SegThy — Thyroid and Neck Vessel Segmentation (3-D Ultrasound + MRI)

SegThy pairs electromagnetically tracked freehand 3-D ultrasound with T1-VIBE MRI of the same healthy volunteers, and annotates the thyroid gland, both common carotid arteries and both internal jugular veins by hand in each.

Two things make it unusual. It is one of very few public datasets with 3-D reconstructed freehand ultrasound rather than 2-D B-mode frames — the sweeps are tracked and resampled to a 0.12 mm isotropic volume. And 14 subjects carry multi-class labels in both modalities, which is what makes it a genuine MR/US registration benchmark rather than two unrelated collections.

What this mirror contains — read first

⚠️ The 504-volume network_results/ tier is deliberately NOT mirrored. It is 94 % of the upstream archive (17.3 GiB) and it is not ground truth. Upstream's own Readme: "contains the US scans and corresponding thyroid segmentations from the trained QuickNAT" — these are CNN predictions, and the object of study in the paper, not a reference standard. Three independent reasons to exclude them:

1. They are model output, so evaluating against them measures agreement with QuickNAT, not with a human.
2. Their labels do not overlay their own images. The label arrays are stored in the network's resampled frame — (ceil₄(Z_image), 448, 384), axes permuted with Z first — so shapes disagree with the image in every pair sampled (12/12). They cannot be used without inverting an undocumented resampling.
3. 31 of the 32 gold scan IDs reappear inside them with conflicting labels, so loading both tiers silently double-counts almost the whole gold set.

Get them from the source if you want them. Everything mirrored here is 100 % manually annotated.

⚠️ The public release is Sub-dataset 1 only. The project describes two sub-datasets; only the 28-volunteer one was ever released. Sub-dataset 2 (186 routine-care patients) is not downloadable and is not here. The homepage also announces trachea and thyroid-nodule labels as "currently being extended" — those have never shipped, and the 2022 Readme's promise that the remaining MRI vessel labels would arrive "in the following weeks" is still outstanding at the 2025-05-09 file date (14 of 28). This dataset has no nodules and no pathology; the cohort is healthy volunteers.

⚠️ Upstream's US Readme says "The label maps are binary". That is false for the data mirrored here. The same paragraph also states the vessels are manually segmented, contradicting itself. Verified on all 32 gold masks: left-lobe scans carry {0,1,2,3}, right-lobe scans carry {0,1,4,5}. A loader that assumes binary will silently relabel every carotid and jugular voxel as thyroid. The "binary" sentence describes the excluded network_results/ tier.

Dataset Details

Field	Value
Modalities	Tracked freehand 3-D ultrasound (Siemens Acuson NX-3, VF12-4 12 MHz, piur tUS) · T1-weighted VIBE MRI (Siemens Biograph mMR 3 T)
Body part	Neck — thyroid gland, common carotid arteries, internal jugular veins
Cohort	28 healthy volunteers, single centre (TUM / Klinikum rechts der Isar)
Annotation	100 % manual, single tier — no rater or tier to choose
US spacing	0.12 mm isotropic (0.11 mm for subjects 006 and 026 — the Readme's flat "0.12" is approximate)
MRI spacing	0.625 × 0.625 × 1.0 mm, all 28 volumes
US shapes	variable, 457–617 × 363–526 × 359–696
MRI shape	320 × 320 × 80, all 28 volumes
Dtypes	US image + mask uint8 · MRI image uint16, masks uint8
Intensity	US 0–247 (as documented) · MRI 0–1666
Split	none upstream — one train split per config
License	CC BY, commercial use explicitly permitted — shipped inside both archives
Paper	Krönke et al., PLOS ONE 17(7):e0268550, 2022 · doi:10.1371/journal.pone.0268550

Label map

Identical in both modalities, and confirmed against upstream's MRI Readme ("the thyroid value being 1, the jugular vein being 3 and 5, the carotid being 2 and 4"):

Value	Structure
0	background
1	thyroid
2	carotid artery, left
3	internal jugular vein, left
4	carotid artery, right
5	internal jugular vein, right

Classes are disjoint — every voxel carries exactly one label, so this is a single label map and not a fan-out into overlapping binary channels.

⚠️ In the us config the class ID depends on which lobe was scanned. Each US volume covers one lobe, so a left scan contains only {0,1,2,3} and a right scan only {0,1,4,5}. The three missing classes are missing because that anatomy is outside the field of view, not because it was left unannotated — so zero-filling them is correct here. Contrast this with the MRI configs, where both sides are in every volume.

The one genuine annotation gap is 028_P3_1_left, whose alphabet is {0,1,2}: the left jugular is in the field of view but was not drawn, and it is annotated in all 15 other left scans. It is flagged annotation_gap in the jsonl.

