Datasets:

le-dang1haha
/

FundusShift

label class label 3 classes	image_id stringlengths 16 16	label_name stringclasses 3 values	source stringclasses 7 values	family stringclasses 7 values	region stringclasses 2 values	country stringclasses 5 values	device stringclasses 7 values	licence stringclasses 4 values	native_label stringclasses 42 values	normal_scope stringclasses 2 values	group_id stringlengths 17 17	orig_sha256 stringlengths 64 64	quality_flags stringclasses 6 values	dr_grade int8 -1 4	orig_width int32 702 5.18k	orig_height int32 706 3.46k	fov_fraction float32 0.12 1	brightness float32 13 229	contrast float32 3.3 106	sharpness float32 1.6 503
2G	000c721982610a31	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g000c721982610a31	21454ac5b4c31beb21eff4443a12255b04400e3b2a287499bef6302c4b8341e0		-1	2,004	1,690	0.662704	99.486099	13.938952	41.705132
2G	0026ce97338dd7e7	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0026ce97338dd7e7	81fe5b9aec288ca60590866d501aa54af137cad0648cd087cba9f66c7a324943		-1	2,004	1,690	0.837978	109.771225	26.566496	62.367756
2G	0067a68ce0f5543e	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0067a68ce0f5543e	174e2e675716d40ff18718d4dcadf208e6cc54d944623b5985f18e6b6c32bceb		-1	2,004	1,690	0.837306	106.966293	19.547321	50.943665
1D	00706eb9f5a64ffa	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g00706eb9f5a64ffa	6e97545f5a8ddb1eaf659e16e7efe06a2da9872bf5299297bd8e74c551430caa		-1	2,004	1,690	0.662685	102.946175	11.9674	44.606979
0N	00770d728d3e906a	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g00770d728d3e906a	eb03d5227d40b52578b2473f284d6f93c65b50a8d9c320c5e4bcc67e32d3637e		-1	2,004	1,690	0.839355	132.404572	18.521046	99.344116
1D	007d179f77abacdf	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g007d179f77abacdf	5ec9ec1dc4558f7626f2ab43ecdf67c2a1e7d476d81b7b8fd4081dd462d8f356		-1	2,004	1,690	0.836383	92.701439	25.509893	60.16074
0N	007ee93dcd4675c5	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g007ee93dcd4675c5	f6f82fa4839486bd04f0858bf2e17543782aae051e5f5307b447f100fb7a73e4		-1	2,004	1,690	0.838157	123.455185	25.541014	85.881485
0N	00aedf528c31bd7b	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g95354c07b3680d06	5ed5abaea9e2c99f9703a15ecdecb41d0973edbac189ab09ee5055942abfe1c1		-1	2,004	1,690	0.835623	120.003929	30.625767	54.546806
2G	00b3c2e92115a8a9	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g00b3c2e92115a8a9	4a8b7354cd8b905337defef3c956d5135750e891ca8038887c6c2f73f0d1534c		-1	2,004	1,690	0.664047	96.423088	14.425532	41.816784
2G	00b7665d1d5bc084	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g00b7665d1d5bc084	907e72b11ad0a2f43c7321d2429a914fe358720d522fff7ccad3182ed855fe2a		-1	2,004	1,690	0.663938	96.371391	17.096872	33.843853
0N	00f682288d39b165	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g685c2828883bea38	78fe22e713654fd117ff313e9ce9aade162f7d0d9489e118f8255cba82b2a504		-1	2,004	1,690	0.836657	128.154831	31.272835	55.852257
0N	0106e6bea6a7382c	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0106e6bea6a7382c	d118c729197d9dbe2be4f16ff433b7cb32ab71c5468d41de714a858545717ba7		-1	2,004	1,690	0.838607	131.129486	29.680752	102.164833
2G	010ac607628802cd	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g010ac607628802cd	12531b545850dc8955b81d5752f06febad2d79e5af1afeb1a27debad1d296678		-1	2,004	1,690	0.838167	108.704292	33.148201	105.512199
1D	011b447278e9ab51	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g011b447278e9ab51	45538c13b890f0930d8b90c77a51441e1107b553bfed8131b42d1236725eeee3		-1	2,004	1,690	0.837395	125.434746	15.71863	80.824333
1D	01248f43d03b0db1	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g01248f43d03b0db1	463998460b8b707feaac3a8270ef31abcc1c78bf2664b9c13feb601088aaad75		-1	2,004	1,690	0.663564	102.369873	12.631478	50.857609
0N	01309702143e5e6e	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g6a28040d9c2e18cd	da54567b2d363315fa2c64d154fccd3cdc45c0b42750114732439c51d80170db		-1	2,004	1,690	0.83762	129.750183	23.263138	114.07869
0N	017aa941098fecbc	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g017aa941098fecbc	ef2d08175145e324a671da122cc9ea80fa850888bde7a9fa887db2d491cb91b0		-1	2,004	1,690	0.839373	124.026619	29.205923	89.635246
1D	0182a0a5b019588c	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0182a0a5b019588c	f36920e799be80a2e191c170df720f061704b1e9a7d370dc37347dc04e5374b9		-1	2,004	1,690	0.838867	133.499725	22.848591	148.269516
1D	01985d43b4e0dd5e	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g01985d43b4e0dd5e	247fd479c18c54868e5edf3757463d6010f624e1b92495d801dceeb09071f855		-1	2,004	1,690	0.838163	120.259567	21.887814	98.541458
1D	019f41b78a295f30	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g019f41b78a295f30	738c0cecb3a0281d44266cb92c9e67b583b63e93b12eb5825ec304939854ddde		-1	2,004	1,690	0.8375	114.234047	22.824318	51.210358
1D	01b663544d0d2831	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g01b663544d0d2831	fc64fa944251b618398524692a7b04e9e52acf2baceacba767025dfc5af0b5a1		-1	2,004	1,690	0.837725	108.7798	27.157665	62.832111
0N	01b86cc5b31a64c2	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g01b86cc5b31a64c2	d6db72aad6072bf83a1583acea632c992d3ee0f7c9a7ad818b75d5027143e068		-1	2,004	1,690	0.663157	109.267815	15.338546	65.443512
1D	01cde9317fd71833	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g01cde9317fd71833	88b69b0581a75d4f0ddf30026c7bca2c051b5f6e16d06899fc14e8b5e05b31ed		-1	2,004	1,690	0.664483	110.19091	14.582842	59.057594
