fidoptKiTS23
A preoperative contrast-enhanced abdominal CT of a patient operated on or ablated for a suspected kidney cancer in the M Health Fairview system, Minnesota, with the kidneys, tumours and cysts segmented (KiTS23 training set).
28 test casesabdomenimage CTresearch only
Read before comparing (8)
- Pool. KiTS23's 489 training cases are patients who had a cryoablation or a partial or radical nephrectomy for a suspected renal malignancy from 2010 to 2022, each the most recent contrast-enhanced preoperative CT, in the late arterial (corticomedullary) or nephrogenic phase; the challenge gives no phase a case. The pool is cases 00210 to 00299 (KiTS19's test set, its labels public since KiTS21) and 00400 to 00588 (new in KiTS23), 279 patients; no two share sex, age and BMI in kits23.json.
- Truth. On the cases new in KiTS23 each region was drawn once by a trainee; on cases 00210 to 00299 (9 of the 28 test CTs), annotated for KiTS21, three laypeople drew each region from a trainee's guidance pins and the label is their majority vote. The postprocessing removes perinephric fat by a Hounsfield threshold, and an nnU-Net's cross-validation flagged cases for expert review. Graded semi-automatic, as KiTSCT. KiTS21 found that three drawings a region barely helped methods, so KiTS23 dropped them; no agreement between readers is published for these cases.
- Labels. kidneys excludes the tumours and cysts, as in KiTSCT; kidney_cyst is TotalSegmentator's kidney_cyst_left with kidney_cyst_right, as in RCCAID. KiTS ranks hierarchical classes (kidney with masses, masses, tumour); the benchmark scores each label apart, so its numbers are not KiTS23's.
- Cases. The test tumours hold 1.4 to 2,591 mL (median 29 mL; case_00430's fills much of the abdomen); 15 of the 28 test CTs hold cysts. Slices 0.5 to 5 mm apart, pixels 0.52 to 0.98 mm; the scanners' padding at -1024 or -2048 HU stays.
- Images. The repository's NIfTI files byte for byte, in Hounsfield units. Their headers put the slice axis first and the origin at 0: positions are not the scanner's, but the orientation is anatomical (liver on the patient's right on the cases viewed), and no method needs more.
- Start. A mask of the whole image scores a Dice at the median of the test CTs of 0.012 for the kidneys and 0.001 for the tumours; an Otsu mask 0.024 and 0.003. TotalSegmentator 2.18.0 (total) on test case_00275 scores 0.837 for the kidneys (629 mL against 559 mL: its kidneys keep part of the tumour), 0.261 for the cysts, and has no tumour class.
- Overlap. KiTSCT, SAROSBodyCT and AorticTreeCTA hold C4KC-KiTS patients, cases 00000 to 00209 here, all left out. RCCAID's TCGA kidneys and CPTACCCRCC_Seg's CPTAC patients come from other centres. No other dataset of the corpus names KiTS21 or KiTS23.
- Outside numbers. None on these cases: KiTS23 ranks on its hidden test set, where its winner reached a Dice of 0.835 averaged over the three hierarchical classes and 0.758 for the tumour (the challenge's results; Myronenko et al. 2023).
No method has a result on every case of this dataset yet.
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
- 28 test CTs drawn with seed 2026 among the 254 cases outside development, KiTS19's former test cases and the cases new in KiTS23 in proportion (9 and 19), and 25 development CTs. Cases 00000 to 00209, TCIA's C4KC-KiTS, are left out of the pool: KiTSCT holds them. KiTS23's 110 test CTs are not public.
- Truth
- contours by a program, verified by people KiTS23's three classes, 1 the kidneys (parenchyma and the non-adipose tissue of the hilum, without tumours and cysts), 2 the tumours suspected malignant, 3 the cysts: 3D boxes and 2D contours drawn by trained annotators (residents and students, experts consulted in doubt) on the public ULabel platform, rasterised by a script of Hounsfield thresholds and geometric rules, reviewed by trainees, and the cases where a cross-validated nnU-Net disagreed reviewed by experts; each region drawn once on the cases new in KiTS23, and on cases 00210 to 00299 (KiTS21's process) drawn by three laypeople from a trainee's guidance pins, the label their majority vote
- Source
- kits-challenge.org/kits23
github.com/neheller/kits23
huggingface.co/datasets/neheller/KiTS-Challenge-Imaging
arxiv.org/abs/2307.01984
arxiv.org/abs/2310.04110 - Licence
- CC BY-NC-SA 4.0
25 cases for tuning, never tested
case_00437, case_00217, case_00536, case_00438, case_00232, case_00567, case_00501, case_00251, case_00433, case_00521, case_00228, case_00543, case_00469, case_00264, case_00556, case_00554, case_00517, case_00256, case_00514, case_00465, case_00214, case_00574, case_00570, case_00238, case_00426