fidoptSTS-3D-Tooth
Dental cone-beam CT crops of two patients, the same side or region of the jaws in both, with the teeth labelled by hand (the 22 labelled crops of STS-3D-Tooth), paired by us.
Read before comparing (9)
- Truth. Every tooth in one label, without tooth numbers. Dental practitioners drew the teeth in ITK-SNAP or corrected a 2D network's proposals by hand, and experienced dentists verified the work; the release does not say which crop went which way. Graded semi-automatic, with no agreement between raters.
- One label. The overlap rewards teeth on teeth, not each tooth on its own: a field that slides a row by one tooth loses little. The pulp is included; the jaw bones, canals and soft tissue are not labelled. Ten masks hold 800 to 14,863 specks under 1 mm³, kept as delivered: they count in HD95.
- Fields. The two fields show the same side, not the same teeth: each image's teeth are cut to the other's round field through the start, 0 to 63 % of a fixed mask on the test pairs. A tooth one patient lacks stays, without a counterpart.
- Tooth landmarks. TotalSegmentator's task teeth, run by us, numbers the teeth (Dice 0.59 to 0.92 against the experts, median 0.84) and mixes the sides on ROI_L_007 and ROI_L_009. A test pair shares 4 to 15 teeth; their centres are the landmarks and the numbered teeth the label set fdi, both automatic, shown, never ranked.
- Alignment. The crops carry no scanner frame: as delivered their centres coincide (teeth Dice 0.04 to 0.30). The corpus gives each pair one elastix translation on mutual information inside each crop, no label used, written in the moving image's, masks' and landmarks' headers; native voxels kept.
- Baseline. With no registration the 10 test pairs score a Dice of 0.45 (0.24 to 0.53), an HD95 of 6.3 mm (4.3 to 10.4), a tooth-centre error of 6.1 mm (1.9 to 10.8) and an fdi Dice of 0.25 (0.01 to 0.49). A rigid fit on the tooth centres left 2.9 mm, but handed the network's teeth to the start.
- Images. 512 x 512 x 400 voxels of 0.156 x 0.156 x 0.15 mm (0.078 mm in plane on ROI_L_009 and ROI_L_017), the field asked on the fixed grid, 105 million voxels. Values from -1000 to 3095 on a scale like Hounsfield units, not calibrated. The spacing is real: the teeth measure their size.
- Cleaning. prepare.py clears what the masks paint beyond the teeth where the images hold the constant 2 (the top slice of 12 crops, the edge row and column of ROI_L_017's slices), and two whole slices of ROI_L_001 and most of one slice of ROI_L_017 in label 1.
- Outside numbers. None: STS-Tooth is a segmentation dataset and the pairing is ours.
| Accuracy | Smoothness | Cost | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Change | HD95 mm | TRE mm * | rTRE | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| 1 | NiftyRegCPUapp | 0.711 ±0.073 | +0.266 | 5.28 ±2.22 | 5.57 ±1.69 | 0.0435 ±0.0132 | 1.451 | 16.32 | 486 | 9.3 | ||
| 2 | greedyCPUapp | 0.583 ±0.089 | +0.138 | 5.93 ±1.56 | 5.72 ±2.21 | 0.0447 ±0.0173 | 0.062 | 0.00 | 454 | 19.2 | ||
| No registration | 0.445 ±0.063 | 6.31 ±1.49 | 6.11 ±2.11 | 0.0478 ±0.0165 | 0.000 | 0.00 | ||||||
The order averages each method's ranks on Dice, HD95 mm, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs 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 Laptop.
* shown, not ranked: on this dataset it is checked against the output of a program.
Dice: how much the organ contours overlap, 1 is perfect. HD95: distance between the contours' surfaces. TRE: distance between matching landmarks. rTRE: the TRE over the length of the image's diagonal, to compare datasets of other sizes and units. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 10 test pairs drawn with seed 2026 among the 11 pairs the 22 crops make; the eleventh, 003_015, is the development pair. Each crop is one patient and in one pair, with a crop of the same side or region.
- Truth
- contours by a program, verified by people landmarks by a program fdi by a program the dataset's tooth masks, every tooth in one label (no tooth numbers): drawn slice by slice in ITK-SNAP by dentists, or proposed by a 2D network and corrected by hand, then reviewed by senior dentists the centre of each tooth that both images hold, as TotalSegmentator (task teeth) run by us numbers them, 4 to 15 a pair: their error is shown, never ranked
- Source
- zenodo.org/records/10597292
huggingface.co/datasets/MedOtter/STS-3D-Tooth
doi.org/10.1038/s41597-024-04306-9
arxiv.org/abs/2407.13246
arxiv.org/abs/2206.08778 - Licence
- CC BY 4.0
1 pair for tuning, never tested
003_015