1D	01ec45598f8bd624	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g01ec45598f8bd624	05170cb220a3a603094c7920d384383d03940e84b5c54553f69738464963759e		-1	2,004	1,690	0.663957	94.812424	8.971308	42.379425
0N	01f06ec94a4bd5fe	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g01f06ec94a4bd5fe	4af2bb379ff724a18127a0ebae7fc94be35a3529eacc05fdf156bf5442a68f3d		-1	2,004	1,690	0.838203	118.357971	20.780241	61.600952
1D	01f580e533c73c24	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g01f580e533c73c24	7940637766cd820eaded492e696bbe54800f586562b5f1bb54c2398d434c8461		-1	2,004	1,690	0.838103	110.761429	31.955399	48.060658
1D	0213e5807b887995	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0213e5807b887995	8a493a90f6f503bdaa2a1e22881f79ac6f1f4779739e1660cca718e05e5e6711		-1	2,004	1,690	0.838123	114.465462	24.723019	89.496498
0N	021900aa649da1e0	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g021900aa649da1e0	8fe0afa660412514286e4c6f7809aff269ccbb6fe651859fce7863862445c266		-1	2,004	1,690	0.837475	116.711006	32.20496	48.958923
1D	021de8fe8a59f3d0	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g021de8fe8a59f3d0	4015abc23aaa35e3f36994eb24a42221dc7a94879550d227dd0d027d9fc5b2bf		-1	2,004	1,690	0.66385	94.11068	15.11403	39.959568
2G	0272df3b6603e236	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0272df3b6603e236	812986e5858f403158f928b854f40e75d90cf2cd07bd93b8fb13d3a0e8daf662		-1	2,004	1,690	0.664252	103.925034	13.729239	54.837395
2G	0293d195f17e4371	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0293d195f17e4371	fbf7687262f9247fe9f3f20bc46ab4dd7ccdae8c552018ca49ab85acc89b7d5b		-1	2,004	1,690	0.83801	109.101799	24.815151	43.355877
0N	0295f2979d31f516	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0295f2979d31f516	cc9571372f1692739e20e44f303eaa55aef16b8e380c9a539a5465ef3a0b7cb2		-1	2,004	1,690	0.663389	102.466682	18.63496	49.48642
1D	03150f2f8e992922	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g03150f2f8e992922	1c583e1182cdda0223063faed11c672e8c582f24aa12637333ab6b46896471d2		-1	2,004	1,690	0.837592	105.765839	16.116419	64.341682
0N	031c3fa40de2c8c6	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g031c3fa40de2c8c6	937d9d475174e3eb574fb0a1893a60a213aecd7f55b758dd3aa896878d9b3f7a		-1	2,004	1,690	0.838516	117.589233	19.839237	76.462021
1D	032b787784d2c684	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g032b787784d2c684	bbf80ed2395a09f2b35f2f451f7e947362d3b4730eaef89cb64e22881170a651		-1	2,004	1,690	0.838164	113.720116	23.070253	120.422157
2G	032f4b59a94eba9e	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g032f4b59a94eba9e	6e8fba8434a5479d19066151b13746923dbc8aacce964320e420bfcfff8e0897		-1	2,004	1,690	0.664106	95.96701	13.112287	39.097786
2G	034dc36f707f2e49	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g034dc36f707f2e49	ffe570a3eab0f12cc1cd64561d3e3d385c0320769aa990d75233b278c9625071		-1	2,004	1,690	0.838036	119.782288	22.779039	86.743614
2G	03632f9033f8a1d2	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g03632f9033f8a1d2	94dd814a54604d9952b0b12fce616959010d2e18b68e68e2afbd8d5991d7e6a1		-1	2,004	1,690	0.837916	127.024231	23.708832	71.818008
0N	0367413c02d47344	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0367413c02d47344	8b6977dda508fd2b454659a4a45d7b9d654899f20446d09e6082b3b5e80adde0		-1	2,004	1,690	0.837232	120.134209	24.356506	86.572533
2G	036da3f3cc71fa53	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g036da3f3cc71fa53	ba7e2ec3f204206c94927b354f118628347c3473a8875dcd0c01492ad7403b6c		-1	2,004	1,690	0.838096	121.278809	31.423954	70.443207
0N	03955174543a151d	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g03955174543a151d	1d826794a214f760f3afb5054c19bff7e6b65685418cb478392a7f2da146ff7d		-1	2,004	1,690	0.838482	126.693649	23.33029	115.294731
2G	0395b038d47dbf4c	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0395b038d47dbf4c	0c6f572b1542725a7dbac228aac38fbd5868e7ecc550168e1ad4ebbf8a6fcf55		-1	2,004	1,690	0.839295	142.711014	23.262722	133.535843
2G	03b92aa341988b19	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g03b92aa341988b19	d96c697145988f7f777d00eca45b7faab28a8f1c981f3856f1f4ae0c4918ef74		-1	2,004	1,690	0.838318	111.522163	25.167191	64.072105
2G	03fe496b8a4f4531	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g03fe496b8a4f4531	4382c56da5b9e9f530ecef8a558af5a23042a8e06d37d4b4f29e7b5a192f8d2c		-1	2,004	1,690	0.838312	123.120659	20.727167	90.62822
2G	042110baae4e6ca0	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g042110baae4e6ca0	7cd80109b2e18d874b170aa860aea1d26dfa102dd03f196ae89a91a9fb44edf3		-1	2,004	1,690	0.662875	97.408615	19.963165	69.173737
1D	04326f3f1e85aa40	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g04326f3f1e85aa40	34994fe171053be3eb332961e901a7b4f8490e49d09b8f8b62a09f942f70c7bb		-1	2,004	1,690	0.664294	95.749115	15.2903	41.335648
0N	0448ea3b3ca95682	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0448ea3b3ca95682	b5bb5aa658fbd69699cd4c75514d89747b0ced6ce0fe745147e086a38f62cde2		-1	2,004	1,690	0.838967	121.768471	23.614403	65.37188
2G	0451c9bd992056a5	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0451c9bd992056a5	d3be2075f6f71a31109f2d59ee4360b274c336f8a7c75d826149d006424bebea		-1	2,004	1,690	0.838189	126.139717	30.274935	62.587364
0N	0456d3a8cde03b01	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0456d3a8cde03b01	cac335a92e4bd047d25a002d9cbef9920889a74d4871b84494c2942ce101f4b4		-1	2,004	1,690	0.664292	107.207176	12.109125	53.282928
