fidoptHaN-Seg
Planning CT of the head and neck of one radiotherapy patient at the Institute of Oncology Ljubljana, with 30 organs at risk delineated by experts and curated (HaN-Seg, public set 1).
Read before comparing (9)
- Case. The planning CT alone, the deposit's NRRD file byte for byte, in Hounsfield units the provider capped at 3000 (metal). The field of view, up to 800 mm across, holds the couch. No mask is given.
- MR. Each patient also has a T1-weighted MR, taken 0 to 11 days later in the same mask, on its own grid and not registered to the CT. The labels were drawn with its help, so a CT-only method is at a disadvantage on the soft-tissue organs: brainstem, optic chiasm and nerves, pituitary, glands.
- Truth. One delineation a patient by one of two annotation teams, curated; graded expert, with no second reader on these cases. The provider's variability study (Podobnik et al. 2024) measures the spread between observers on other patients. The pharyngeal constrictor muscles were left out by the provider as unreliable.
- Labels. The deposit's TG-263 names map to ours. Eye_AL/AR and Eye_PL/PR are the anterior and posterior eyeball segments, whose union is the eyeball; the carotid arteries are cut below the pituitary gland by the curator.
- Small structures. The 30 masks never overlap and all are present on every case. Seven hold under 1 mL as a rule (arytenoids, cochleae, lacrimal glands, optic chiasm, optic nerves, pituitary gland): there one voxel's shift weighs heavily on the Dice, and the surface Dice at 1 mm decides.
- Checks. Each label map lies on its CT's grid; every left structure lies on the patient's left of the brainstem and every right one on the right. The masks sit on the anatomy in the slices looked at on case_07, case_21 and case_42.
- TotalSegmentator. Version 2.18 (head_glands_cavities) on case_21 matches our sides. Its Dice with the truth, 0.59 and 0.58 for the parotids and 0.82 and 0.84 for the submandibular glands, shows a difference of protocol more than an error of the truth: the truth is not that model's output.
- Outside numbers. None on these cases: the challenge scored set 2 only. The provider's nnU-Net was trained on set 1, which holds these cases.
- Overlap. No descriptor of the corpus names HaN-Seg, Ljubljana or Slovenia; its other head and neck sets are other cohorts (PDDCA, HNTSMRG, HNMRLinacADC, HNCBCTCT, SynthRAD2023 and SynthRAD2025).
| Accuracy | Cost | |||||||
|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Surface Dice | HD95 mm | Time s | RAM GB | GPU GB | Images |
| 1 | KonfAI segmentation app · TotalSegmentator (craniofacial_structures), the authors' packageGPUapp1 of 30 structures | 0.943 ±0.011 | 0.956 ±0.017 | 1.72 ±0.79 | 53 | 33.0 | 6.0 | |
| 22–4 | KonfAI segmentation app · TotalSegmentator (head_glands_cavities), the authors' packageGPUapp6 of 30 structures | 0.644 ±0.037 | 0.578 ±0.041 | 12.61 ±1.51 | 41 | 25.4 | 4.3 | |
| 32–4 | KonfAI segmentation app · TotalSegmentator (total), the authors' packageGPUapp2 of 30 structures | 0.599 ±0.022 | 0.439 ±0.049 | 38.22 ±4.17 | 26.3 | 8.0 | ||
| 42–4 | KonfAI segmentation app · TotalSegmentator (total), KonfAI appGPUapp2 of 30 structures | 0.597 ±0.024 | 0.433 ±0.048 | 38.31 ±4.22 | 58 | 15.2 | 4.4 | |
| 5 | CT-FM whole-body segmentationGPUapp2 of 30 structures | 0.404 ±0.087 | 0.303 ±0.082 | 55.69 ±19.45 | 29.4 | 6.8 | ||
| 6 | KonfAI segmentation app · IMPACTSeg body, KonfAI appGPUapp1 of 30 structures4 failed | 0.198 ±0.123 | 0.054 ±0.040 | 74.70 ±12.39 | 17 | 2.3 | 7.1 | |
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.
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
- Test cases
- 10 test cases drawn with seed 2026 and 10 development cases, one a patient, from the 41 patients of public set 1: the deposit's 42 less case_19, whose optic chiasm has no mask and whose brainstem, optic nerves and pituitary gland are marked partial in OAR_data.csv. The provider's test set (set 2, 14 patients) is not released.
- Truth
- contours by experts the deposit's 30 binary masks on the CT, one label each: drawn by hand in ARIA, on 29 patients by an RT technologist trained for organs at risk (the MR registered to the CT as support), on 27 by radiation oncologists (organs better seen on the MR drawn there and propagated to the CT), all supervised by a senior radiation oncologist, then curated in Seg3D by a medical imaging researcher (jagged edges, isolated pixels, wrong labels; the carotid arteries cut below the pituitary gland) (Podobnik et al. 2023)
- Source
- zenodo.org/records/7442914
doi.org/10.1002/mp.16197
lit.fe.uni-lj.si/en/research/resources/HaN-Seg
doi.org/10.1016/j.radonc.2024.110410
han-seg2023.grand-challenge.org - Licence
- CC BY-NC-ND 4.0
10 cases for tuning, never tested
case_29, case_38, case_10, case_28, case_34, case_39, case_22, case_24, case_05, case_15