The three configs

Config	Volumes	Subjects	Classes	Notes
us	32	16	thyroid + vessels of the scanned side	2 per subject (left + right lobe)
mri_multiclass	14	14	all 5	the richest tier
mri_thyroid	28	28	thyroid only	widest subject coverage

⚠️ The two MRI configs are not nested, and their thyroid labels disagree

The 14 mri_multiclass subjects are a strict subset of the 28 mri_thyroid subjects and share the same underlying MRI scans — so do not evaluate on both and pool the numbers; 14 subjects would be counted twice.

More surprisingly, for those 14 shared subjects the binary mask is not the multiclass thyroid channel. They are two independent delineations:

	Dice(binary, multiclass==1)
median	0.912
range	0.750 – 0.999

The binary-only voxels land on multiclass background, not on vessel labels, so this is not vessels being absorbed into the gland — it is ordinary inter-delineation disagreement, at roughly inter-observer magnitude.

Which one to trust? Independent evidence favours mri_multiclass. Summing the two US lobe volumes per subject and comparing against the MRI gland volume — the very quantity the paper studies — gives:

MRI tier	median US-sum / MRI ratio	n
mri_multiclass	0.947	14
mri_thyroid	1.264	15

Two independent modalities and annotators agreeing to ~5 % is a strong signal; the binary tier is ~26 % smaller than the same subjects' ultrasound, i.e. it appears to under-segment. Use mri_multiclass when both are available. A Dice measured on one MRI config is not comparable to a Dice measured on the other.

⚠️ Subject 005's binary mask is 4-D upstream

005_MRI_thyroid_label.nii.gz ships as (320, 320, 80, 2) — the only non-3-D file in the whole release (1 of 134). Volume index 1 is entirely empty (0 foreground voxels); index 0 holds the real 20,294-voxel mask, whose size is unremarkable among its peers (median 17,453, range 7,251–42,819) and whose Dice against the multiclass thyroid (0.723) sits in the same band as every other subject.

np.squeeze does not fix this — the trailing axis has length 2, not 1, so it survives the squeeze and breaks any loader that assumes rank 3.

This mirror therefore ships [..., 0] as the canonical mask, with the affine and header preserved, flagged mask_repaired: true in the jsonl. The untouched original is kept at originals/005_MRI_thyroid_label.original_4d.nii.gz, and both digests are in metadata/manifest.csv, so nothing is lost and the change is auditable. Every other file is bit-identical in payload to upstream.

Splits, grouping and leakage

No split ships upstream — the archives are flat folders with no split file. The paper used an internal 26/6/6 lobe split that is not distributed. Rather than invent a boundary, each config ships one train split and the loader declares a single-split fallback.

Group on subject_id, never on case_id. A subject contributes up to two us rows (left and right lobe of the same neck, same session) plus an MRI row in each MRI config. Splitting those across a train/test boundary leaks. subject_id is the 3-digit volunteer number and is consistent across all three configs and both modalities — subject 007 is the same person everywhere.

Choosing a slicing axis — axis 2

For 2-D slice-wise use, axis 2 carries by far the most annotated slices:

Config	axis 0	axis 1	axis 2
us	71.7 %	39.1 %	81.4 %
mri_multiclass	39.7 %	17.2 %	100.0 %
mri_thyroid	18.0 %	13.0 %	48.1 %

(median fraction of slices containing any foreground). mri_multiclass reaches 100 % because the carotids and jugulars run the full superior–inferior extent of every volume, while the thyroid alone covers about half of it.

Cohort anomalies

• Subject 029 appears only in the US gold tier (2 volumes). Every source — homepage, paper, both Readmes — says 28 volunteers and numbers them 001–028. There is no 029 MRI and no 029 anywhere in network_results. Its scans are well-formed and normally annotated; it is mirrored as-is and flagged in the cross-reference.
• Subject 004 has an MRI and a binary thyroid mask but no vessel labels and no US gold data.
• 011 and 029 have US gold but no mri_multiclass, which is why the both-modalities multi-class intersection is 14 subjects rather than 16.