0N	04598d7d8746d7ef	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g04598d7d8746d7ef	be09da53b5360ec47c6fdf77ef7a4a560bd1e1c59c560cda363686d38bbdb2ef		-1	2,004	1,690	0.663508	96.713982	13.710096	55.351761
1D	045d7f9fe46921c6	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g045d7f9fe46921c6	259d80adc12109132cffe05541cd38d94e7036a492c9edeb372020b46be6eda1		-1	2,004	1,690	0.836928	119.576241	32.376991	77.58358
0N	046907105772d460	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g046907105772d460	eb19e219a059e07d432905c5bcde75fbd1ddd5075cc21f7894ce3df0911ded3f		-1	2,004	1,690	0.837821	121.489456	31.496807	123.834427
1D	046e1da3fa31099b	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g046e1da3fa31099b	78cab409dba549cb8f6a9846a1f64cba2e462c6a1688673b47d677ae66a78f59		-1	2,004	1,690	0.837495	78.734329	20.326866	33.286625
2G	047463ccd2a381bc	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g047463ccd2a381bc	4c31e0cddaf96d39207d6d85c6dde683611fc429d7870e94d2cf417f7b1d7bf6		-1	2,004	1,690	0.837983	122.056061	23.589966	97.644844
1D	047fcc960de25971	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g047fcc960de25971	4f0c5ca93c74ae3d1d7376d0a0c188b4ca0885283f3482b52816261f97cbbae6		-1	2,004	1,690	0.664074	96.213493	10.604996	47.779434
1D	0486ab4d1c30d168	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0486ab4d1c30d168	1c68de53bb94608f023cb127848bf556fe01b28f1be0e9982c55b56fea6c5f53		-1	2,004	1,690	0.663669	90.545822	13.433776	40.820873
0N	04b3cba322af3238	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g04b3cba322af3238	8ff505fcf990168cb443ccb5a19b8182d54f1141c86a3bb56a3333874ab1db8c		-1	2,004	1,690	0.839279	133.151077	25.002441	85.388092
2G	04b4ee08c8b79aec	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g04b4ee08c8b79aec	d0889b2f65b0a21014a4bfd9c8e70fdaa50be444787d1e9039c8190878001670		-1	2,004	1,690	0.663864	105.026344	15.811504	51.376579
0N	04c9727d83eb302b	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g04c9727d83eb302b	8af11848206ca2a4f86411321335d0ad3ea33fcb45d5d82a8fecc09467319c94		-1	2,004	1,690	0.83811	119.645767	19.80891	72.969139
2G	04d0dbac8d357578	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g04d0dbac8d357578	8c6e25daac6adc53ecb3854571369790f1efabdeeccc4c64118ca1d875ae0c58		-1	2,004	1,690	0.840015	116.251686	30.631969	57.124714
0N	04d8a9d54fe8fd25	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g021900aa649da1e0	1f64a0c71120ad263c5738ff2e62e3096e56d2beed0f66cf0e542e7979769b49		-1	2,004	1,690	0.836318	109.666367	33.601719	44.473892
1D	04f7aa9ee3e21639	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g04f7aa9ee3e21639	3fc4725f60223614ca50afd0e5d7d23fb5844b1bb448d9a882744f42f3332cbc		-1	2,004	1,690	0.838765	133.473206	23.627996	92.027641
1D	0511a2fe5bdcc735	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0511a2fe5bdcc735	a68cdab322b4c32dcdf80974fc3027c6c772aec5ee921950ce5b96e3711c52d9		-1	2,004	1,690	0.838759	122.171509	19.56391	106.667961
1D	051826635ed1b0ce	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g051826635ed1b0ce	d25a037c25049f3c3c07c3e777a1eb2f35d2f7fb7cb6cb5aede4882b53df6d80		-1	2,004	1,690	0.839102	105.643623	17.934385	55.677933
0N	051bf32b5837f290	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g051bf32b5837f290	c2ce5bcc12bc9caa0a89529b7a090fbc66138c678d15592874dd0deeaac0b7d3		-1	2,004	1,690	0.837193	111.431877	29.709568	74.189919
2G	053c311d9992c571	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g053c311d9992c571	f7b7114a2da3623a78699e348348197a8d8c5b30d835b89bd8a994fcf4c9f5bf		-1	2,004	1,690	0.839342	124.210869	25.455153	75.418449
1D	055fd635f6c9c4a4	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g055fd635f6c9c4a4	906b0396e4df400a4a0fcdf37fa5cbce7839cd62fa0b2378650f597a05e7cbab		-1	2,004	1,690	0.837238	123.863556	25.167898	72.197029
2G	05604c59e65b85be	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g05604c59e65b85be	8a8f40404f2eac8b1c7eb46987f358d2363783aba3e473f35755812c632094e9		-1	2,004	1,690	0.66361	102.591469	12.483061	43.219444
1D	0561caae12b25c06	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0561caae12b25c06	fa880817d6294c0b161bc064a50b7303aa3bcec6b3998513f15787d7b1c1e23f		-1	2,004	1,690	0.664123	110.344574	12.937005	66.868118
2G	056c81dc5228c8e3	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g056c81dc5228c8e3	6da6da3244c98ff76c7948980e52e6381add8b97c7b12cf1fe750244b0502c00	blurry_for_source	-1	2,004	1,690	0.83524	89.121017	44.544308	24.551519
0N	0572249d777e3e8f	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0572249d777e3e8f	79e53849af5588927e65225e502f02c535d707471c11b16cbbc6f18a31c99f3a		-1	2,004	1,690	0.83871	144.128189	31.889505	96.600594
0N	0573283ca2984259	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0573283ca2984259	345bec04ef4b0e3774fd6358a31bd5e281a42706c45877fa4a123c57a52c6020		-1	2,004	1,690	0.665042	103.316437	12.201512	36.324017
1D	059127483e8f4363	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g059127483e8f4363	ad393d97b44b2a7f0ff4032c36c596d7db1f8a436813a06d8fa0d7a4e54d4616		-1	2,004	1,690	0.663374	105.622604	14.845	48.671429
2G	0596adacbe8fee93	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0596adacbe8fee93	7304b94016a9d300bffc8bbc75e1d6e6951aa815998625aed0a970ee9ff2aa27		-1	2,004	1,690	0.83851	123.332504	26.153585	106.737
1D	05a9a2504c9e3d2c	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g05a9a2504c9e3d2c	c5ccd18232a5dd56be4f7b870f4a2abba551ac81ba4e6b97d9bf7ee7e3002e69		-1	2,004	1,690	0.838436	123.880638	20.060862	100.015854
