fidopt Interactive view
Abdomen CT, segmentation

RCC-AID

Contrast-enhanced abdominal CT of a patient with a clear cell, papillary or chromophobe renal cell carcinoma (TCGA-KIRC, TCGA-KIRP, TCGA-KICH), kidneys, masses and cysts segmented by a model and corrected by students (RCC-AID).

10 test casesabdomenimage CT
Look at the images
Read before comparing (8)
  • Pool. RCC-AID kept 129 contrast-enhanced CTs of 91 patients from TCGA's 1,915 renal CT scans, by metadata and a reader study; 38 of the 129 hold a cyst. A preprint counts 142 CTs of 101 patients in an earlier record; the files used are IDC release 25's 129, read as DICOM label maps.
  • Cases. A case is a patient's earliest study, then the venous, arterial or delayed phase in that order, then the file name, cut to the slices the segmentation covers. The 10 test CTs are 5 venous, 3 arterial and 2 delayed, slices 1 to 5 mm apart, in Hounsfield units with the scanner's padding at -3024 or -3027.
  • Truth. A model's masks corrected by two trained students, a radiologist seeing only the cases they marked uncertain; no agreement between readers is published. It passes the bar of a main leaderboard with these limits. The kidney label stops where the annotated slices stop and is one class, its sides left to the scorer's merge of kidney_left and kidney_right.
  • Scans. They are what the clinics acquired, not always before surgery: development case TCGA-CW-6088 follows a right nephrectomy. The test masses hold 7.4 to 440 ml (median 56); one test case holds a cyst (TCGA-KM-8442, 1.2 ml).
  • Checks. Seven cases looked at one by one show the labels on the kidneys, the masses on renal masses and the cysts on fluid-dense renal lesions. TotalSegmentator 2.18.0 on TCGA-B9-4617 scores a kidney Dice of 0.94, and on it and TCGA-BP-4352 puts the mass in the kidney on the side the label gives.
  • Room. Moved by one voxel in plane, the test masks keep a Dice of 0.95 to 0.99 for the masses and 0.96 to 0.98 for the kidneys; moved by one slice, 0.70 to 0.98 and 0.87 to 0.97. The surface Dice is at 1 mm, the provider giving no tolerance.
  • Overlap. Checked by the patient IDs of every descriptor and prepare.py: no other of the 91 appears elsewhere. The 9 with an IDC MR are left out because TotalSegMR and LiverSegmentsMR hold IDC MR whose patients are not all identified. KiTSCT and AorticTreeCTA (C4KC-KiTS) and CPTACCCRCC_Seg (CPTAC-CCRCC) are other collections with other patient IDs; SAROSBodyCT draws no TCGA kidney collection.
  • Public labels. The annotations are public since 2026: a network trained on them may have seen the test cases.
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBImages
11–4KonfAI segmentation app · MRSegmentator, KonfAI appGPUapp1 of 3 structures0.851 ±0.0670.759 ±0.06314.88 ±10.887.06.1
21–4MONAI wholeBody CTGPUapp1 of 3 structures2 failed0.678 ±0.2610.600 ±0.23214.45 ±11.33163.77.7
31–4KonfAI segmentation app · TotalSegmentator (total), KonfAI appGPUapp2 of 3 structures0.599 ±0.1700.523 ±0.14016.38 ±11.763.26.4
41–4KonfAI segmentation app · TotalSegmentator (total), the authors' packageGPUapp2 of 3 structures0.554 ±0.1620.485 ±0.13516.91 ±12.9214.53.9
5KonfAI segmentation app · TotalSegmentator (kidney_cysts), the authors' packageGPUapp1 of 3 structures0.005 ±0.0140.020 ±0.05071.392416.62.5

The order averages each method's ranks on Dice, Surface Dice, HD95 mm, 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 Workstation.

1 result not ranked: an older run, cases still missing, or a failure on every case
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBMachineImages
MONAI wholeBody CTGPUapp1 of 3 structuresolder run0.847 ±0.0540.727 ±0.06517.96 ±12.5992.48.3Laptop

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

The first test case of RCC-AID
The first test caseClick to look at every case.
Test cases
10 test patients drawn with seed 2026 by collection (8 TCGA-KIRC, 1 TCGA-KIRP, 1 TCGA-KICH) among the 78 of the 91 annotated patients the corpus cannot hold elsewhere, one CT each, then 10 development patients, the last of the same seeded order. Left out: the 4 patients whose CT and MR L2R_AbdomenMRCT holds and 9 others with an MR series in the Imaging Data Commons. The provider gives no split.
Truth
contours by a program, verified by people RCC-AID's label map, one class each: 1 the kidneys without the masses and cysts, 2 the renal masses, 3 the renal cysts. A segmentation model made the first kidney and lesion masks; two trained students checked and corrected them and annotated the tumours and cysts, a board-certified radiologist checking the uncertain cases (the paper's abstract). IDC converted the NIfTI masks to DICOM label map segmentations, every segment marked SEMIAUTOMATIC.
Licence
CC BY 3.0 (TCIA TCGA CTs) and CC BY 4.0 (RCC-AID annotations, IDC's DICOM conversion)
10 cases for tuning, never tested

TCGA-CZ-5451, TCGA-BP-4789, TCGA-B9-4117, TCGA-BP-4965, TCGA-BP-4971, TCGA-KN-8424, TCGA-B0-5097, TCGA-BP-4995, TCGA-CW-6088, TCGA-BP-4768