Structure

us/train/images/001_P1_1_left.nii.gz          #  32 tracked 3-D US volumes
us/train/masks/001_P1_1_left.nii.gz           #  32 masks, same grid
mri_multiclass/train/images/001.nii.gz        #  14 T1-VIBE MRI volumes
mri_multiclass/train/masks/001.nii.gz         #  14 five-class masks
mri_thyroid/train/images/001.nii.gz           #  28 T1-VIBE MRI volumes
mri_thyroid/train/masks/001.nii.gz            #  28 binary thyroid masks

us_train.jsonl                                # per-case metadata, one file per config
mri_multiclass_train.jsonl
mri_thyroid_train.jsonl

metadata/manifest.csv                         # payload_sha256 + bytes + shape, all files
metadata/subject_crossref.csv                 # which tiers each subject appears in
originals/005_MRI_thyroid_label.original_4d.nii.gz   # untouched 4-D original

Readme_MRI.txt                                # upstream, verbatim
Readme_US.txt                                 # upstream, verbatim
README.md
LICENSE.txt

The preview_* configs the Dataset Viewer shows are thumbnails, not the data. They hold one rendered PNG triplet (image / colour-coded mask / overlay) per volume, purely so the collection can be browsed in the Hub UI. Load the .nii.gz files listed in the jsonl for any real use — the previews are 8-bit, single-slice and lossy.

US case IDs follow upstream's grammar NNN_PX_S_side: subject, physician (P1–P3), repeat index, lobe. The gold tier holds exactly one scan per subject per lobe, so the physician and repeat fields vary between subjects but never within one.

jsonl columns

Column	Meaning
case_id	unique within the config
subject_id	3-digit volunteer number — the grouping key, consistent across configs
config, modality	us/mri_multiclass/mri_thyroid, US/MRI
image, mask	repo-relative paths
split	always "train" (no upstream split exists)
side, physician, scan_index	us only, parsed from the filename
shape_xyz, n_slices, spacing_mm	geometry, read per file
axcodes, sform_code, qform_code	header provenance
image_dtype, mask_dtype, intensity_min, intensity_max
mask_values	the label alphabet actually present
class_voxels, class_volume_ml	per class, keyed by label ID
thyroid_volume_ml	class 1 volume — clinically the quantity of interest
fg_voxels, n_voxels, foreground_fraction
fg_slice_fraction, leading_bg_slices	per axis ("0","1","2")
annotation_gap	non-null only for 028_P3_1_left
mask_repaired, mask_repair_note	true only for mri_thyroid subject 005
has_us_gold, has_mri_multiclass, has_mri_binary	cross-tier availability
payload_sha256	digest of the uncompressed NIfTI bytes — comparable to upstream regardless of gzip framing
image_sha256, mask_sha256, image_bytes, mask_bytes	as stored here

Overlap and contamination

• ⚠️ SegThy is inside UltraSam's US-43d training corpus (arXiv 2411.16222, listed as Segthy-Dataset thyroid). Benchmarking UltraSam or its derivatives on SegThy is contaminated.
• No patient overlap with TN3K / TG3K / DDTI / TNSCUI. Those are 2-D B-mode nodule datasets from Chinese and Colombian hospitals; SegThy is 3-D tracked freehand US plus MRI of healthy German volunteers, targeting the gland and vessels. Different target, cohort and acquisition. No cross-reference ID exists or is needed.
• No TCIA, Medical Segmentation Decathlon or BraTS lineage — single-centre TUM acquisition under TUM Ethics Commission approval.
• Clean with respect to MedSAM, MedSAM2, SAM-Med2D and SAM-Med3D.
• Reference ceiling: Munsterman et al., WFUMB Ultrasound Open 2:100055 (2024), doi:10.1016/j.wfumbo.2024.100055, trained 2-D/3-D U-Nets on the SegThy tracked sweeps and report median Dice 0.934 / 0.924 / 0.897 for thyroid / carotid / jugular.

Source & Citation

• Official, author-hosted: https://www.cs.cit.tum.de/camp/publications/segthy-dataset/
• Direct: https://www.campar.in.tum.de/public_datasets/2022_plosone_eilers/
• Avoid the Academic Torrents copy — it carries US_data.zip only and its size does not match the current 2025-05-09 revision.

@article{kronke2022segthy,
  author  = {Kr{\"o}nke, Markus and Eilers, Christine and Dimova, Desislava and
             K{\"o}hler, Melanie and Buschner, Gabriel and Schweiger, Lilit and
             Konstantinidou, Lemonia and Makowski, Marcus R. and Nagarajah, James
             and Navab, Nassir and Weber, Wolfgang and Wendler, Thomas},
  title   = {Tracked 3D ultrasound and deep neural network-based thyroid
             segmentation reduce interobserver variability in thyroid volumetry},
  journal = {PLOS ONE},
  volume  = {17},
  number  = {7},
  pages   = {e0268550},
  year    = {2022},
  doi     = {10.1371/journal.pone.0268550}
}

Note: the in-archive Readme credits "Lilit Mirzojan"; PLOS ONE publishes the same author as "Lilit Schweiger".