1D	05b153b01ccee39e	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g05b153b01ccee39e	1ab320ebd1360dae609886146f70a1ba1a90522eb89ac8bb70f0917ed35b6823		-1	2,004	1,690	0.837429	125.672333	37.118023	72.12146
0N	05c3e92622b94e8e	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g05c3e92622b94e8e	dc7c7e5a6f5733cd0e95b05ffbc460b7dce6fe40076e29ad7c283a69237cb82f		-1	2,004	1,690	0.837633	132.567932	29.768036	69.446388
1D	05c8b8ff1fff9fd6	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g05c8b8ff1fff9fd6	c4746f3d943d9708f9a93b562523ac4b5b51bef07477f822b2f83335fe1c4db2		-1	2,004	1,690	0.836472	67.11145	14.238746	29.385721
2G	06097f739bb3aaaa	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g06097f739bb3aaaa	25fb87878c918c38fc7b161a1785b3a31d2effc65a763e639b811c8b8a12f87e		-1	2,004	1,690	0.837525	125.856514	30.617916	56.898323
2G	063a9b4a2fcf775f	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g063a9b4a2fcf775f	166a711871043d0e45342a7856be6803a05575aba4eb07bb797f203bb4b59142		-1	2,004	1,690	0.839653	151.508301	25.372097	96.775887
1D	065a5abda69cf5b0	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g065a5abda69cf5b0	5ee19d06b62560c3aa1d48d8c485280f5ea78a0a69bdc53e2a868525cb587c20		-1	2,004	1,690	0.839358	147.569122	20.164736	184.118134
2G	065e10d4898b8c93	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g065e10d4898b8c93	1df060b0eb7497eb1a6494ff761ccda4f2e617d3587135c40f98d7ed22150471		-1	2,004	1,690	0.6635	105.34864	17.991444	63.148064
1D	0667c9499f7bcc34	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0667c9499f7bcc34	70df52ad1d04b879f1c36f9c8b8a09f52b17616a333b347a7b6a2c3690d978bd		-1	2,004	1,690	0.664941	123.347374	16.635956	70.408714
2G	068edb64999bfe02	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g068edb64999bfe02	4385db57c18646e78bd9f09d31f285df6f423dc089ff3977e72d0950efba04b5		-1	2,004	1,690	0.663372	95.710587	17.822456	50.642681
0N	0693dc0bb04e3132	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g0693dc0bb04e3132	49afffe3c60822e5bff231eee1a9a32f13cef4630a2cddc7aa0423f230b88d85		-1	2,004	1,690	0.664311	114.130692	20.073551	60.997696
1D	069e263679b630df	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g069e263679b630df	9144177eb55d68af8fe927d44e6fec0764c92282ef08a3c0601b13f1a4fbd7fd		-1	2,004	1,690	0.837505	87.697769	16.666107	31.507084
1D	06a6fedcafc7ed46	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g06a6fedcafc7ed46	1c600c667569ddd4a84b77d681d5d21acd7f8ca4f426570fb5494f641e33298e		-1	2,004	1,690	0.661316	94.203888	12.987561	34.315678
1D	06b2f7652b9ee941	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g06b2f7652b9ee941	bf3d95590445867d533d05e4734db3bae5aaa345d9757cc07df0dae8757372c9		-1	2,004	1,690	0.663627	105.121788	15.755864	64.985893
1D	06bfbc6c765891c4	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g06bfbc6c765891c4	693fcfe08954deabb1779f82a2e3b40bbdf22da089078db4e77ca99aefd78726		-1	2,004	1,690	0.662998	102.820595	21.327602	57.270706
0N	06c880480b49c1f0	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g06c880480b49c1f0	47a4a37855c7f602ee9e90fd09d1f4238b02b1d72abe7d8478d3dfa40a11c30e		-1	2,004	1,690	0.663641	113.388351	9.709654	60.516121
2G	06df83ddcb25df29	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g06df83ddcb25df29	7dbebc348098e7c9795a9de007c20dfc1e5798eff5bedf704bd18dc35ea3efc5		-1	2,004	1,690	0.662794	99.443886	24.83952	42.728447
0N	06e40938273a85c6	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g06e40938273a85c6	1a0fd3ee5d63378f7a79b9910f7f08ed947810f4784d802e26d52ca966f92e50		-1	2,004	1,690	0.837513	119.273232	24.320606	82.621918
0N	06eb425c284863c9	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	ga8a55f6803c905fe	03b2ada63e914940743748f037a19856eb2b7e2404a46903d63c4b4e85093893		-1	2,004	1,690	0.836465	108.162521	24.81954	50.626812
0N	073cc7e7bdf4b76c	N	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Healthy	full	g073cc7e7bdf4b76c	44dc423f9ecbd8cafa5ca7db1cacdbef762b3c6409c950b5ef355ceb2164e0ff		-1	2,004	1,690	0.663164	105.866745	9.892383	45.932339
2G	074d49cad18be9e3	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g074d49cad18be9e3	0444afba2a05bbfaf08af66df61340a1125f0c4115690e679fd73c1e3e3acd9b		-1	2,004	1,690	0.839011	128.695511	23.453127	104.487572
1D	0752c8e24a7a6300	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g0752c8e24a7a6300	9cd4609510a1a41bc21d09302eb2c66a62836e7ac8704f631b83bd488351314e		-1	2,004	1,690	0.838329	92.116051	18.896826	54.588985
1D	075505ab0129fda9	D	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Diabetic_Retinopathy	full	g075505ab0129fda9	2b58e2b23fa5add179a232a32de961f594d35466b45eb8e9737154810e27ef09		-1	2,004	1,690	0.663871	96.195107	12.845364	40.441212
2G	079176ec65d30be9	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g079176ec65d30be9	5d30bfba64aa71ac57c41cf5a29b3782b00c2aaf7f3545c679754b9683aff791		-1	2,004	1,690	0.663999	97.967407	9.603223	31.92852
2G	0798ac1d2475a681	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g0798ac1d2475a681	d029e204f75ae30da49a844c8eb4778db113c54e3fd5cfa2806a003eafe86f3d		-1	2,004	1,690	0.664037	104.961693	12.970076	41.632759
2G	07c1866a5a58998b	G	bangladesh_eye	FARIDPUR_BANGLADESH	South Asia	Bangladesh	not reported	CC BY 4.0	Glaucoma	full	g8c93c39bd13a5381	d4f15900db4c8a9e4f21fccfbc59976d71d15bb76ff507ab934b41cbf764e7b6		-1	2,004	1,690	0.837462	119.488655	27.151337	102.295883

