

This is a subset of the full figshare release, restricted to what a segmentation benchmark needs. Present: tier 1.0 (original images), tier 5.0 STAPLE/Majority/Median fusions, tier 6.0 metrics and decisions. Not mirrored: tier 2.0 (expert contour markups), tier 3.0 (per-expert binary masks — note Expert 3 has no Bosch folder), tier 4.0 (OD+OC fusion images), and the Mean and Overlay variants of tier 5.0. Those total ~200 GB of mostly uncompressed TIFF and are available from the original figshare deposit.

Caveats

• Patient overlap across the official split. Only the Bosch subset carries patient IDs, and at least one patient straddles the boundary (P10_Image1 in train, P10_Image2 in test). Forus and Remidio splits are sequential cuts by capture order, not patient-aware, so fellow-eye pairs may also straddle it. The official split is preserved as published.
• Source tier-6 CSVs are partly duplicated across the train and test folders (the Bosch CSV is dataset-wide in both); they are merged by (camera, image_id).
• Expert decision strings upstream include a misspelling (GLAUCOMA SUSUPECT); these are normalised to GLAUCOMA SUSPECT.
• No overlap with REFUGE, DRISHTI-GS, RIM-ONE, ORIGA, PAPILA, IDRiD, DRIVE or RAVIR — Chaksu was newly acquired at KMC/MIT Manipal.

Citation

@article{kumar2023chaksu,
  title   = {Chaksu: A glaucoma specific fundus image database},
  author  = {Kumar, J. R. Harish and Seelamantula, Chandra Sekhar and
             Gagan, J. H. and Kamath, Yogish S. and Kuzhuppilly, Neetha I. R. and
             Vivekanand, U. and Gupta, Preeti and Patil, Shilpa},
  journal = {Scientific Data},
  volume  = {10}, number = {1}, pages = {70}, year = {2023},
  doi     = {10.1038/s41597-023-01943-4}
}

Original data: https://doi.org/10.6084/m9.figshare.20123135 (CC BY 4.0).