fidoptTotalSegmentator MRI, any region
A routine clinical MR of any region, sequence and scanner, adult or child, from the University Hospital Basel or the Imaging Data Commons, the 50 structures of TotalSegmentator's task total_mr labelled (TotalSegmentator MRI dataset 3.0.0).
61 test caseswhole bodyimage MR
Read before comparing (8)
- Source. Version 3.0.0 holds 1296 MR images 'randomly sampled from clinical routine': version 2.0.0's 616, from Basel and the Imaging Data Commons, adults only, and 680 more, among them whole-body images, children and images 'where the model failed before'. meta.csv splits them into 1184 training and 112 test images, 2.0.0's 55 test images unchanged.
- Pool. Of the 112 test images 61 remain. Left out of both pools: the images TotalSegMRI and LiverSegmentsMR hold under other ids, found by grid and voxels; the 36 that share a grid with another, some across the two splits; those whose masks are all empty (6 test images); the three with broken masks (see Labels); and those whose age, sex, institute and study type both splits hold (32 test images), meta.csv naming no patient.
- Images. The values are as delivered, 16-bit integers on no common scale, but for s1289, whose values lay between -32768 and -30965, as if unsigned values had been moved into a signed type: prepare.py adds 32768 back. On 95 images, s1176 among them, prepare.py sets the qform_code so that ITK reads the rotation the sform holds.
- Truth. For version 2.0.0, two board-certified radiologists segmented the first images by hand or refined MRSegmentator's or SynthStrip's output; retrained nnU-Nets predicted the next ones, refined by hand where needed, until all 616 were reviewed. Version 3.0.0 'improved the label quality' and does not say how its 680 new images were annotated. No agreement between readers is given: graded semi-automatic.
- Labels. Each image holds the structures of its field of view, cut where it ends, and a diseased organ keeps its label. Where masks overlap at their borders, prepare.py gives the voxel to the smaller structure, as TotalSegMRI does. Where that rule fails, the flagged images were looked at one by one: s1094, s1268 and s0892, whose masks lie on another organ's tissue, are left out; s0327 and s0367 are kept.
- TotalSegmentator. Its MR models learned this dataset's training split, so its line here is not a fair comparison: total_mr's '1088subj' weights, the default in TotalSegmentator 2.18, held out version 2.0.0's 55 test MRIs and never saw 3.0.0's added images. The labels began as predictions of networks of its kind, so the truth shares its view of each boundary.
- Overlap. By IDC's index, s0031 matches series of TCGA-KIRC, TCGA-LIHC and ACRIN-NSCLC-FDG-PET, s0042 of TCGA-KIRP, TCGA-LIHC, CPTAC-CCRCC, CPTAC-PDA and CMB-CRC, and s0309 of TCGA-KIRC alone (TCGA-BP-4761), none of L2R_AbdomenMRCT's eight TCGA patients. s0042 is an MR series of CPTAC-CCRCC's C3L-00800, whom CPTACCCRCC_Seg leaves out. In the other test images of the Imaging Data Commons, a patient of SAROSBodyCT or L2R_AbdomenMRCT cannot be ruled out, as another series, an MR.
- Outside numbers. On version 2.0.0's internal test set the paper reports a Dice of 0.862 over the 50 structures for TotalSegmentator MRI at 1.5 mm (0.779 at 3 mm; MRSegmentator 0.759 against 0.862 on the 40 structures it shares): room is left. The surface Dice here is taken at 3 mm, the tolerance of the paper's normalised surface distance.
| Accuracy | Cost | |||||||
|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Surface Dice | HD95 mm | Time s | RAM GB | GPU GB | Images |
| 1 | KonfAI segmentation app · TotalSegmentator (total_mr), the authors' packageGPUapp | 0.718 ±0.046 | 0.765 ±0.047 | 12.25 ±2.94 | 31 | 14.2 | 3.3 | |
| 2 | KonfAI segmentation app · TotalSegmentator (total_mr), KonfAI appGPUapp | 0.631 ±0.054 | 0.675 ±0.056 | 15.21 ±3.52 | 25 | 2.5 | 3.5 | |
| 33–4 | KonfAI segmentation app · IMPACTSeg body, KonfAI appGPUapp2 of 50 structures | 0.105 ±0.052 | 0.086 ±0.043 | 63.47 ±26.49 | 22 | 2.0 | 3.1 | |
| 43–4 | KonfAI segmentation app · MRSegmentator, KonfAI appGPUapp40 of 50 structures1 failed | 0.039 ±0.041 | 0.039 ±0.043 | 149.73 ±16.65 | 23 | 3.0 | 5.9 | |
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 61 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.
3 results not ranked: an older run, cases still missing, or a failure on every case
| Accuracy | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Surface Dice | HD95 mm | Time s | RAM GB | GPU GB | Machine | Images |
| KonfAI segmentation app · IMPACTSeg body, KonfAI appGPUapp2 of 50 structuresolder run | 0.105 ±0.052 | 0.086 ±0.043 | 63.48 ±26.50 | 7 | 1.8 | 4.3 | Laptop | ||
| KonfAI segmentation app · MRSegmentator, KonfAI appGPUapp40 of 50 structuresolder run | 0.039 ±0.041 | 0.039 ±0.043 | 149.73 ±16.65 | 11 | 2.5 | 9.2 | Laptop | ||
| KonfAI segmentation app · TotalSegmentator (total_mr), KonfAI appGPUappolder run | 0.628 ±0.054 | 0.672 ±0.056 | 15.26 ±3.52 | 9 | 2.2 | 5.6 | Laptop | ||
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 caseClick to look at every case.
- Test cases
- The test cases are the 61 images of meta.csv's test split that prepare.py leaves: the provider's test split taken whole, s0892 left out for a misplaced mask. The 30 development images are the last of the training pool's 791 in the order of sha256("2026:TotalSegMR:<image_id>").
- Truth
- contours by a program, verified by people the dataset's masks of the 50 structures of TotalSegmentator's task total_mr on each image, merged into one map in its numbering (a voxel two masks claim, at most 0.44 % of a test image's labelled voxels, goes to the smaller structure, as TotalSegMRI merges them); a test image holds 3 to 43 of them, those of its field of view
- Source
- zenodo.org/records/22688334
doi.org/10.1148/radiol.241613
arxiv.org/abs/2405.19492
zenodo.org/records/14710732
github.com/wasserth/TotalSegmentator
learn2reg.grand-challenge.org/Datasets
github.com/ImagingDataCommons/idc-index - Licence
- CC BY 4.0 (the record); the images from the Imaging Data Commons keep their collection's licence, CC BY 3.0 or CC BY 4.0
30 cases for tuning, never tested
s0401, s0951, s0247, s0359, s0789, s1176, s1168, s0146, s0244, s0611, s0211, s0147, s0361, s0672, s0427, s0351, s0566, s1191, s1252, s0231, s1105, s0621, s0363, s0574, s1249, s1146, s0527, s0212, s0435, s1284