End of preview. Expand in Data Studio

FundusShift

A source-held-out benchmark for colour fundus screening under acquisition shift.

FundusShift harmonises 15 openly licensed colour fundus photograph collections into one screening task and evaluates models on whole collections that never appear in development. It keeps the native annotation and the collection identity next to the harmonised label, removes duplicates and patient overlap across splits, and ships disjoint calibration halves for temperature scaling and referral thresholds.

35,280 images are released (inventory: 40,414 images; every excluded image is listed with its reason in metadata/splits_final.csv).

Paper and code

When Fundus Classifiers Meet a New Source: The FundusShift Benchmark and a Low-Capacity Baseline (D. T. M. Le, MBZUAI). Code, saved results and supplement: https://github.com/dangaura-tech/fundusshift (archive, all versions: https://doi.org/10.5281/zenodo.23911966). metadata/splits_final.csv is the exact split used in the paper.

Tasks

task	label	train on	evaluate on
1. Screening (main)	`label` ∈ {N, D, G}	train (+ validation for model selection)	`test`, `test_ood`, per collection and per region
2. Open-set referral	out-of-label-space conditions (`test_openset`, label = -1; condition in `native_label`)	same model as task 1	AUROC / referral rate for AMD, pathologic myopia and CSR images from held-out collections vs `test`
3. DR grading	`dr_grade` ∈ {0..4} (ICDR; -1 if not graded)	dev images with `dr_grade >= 0`	`test` images with `dr_grade >= 0` (Paraguay, mapped from its 7-level scale)
4. Calibration & selective prediction	task-1 labels	temperature on `cal_A`, threshold on `cal_B`	ECE, AURC, coverage/risk on `test` / `test_ood`

Task 1 labels

One label per photograph:

id	label	meaning (describes the photograph, not a confirmed diagnosis)
0	N	apparently normal fundus: the source says normal/healthy, or DR grade 0 in a DR-graded collection (`normal_scope = dr_only`)
1	D	diabetic-retinopathy-like findings (ICDR 1-4, Davis SDR-PDR, or a DR class)
2	G	glaucomatous optic neuropathy / glaucoma (in Chákṣu: 5-expert majority "glaucoma suspect")

A native label is mapped only when it supports the category. Labels that only say "not X" (GON-, PAPILA suspect) are excluded rather than turned into N. Images with findings from two classes, or with conditions outside N/D/G (myopia, RVO, macular scar, ...), are not used.

test_openset holds photographs from held-out collections whose condition is outside the label space (label = -1): AMD (FIVES; AFIO geographic and neovascular AMD), pathologic myopia (PALM) and central serous chorioretinopathy (AFIO). A safe screening model should refer them instead of forcing N/D/G.

