TrackRAD2025 — labeled subsets

Real-time tumour tracking for MRI-guided radiotherapy: 2D+t sagittal cine-MRI acquired on two MR-Linac platforms, with a single binary target contoured on every frame.

This is a mirror of the labeled portion of the official LMUK-RADONC-PHYS-RES/TrackRAD2025 release (DOI 10.57967/hf/4539).

Scope — read this first

The upstream repo is 269 GB, of which 268 GB is the unlabeled_training_data pool (477 patients). Those cases ship no targets/ directory at all and carry zero segmentation ground truth, so they are excluded here. This mirror is the 1.0 GB labeled portion, which is the complete usable set for segmentation:

Split	Upstream subset	Patients	Frames	Size
train	trackrad2025_labeled_training_data	50 (A25 / B15 / C10)	5,027	576 MB
val	trackrad2025_labeled_pre-testing_data	8 (A2 / B3 / C3)	607	74 MB
test	trackrad2025_labeled_testing_data	30 (A5 / B8 / C11 / X6)	2,298	324 MB
Total		88	7,932	976 MB

val is the challenge's own "pre-testing" set — the 8-case public validation cohort used during the pre-test phase. Splits are the official challenge splits, preserved as-is.

Cohort D (D_001–D_020) is permanently withheld by the organisers for privacy. Public = 88 of the paper's 108 labeled patients; the missing 20 are exactly Cohort D.

Acquisition

Modality	2D+t sagittal cine-MRI
Scanners	0.35 T ViewRay MRIdian (bSSFP, 4 / 8 Hz) — 32 cases · 1.5 T Elekta Unity (bFFE, 1.3–3.5 Hz) — 56 cases
Regions	abdomen 47 · thorax 27 · pelvis 14
Centers	6, letter-anonymised A–F
Image dtype	uint16, resampled to 1 × 1 mm in-plane
Mask dtype	uint8, strictly {0, 1} — 1 class
Frames / case	44 – 248 (median 97)

The target is normally the GTV, but where tumour contrast was clinically too low the organisers contoured a surrogate structure instead (e.g. the whole liver in place of a low-contrast liver lesion). Treat the class as "the tracked structure", not strictly "tumour".

Layout

{subset}/{case_id}/images/{case_id}_frames.mha        uint16 (H, W, T)
{subset}/{case_id}/targets/{case_id}_labels.mha       uint8  (H, W, T)  <- ground truth
{subset}/{case_id}/targets/{case_id}_first_label.mha  uint8  (H, W, 1)  <- input prompt
{subset}/{case_id}/targets/{case_id}_labels2.mha      uint8  (H, W, T)  <- 2nd observer, 14 cases
{subset}/{case_id}/b-field-strength.json              0.35 | 1.5
{subset}/{case_id}/frame-rate.json                    Hz
{subset}/{case_id}/scanned-region.json                thorax | abdomen | pelvis
train.jsonl  val.jsonl  test.jsonl                    per-case manifests (repo-relative paths)

Note the file is b-field-strength.json, not field-strength.json — the upstream card's folder diagram has this wrong.

Gotchas — all verified against all 88 cases

1. Axis order is (H, W, T) — time is the LAST numpy axis. SimpleITK's GetSpacing() returns (5.0, 1.0, 1.0), and that leading 5 mm is slice thickness parked on the time axis — an artifact of writing 2D+t as a 3D volume. Do not read it as a 3D volume with 5 mm z-spacing. The official evaluator does transpose(2, 0, 1) to reach (T, H, W).

2. Ground truth is DENSE, not sparse. Despite being a tracking challenge, every frame is labeled: T(labels) == T(frames) in 88/88 cases, zero mismatches.

3. _first_label.mha is byte-identical to labels[..., 0] (verified 88/88). It is the algorithm's input prompt (Grand Challenge interface mri-linac-target), not an extra annotated frame. The dataset paper's phrase "the labels of the remaining frames" is wrong — concatenating first_label onto labels yields an off-by-one T+1.

4. In-plane size is ragged across cases: 270, 350, 423, 424, 425, 426, 437, 450 (all square). Do not np.array() a mixed batch.

5. Labels are already {0, 1}. No min-max normalisation or > 0.5 threshold — a binarisation recipe applied here is a no-op at best.

6. 3 cases contain empty frames (target out of plane): A_013 14/100, A_018 12/100, B_018 35/70. 7,871 of 7,932 frames have foreground. The official evaluator excludes empty-GT frames from metrics.

Ground-truth tiers

File	Role	Coverage
_labels.mha	Primary observer, per-frame	all 88 — use this
_labels2.mha	Second independent observer	14 cases, all Center C (655 frames)
_staple_labels.mha	Official gold standard	not distributed by the organisers

The official evaluation/evaluate.py reads _labels.mha and then overrides it with _staple_labels.mha where present — but the STAPLE files were never released. To match official scoring on the 14 two-observer cases you must recompute STAPLE yourself from labels + labels2; this is not cosmetic (measured observer-1 vs observer-2 DSC on C_016 = 0.871). This mirror therefore uses _labels.mha uniformly so the GT tier is consistent across all 88 cases, and ships labels2 alongside for anyone who wants to reconstruct STAPLE.

Center D used 5 observers, which is the source of the paper's "+8000 multi-observer frames" — but D is withheld, so only ~655 multi-observer frames are actually public.

Grouping / leakage

Group on case_id; each case is one continuous cine sequence.

⚠️ Cohort X is not provably independent of B/C. X_001–X_006 (test split) are Elekta CMM-sequence acquisitions drawn from one of the source centers A–D — necessarily B or C, since X is 1.5 T and labeled. Whether they are the same patients as some B_* / C_* cases is unstated upstream. If you re-split away from the official splits, treat X as potentially non-independent from B and C.

Provenance

Official organiser release, DOI-minted, no third-party re-host. Counts reconcile exactly: 585 = 477 unlabeled + 50 train + 8 pre-test + 50 test; public = 585 − 20 (Cohort D).

The upstream data is itself a preprocessed variant — resampled to 1 × 1 mm, reoriented, with the first 5 frames (0.35 T) / 3 frames (1.5 T) dropped to reach steady state plus any leading frames where the target was not yet visible (so frame 0 always contains the target). Sagittal plane only: 1.5 T acquisitions were interleaved sagittal/coronal(/axial) and only the sagittal series is included.

No documented patient overlap with any other public dataset, and no cross-reference ID column exists (case_id is <center letter>_<3 digits>).

License

CC BY-NC 4.0 — as stated in the dataset paper §2.3 and the upstream card. Note the article is CC BY 4.0; that applies to the text, not the data.

Citation

@article{trackrad2025data,
  title   = {TrackRAD2025 challenge dataset: real-time tumor tracking for
             MRI-guided radiotherapy},
  author  = {Wang, Yiling and Lombardo, Elia and Thummerer, Adrian and
             Bl{\"o}cker, Tom and Maspero, Matteo and others},
  journal = {Medical Physics},
  volume  = {52}, number = {7}, pages = {e17964}, year = {2025},
  doi     = {10.1002/mp.17964}
}

@article{trackrad2025challenge,
  title   = {MRIgRT real-time target tracking: TrackRAD2025 challenge report},
  author  = {Bl{\"o}cker, Tom J. and G{\"o}rts, Pim A. W. and Wang, Yiling and
             Lombardo, Elia and Landry, Guillaume and others},
  journal = {Medical Image Analysis},
  volume  = {112}, pages = {104134}, year = {2026},
  doi     = {10.1016/j.media.2026.104134}
}