iChallenge-PALM19 — PALM: PAthoLogic Myopia

1200 color fundus photographs from 720 myopia-clinic subjects (left eyes only) at Zhongshan Ophthalmic Center, Sun Yat-sen University — the PALM dataset from the ISBI 2019 iChallenge satellite event, with pixel-level annotation of three targets.

• Modality: color fundus photography (2D RGB)
• Organ: retina / eye
• Resolutions: 2124×2056 (Zeiss Visucam 500, n=1047) · 1444×1444 (Canon CR-2, n=153)
• Targets: optic disc · patchy retinal atrophy · retinal detachment
• License: CC0 1.0 (public domain dedication)
• Source: figshare 21299148 (official, author-deposited)

Splits

Patient-disjoint (images from one patient stay in one split). All counts below were re-verified on the archive and match the paper's Table 1 exactly.

Split	Images	PM / non-PM	disc_mask	atrophy_mask	detachment_mask
train (Training)	400	213 / 187	381	311	12
validation (Validation)	400	211 / 189	379	271	6
test (Testing)	400	213 / 187	384	288	6
Total	1200	637 / 563	1144	870	24

Columns

Column	Type	Notes
image_id	string	Stem, e.g. H0001, V0001, T0001 — the pairing key
image	Image (RGB)	Fundus photograph. Original JPEG bytes, stored verbatim (no re-encode)
disc_mask	Image (L), nullable	Optic disc. {0,255}, 255 = disc
atrophy_mask	Image (L), nullable	Patchy retinal atrophy. {0,255}, 255 = lesion
detachment_mask	Image (L), nullable	Retinal detachment. {0,255}, 255 = lesion
has_disc_mask / has_atrophy_mask / has_detachment_mask	bool	Explicit presence flag — see Masks are absent, not empty
pm_label	int32	1 = pathologic myopia, 0 = non-PM
fovea_x, fovea_y	float32	Fovea centre. (0,0) upstream means not visible
fovea_visible	bool	False exactly when the coordinate is (0,0)
device	string	zeiss | canon
shooting_center	string	optic_disc | fovea | midpoint
width, height	int32	Image size (masks always match)
disc_fraction, atrophy_fraction, detachment_fraction	float32, nullable	Foreground area fraction
id_prefix	string	H/N/P (train) or V/T — leaks the class in train, see below

⚠️ Mask polarity was normalised here

Upstream PALM masks are inverted: 0 = foreground, 255 = background. Every mask in this repository has been inverted so that 255 = foreground. Use the ordinary mask > 0.

This was verified rather than assumed: all 2038 source masks decode to exactly the value set {0, 255}, and polarity was confirmed by a corner test — all 8152 frame corners across all masks are 255, none is 0, and the disc's 0-region never touches the frame border in any of 1144 masks.

There is no 128 value in PALM. The 0/128/255 (cup/rim/background) convention belongs to REFUGE, a sibling iChallenge dataset, and must not be applied here.

⚠️ Masks are absent, not empty

56 images have no disc mask, 330 no atrophy, 1176 no detachment. In this 2024 release those files are missing from the archive — there are zero all-background and zero all-foreground masks among all 2038. Absence is stored as a null mask plus a has_*_mask boolean; never infer it from an empty mask.

(The 2019 challenge release instead shipped all-255 empty disc masks for the 56 no-OD images. That trap does not apply to this data.)

⚠️ The three targets overlap spatially

Measured on the 24 detachment cases: 14/24 intersect atrophy (up to 54 % of the detachment area), 5/24 intersect the disc, 3/24 disc∩atrophy. They are therefore three independent binary masks, not one label map. Collapsing them into a single multi-class {0,1,2,3} image would silently drop the overlaps.

⚠️ Filename prefixes leak the class in train

Training stems are H#### (high myopia, non-PM, n=26), N#### (normal, non-PM, n=161) and P#### (PM, n=213) — verified to map 1:1 onto pm_label. validation (V####) and test (T####) are anonymised and carry no leak. Any classification benchmark must read pm_label, never the id. Carried as id_prefix so the leak is explicit rather than hidden in the key.

Ground truth

One tier. Seven ophthalmologists (5–10 yrs) annotated independently and blinded to patient information; masks were merged by majority vote; one senior ophthalmologist (>10 yrs) then quality-checked, discarding faulty individual masks and re-running the vote on the remainder. Fovea coordinates are the mean of the seven, with erroneous ones recomputed by the senior. Disc contours were drawn as fitted ellipses, lesions as free-form closed curves.

The seven individual rater masks are not released — there is no rater or annotation tier to choose between.

Note that ~75 % of PM images are graded "reject" quality by automated IQA (vs 6 % of non-PM). This is intentional — the authors deliberately retained low-quality images to reflect real clinical data — not corruption.

Provenance, naming and cross-dataset overlap

• Provenance: official author-deposited figshare item, named as the data record in the paper. Download md5 28636601413706884ea89ec1be835e94 verified; all counts reconcile with paper Table 1 with zero mismatch.
• Faithful naming: this is the 2024 figshare release, byte-identical to the 2019 challenge files plus the Testing ground truth withheld during the competition — a strict superset of what "PALM19" distributed. Four look-alikes are not this dataset: iChallenge-PM (Baidu/Tianchi — classification only, no masks); ctmedtech/PALM (2019 only, no Testing images, and ships model predictions as if GT); openmedlab PALM19.md (reformatted derivative with fabricated "retinal vessel" labels — PALM has no vessel annotation); IEEE DataPort 10.21227/55pk-8z03 (the 2019 citation, but zero files uploaded).
• Ground-truth tier: single senior-arbitrated majority-vote consensus.
• Cross-dataset overlap: PALM shares institution, grading protocol, cameras and its 400/400/400 structure with REFUGE and ADAM (same iChallenge series), but the cohorts differ (myopia examination vs glaucoma screening vs AMD), PALM is left-eyes-only, and the Canon subsets differ in size (1444² here vs 1634² in REFUGE). No image overlap is documented in either paper and no cross-reference ID exists; the PALM authors explicitly recommend combining PALM with REFUGE and ADAM, implying disjoint cohorts. If PALM and REFUGE ever enter the same benchmark, a perceptual-hash check on the Zeiss 2124×2056 subsets is the one confirmation that cannot be made from the metadata alone. No plausible overlap with DRIVE, IDRiD, PAPILA, RAVIR, Drishti-GS, RIM-ONE, G1020, ORIGA or ODIR-5K (different countries and cohorts).

Citation

@article{fang2024palm,
  title   = {Open Fundus Photograph Dataset with Pathologic Myopia Recognition
             and Anatomical Structure Annotation},
  author  = {Fang, Huihui and Li, Fei and Wu, Junde and Fu, Huazhu and Sun, Xiulan
             and Orlando, Jos{\'e} Ignacio and Bogunovi{\'c}, Hrvoje and
             Zhang, Xiulan and Xu, Yanwu},
  journal = {Scientific Data},
  volume  = {11},
  number  = {1},
  pages   = {99},
  year    = {2024},
  doi     = {10.1038/s41597-024-02911-2}
}

Data citation: Fang H. et al. (2024), figshare, doi 10.6084/m9.figshare.21299148.v1. The challenge page additionally asks for the 2019 record: Fu H. et al., "PALM: PAthoLogic Myopia Challenge", IEEE Dataport, doi 10.21227/55pk-8z03.