Why AMD and pathologic myopia are not training classes. Each class needs at least two independent, cleanly labelled development collections, and the openly licensed pool does not provide them. AMD has 21 AFIO images and an unusable Rajbari set (originals mixed with augmentations). Myopia has only the Faridpur folder, which shares 143 byte-identical files with its Glaucoma folder. They are therefore evaluated as open-set conditions.

Splits

split	N	D	G	OPENSET	total
train	12087	10323	1395	0	23805
validation	1521	1288	174	0	2983
calibration	1510	1285	173	0	2968
test	1477	956	887	0	3320
test_ood	1156	0	187	0	1343
test_openset	0	0	0	861	861
all	17751	13852	2816	861	35280

train / validation / calibration: development collections, split by leakage group (same picture, near copy, or same patient), stratified by (collection, class), 80 / 10 / 10 by images.
calibration carries cal_subset = cal_A (fit temperature) / cal_B (choose acceptance or referral threshold).
test: every core image of 7 held-out collections (FIVES, PAPILA, HYGD, Paraguay, HRF, PALM, AFIO). Each has a different recorded acquisition family from development, but they are not all new countries: FIVES and PALM are from China, as are several development collections. Some FIVES glaucoma images reuse GRAPE photographs; those copies are removed from development. Devices are not reported for every collection, so full camera separation is not claimed.
test_ood: Chákṣu, photographs from three low-cost / handheld devices (Remidio Fundus-on-Phone, Forus 3nethra, Bosch). It is from India, as is IDRiD in development. Report it separately from test.
test_openset: see above.

Collections

source	splits	country	device	N	D	G	OPENSET	inventory	licence
ddr	calibration/train/validation	China (147 hospitals, 23 provinces)	42 camera types	6265	6255	0	0	13673	CC BY-NC-SA 4.0
jichi	calibration/train/validation	Japan	Nidek AFC-230	6295	3120	0	0	9939	CC BY 4.0
bangladesh_eye	calibration/train/validation	Bangladesh	not reported	846	1500	1156	0	5335	CC BY 4.0
deepdrid	calibration/train/validation	China	regular fundus cameras (Shanghai screening programme)	914	1086	0	0	2000	CC BY-SA 4.0
sustech_sysu	calibration/train/validation	China	Topcon TRC-50DX	631	588	0	0	1219	CC BY 4.0
grape	calibration/train/validation	China	Topcon TRC-NW8, Canon CR-2	0	0	586	0	631	CC0 1.0
idrid	calibration/train/validation	India	Kowa VX-10 alpha	167	347	0	0	516	CC BY 4.0
paraguay_dr	test	Paraguay	not reported	708	723	0	0	1437	CC BY 4.0
hygd	test	Israel	not reported	0	0	548	0	747	ODC-By 1.0
papila	test	Spain	Topcon TRC-NW400	333	0	87	0	488	GPL-3.0-or-later
hrf	test	Germany / Czech Republic	Canon CR-1	15	15	15	0	45	CC BY 4.0
palm	test/test_openset	China (Zhongshan Ophthalmic Center)	Zeiss Visucam 500, Canon CR-2	161	0	0	637	1200	CC0 1.0
fives	test/test_openset	China	Topcon TRC-NW8	193	200	193	200	800	CC BY 4.0
afio_composite	test/test_openset	Pakistan (Armed Forces Institute of Ophthalmology, Rawalpindi)	Topcon 3D OCT-2000 (fundus camera)	67	18	44	24	179	CC BY 4.0
chaksu	test_ood	India	Remidio Fundus-on-Phone, Forus 3nethra, Bosch handheld	1156	0	187	0	1345	CC BY 4.0

Regions

region	train	validation	calibration	test	test_ood	test_openset
East Asia	20593	2580	2567	747	0	837
Europe	0	0	0	465	0	0
Latin America	0	0	0	1431	0	0
Middle East	0	0	0	548	0	0
South Asia	3212	403	401	129	1343	24

Leave-one-region-out evaluation is possible with the region column. East Asia is the only region on both sides, through different collections.

Class support is uneven across collections, and several contribute one class only. Report results per collection and class, and report the worst collection, not only pooled accuracy. Pooling gives the largest collection (Jichi) most of the weight and can hide failure on small collection-class groups.

Data engineering

Licence screen. Only collections whose licence allows redistribution of modified images are included (CC BY 4.0, CC BY-SA 4.0, CC0, ODC-By, GPL-3.0). EyePACS, APTOS, Messidor-2, ODIR, BRSET/mBRSET, REFUGE, ADAM, LAG, AIROGS and others are left out because their terms forbid redistribution.
Native → harmonised labels via code/configs/fundusshift_open.yaml, keeping every native token (native_label) and the ICDR grade where available (dr_grade, -1 otherwise).
Exact duplicates: SHA-256 of the original file.
Near duplicates. Generic hashes and DINOv2 embeddings do not separate a re-encoded copy from a different eye on fundus photographs (measured: pHash grouped 9,118 unrelated images; DINOv2 copy cosine overlapped with different-eye cosine). FundusShift therefore uses a vessel-pattern detector: band-pass green-channel signature (32×32) for candidates, then shift-tolerant normalised cross-correlation at 256 px for verification. On 1,500 real images with synthetic copies (rescale 0.3–0.9×, JPEG q60–95, ≤1 % crop, gamma 0.8–1.25), 98 % of copies are recovered, while visually confirmed different eyes reach at most NCC 0.66. Threshold: 0.70. Besides signature candidates, every image is also verified against its 3 most similar images from every other collection. All scored candidates (including the rejected 0.60–0.70 pairs) are in metadata/dup_candidates.csv.
Leakage groups = connected components of {same file, verified near copy, same patient}. A group never spans two splits. Any group that touches a held-out collection is removed from development.
Label conflicts between copies of the same picture remove all copies.
Collection-level exclusions after inspection.
- RFMiD 2.0: each class comes from a different camera (DR 1504×1000, normal 2048×1536, optic-disc cupping 512×512), so trivial image statistics predict the class with AUC 1.00. The collection is kept in the inventory only, so copies of its images are still detected, and none of its images are released.
- Kim's Eye Hospital glaucoma set: its images are 240 px optic-nerve-head crops, not full-field photographs.
- ACRIMA: optic-disc crops (per its readme).
- Rajbari Eye Clinic (Bangladesh): each class folder holds exactly 500 randomly named files, so the 1,222 originals are mixed with augmented copies and cannot be separated.
- AFIO hypertensive retinopathy set: no per-image diagnosis released.
- AIROGS (CC BY-NC-ND) cannot be redistributed in a modified (cropped/resized) form. JSIEC-1000: licence unspecified, and its images are drawn from several sources.
- Aggregated re-packagings (AMDNet23, SMDG, Kaggle "eye diseases", the 22k-image Mendeley merge) are never used, because they re-distribute images of other collections.
Audit, enforced on the final split file:

check	result	violations
leakage group in >1 split	PASS	0
identical file (SHA-256) in >1 split	PASS	0
patient in >1 split	PASS	0
near-duplicate pair across splits	PASS	0
acquisition family on both dev and test side	PASS	0
source on both dev and test side	PASS	0
cal image without cal_A/cal_B	PASS	0

Cross-collection reuse found

a	b	pairs	exact	label_disagree
fives	grape	29	0	0

FIVES re-uses GRAPE photographs in its glaucoma class (brightness-altered copies, confirmed visually). All images of the GRAPE patients involved are removed from development, so FIVES stays a clean held-out collection. Its glaucoma class still comes from the same hospital network as part of GRAPE, so treat FIVES-G as less independent than the other test cells.

Images not used

reason	images
label:excluded_native_label	1820
label:unmapped_condition	1453
duplicate_label_conflict	956
source_excluded_from_core	860
duplicate_of_test_side_image	45

label: reasons come from the label mapping. duplicate_label_conflict is dominated by the Bangladesh collection, where byte-identical files sit in different disease folders (for example Glaucoma and Healthy). Its remaining labels should be considered noisier than those of the other collections.

Shortcut analysis

A random forest on 10 trivial image statistics (resolution, aspect ratio, field-of-view fraction, brightness, contrast, sharpness, mean RGB) identifies the collection with balanced accuracy 0.998 (chance 0.067, grouped 5-fold CV). Within a collection, the same statistics predict the class with the AUCs below. AUC far above 0.5 means the classes were photographed under different conditions, and in-collection accuracy overstates disease recognition:

source	class_vs_rest	n	auc_trivial_features
fives	N	786	0.991
palm	N	798	0.988
sustech_sysu	N	1219	0.923
sustech_sysu	D	1219	0.923
fives	G	786	0.909
ddr	D	12520	0.908
ddr	N	12520	0.908
bangladesh_eye	D	3502	0.864
fives	D	786	0.856
paraguay_dr	D	1431	0.845
paraguay_dr	N	1431	0.845
deepdrid	N	2000	0.841
deepdrid	D	2000	0.841
bangladesh_eye	N	3502	0.825
idrid	N	514	0.77
idrid	D	514	0.77
bangladesh_eye	G	3502	0.77
papila	G	420	0.715
papila	N	420	0.715
jichi	D	9415	0.706
jichi	N	9415	0.706
afio_composite	N	153	0.624
hrf	D	45	0.622
afio_composite	G	153	0.606
hrf	N	45	0.591
chaksu	N	1343	0.587
chaksu	G	1343	0.587
afio_composite	D	153	0.578
hrf	G	45	0.569

Full EDA: metadata/eda/eda_report.md (composition, acquisition statistics, sample grid, quality flags).

Reference result: frozen-feature linear probe

Frozen dinov2-small features (CLS + mean patch token, 224 px), logistic regression with (collection, class)-balanced weights, C selected on validation (code/baselines/linear_probe.py). A difficulty check, not a tuned baseline.

split	balanced acc	macro-F1	recall N / D / G	worst collection (bAcc)
validation (in-distribution)	0.795	0.781	0.83 / 0.78 / 0.77	grape (0.759)
calibration	0.816	0.793	0.84 / 0.77 / 0.84	jichi (0.755)
test (held-out collections)	0.677	0.666	0.66 / 0.86 / 0.51	hygd (0.456)
test_ood (Chákṣu, handheld)	0.517	0.380	0.77 / – / 0.26	chaksu (0.517)

test collection	bAcc
hygd	0.456
afio_composite	0.508
papila	0.560
hrf	0.578
fives	0.635
palm	0.770
paraguay_dr	0.772

Open-set AUROC (max-softmax, test_openset vs test): 0.500.

Balanced accuracy drops by 11.9 points from validation to the held-out collections. Single-class collections (e.g. HYGD) report recall of that class.

Task 3 (DR grading, same frozen features): quadratic weighted kappa 0.735 on validation vs 0.704 on held-out paraguay_dr (n = 1431).

Columns

column	description
`image`	fundus cropped to the illuminated field (black border removed), longest side ≤ 1024 px, JPEG q95
`label` / `label_name`	0/1/2 = N/D/G; -1 / `OPENSET` for `test_openset`
`source`, `family`, `country`, `device`, `licence`	collection identity and acquisition origin
`native_label`	original annotation tokens, `\|`-separated
`normal_scope`	`dr_only` when N means "DR grade 0" in a DR-only collection, `full` otherwise
`dr_grade`	ICDR 0–4 where the collection provides it, -1 otherwise (secondary DR-grading task)
`patient_id`, `eye`	when released by the collection
`group_id`	leakage group (never crosses splits)
`cal_subset`	`cal_A` / `cal_B` in calibration
`orig_sha256`, `orig_width`, `orig_height`	provenance of the original file
`fov_fraction`, `brightness`, `contrast`, `sharpness`, `quality_flags`	image-quality statistics and flags (images are kept)

Protocol and baseline

from datasets import load_dataset
ds = load_dataset("le-dang1haha/FundusShift")
train, val, cal, test = ds["train"], ds["validation"], ds["calibration"], ds["test"]

code/baselines/train_baseline.py fine-tunes an ImageNet ResNet-50 with a (collection, class)-balanced sampler and selects the checkpoint by validation balanced accuracy. code/baselines/evaluate.py implements the protocol: temperature on cal_A, max-probability referral threshold on cal_B (target risk 5 %), then balanced accuracy, macro-F1, ECE, AURC, coverage/risk at the threshold, per-collection and worst-collection balanced accuracy on test and test_ood, and AUROC for referring test_openset.

python code/baselines/train_baseline.py --out runs/resnet50
python code/baselines/evaluate.py --preds runs/resnet50

Limitations

Labels come from the collections. Grading scales differ (ICDR, Davis, 7-level Paraguay scale, expert majority), and N is "no DR" in DR-only collections.
Patient identifiers exist only for some collections. Elsewhere, near-duplicate detection is the only protection against same-eye leakage.
Development glaucoma comes mostly from GRAPE (follow-up visits of few patients) and the noisy Bangladesh collection. Glaucoma recognition is therefore the hardest transfer cell.
There is no AMD class in development (no openly licensed source), so AMD is evaluated only as open-set referral.
Pre-trained backbones may have seen the held-out collections. FLAIR's published pretraining list includes FIVES, HRF, PAPILA, Paraguay and Chákṣu, so results on these are not source-held-out for FLAIR; AFIO, HYGD and PALM are not on that list. Other models were checked by name against their published documents only, not image by image, and some (for example LLaVA-OneVision, Qwen2.5-VL) do not disclose their training data. Check before claiming held-out results.
Not a medical device and not for clinical use.

Licences

There is no single licence for the benchmark. Each image keeps the licence of its collection (licence column), and these terms apply to it in this collection too. Licence texts: CC BY-NC-SA 4.0, CC BY-SA 4.0, CC BY 4.0, CC0 1.0, ODC-By 1.0, GPL-3.0.

OIA-DDR is CC BY-NC-SA 4.0, so it may not be used commercially without written permission from its authors. Use of the full benchmark is therefore non-commercial. DeepDRiD is CC BY-SA 4.0 and PAPILA is GPL-3.0-or-later.
Changes made to every image: cropped to the illuminated field, resized to at most 1024 px on the longest side, re-encoded as JPEG q95. Labels were mapped to N/D/G, and the native annotation is kept in native_label.
Using the benchmark means complying with every collection's licence and citing every collection below.
This section summarises published terms and is not legal advice. Anyone planning commercial use should check with the authors of each collection, in particular for models trained on OIA-DDR.

Citations

OIA-DDR — Li et al., Information Sciences 2019. CC BY-NC-SA 4.0.
PALM — Fang et al., Scientific Data 2024. CC0.
AFIO composite fundus/OCT — Hassan et al., Mendeley Data, doi:10.17632/trghs22fpg.4. CC BY 4.0.
IDRiD — Porwal et al., Data 2018. CC BY 4.0.
DeepDRiD — Liu et al., Patterns 2022. CC BY-SA 4.0.
SUSTech-SYSU — Lin et al., Scientific Data 2020. CC BY 4.0.
Jichi — Takahashi et al., PLoS ONE 2017. CC BY 4.0.
RFMiD 2.0 — Panchal et al., Data 2023. CC BY 4.0. (inventory and duplicate checks only; no images released)
GRAPE — Huang et al., Scientific Data 2023. CC0.
Eye Disease Image Dataset (Bangladesh) — Sharmin et al., Mendeley Data 2024, doi:10.17632/s9bfhswzjb.1. CC BY 4.0.
FIVES — Jin et al., Scientific Data 2022. CC BY 4.0.
PAPILA — Kovalyk et al., Scientific Data 2022. GPL-3.0.
HYGD — Hillel Yaffe Glaucoma Dataset, PhysioNet v1.1.0. ODC-By 1.0.
Paraguay DR — Benítez et al., Data in Brief 2021, Zenodo 4891308. CC BY 4.0.
HRF — Budai et al., Int. J. Biomedical Imaging 2013. CC BY 4.0.
Chákṣu — Kumar et al., Scientific Data 2023. CC BY 4.0.

Downloads last month: 100