MedOtter/Spinal-Multiple-Myeloma-SEG

Spinal-Multiple-Myeloma-SEG

Mirror of the TCIA collection Spinal-Multiple-Myeloma-SEG — 67 patients / 72 dual-energy CT studies of the spine with expert-corrected segmentations of osteolytic multiple myeloma lesions and of individual vertebrae (C1–L5).

Acquired on a Philips IQon Spectral CT (dual-layer detector) at University Hospital Brno, 2020–2023. Whole-body skull-to-knee scans were cropped to the spine before publication.

⚠️ That crop is superior–inferior only. In-plane every slice remains a full body cross-section: ReconstructionDiameter is 422–500 mm over 512 columns, which is also where the per-study in-plane spacing of 0.824–0.977 mm comes from (500/512 = 0.977, 422/512 = 0.824). On the busiest axial slice of a study the spine mask covers only 1.3–1.5 % of the slice area, and lesion foreground is 0.019 % of the volume. Expect to localise the spine within a thorax/abdomen cross-section before the target is even a few pixels wide.

• —Source: <https://doi.org/10.7937/k4qv-hh78> (TCIA, Version 1, 2026-03-27)
• —Authors' code: <https://github.com/MISAG-BUT/Spinal-Multiple-Myeloma-SEG>
• —License: CC BY 4.0 (commercial use and redistribution permitted with attribution)

What is mirrored, and what is not

The TCIA collection ships 8 reconstructions per study (304 GB). This mirror carries the two the authors' own pipeline uses, converted DICOM → NIfTI:

Not mirrored: MonoE 80/120 keV and Calcium Suppression 25/50/75/100 (223 GB), and the 144 DICOM SEG objects (redundant with the NIfTI masks below). All eight reconstructions come from one spectral acquisition and share a voxel grid, so the masks are valid on any of them; fetch the others from TCIA if needed.

Geometry

All four volumes of a study share one voxel grid — identical size, spacing, origin and direction — so no resampling or registration is required. Verified on all 72 studies. Orientation LPS, 512×512 in-plane, 511–1103 slices, in-plane spacing 0.824–0.977 mm (varies per study), slice spacing 0.600 mm.

The official Table_of_acquisition_parameters.tsv reports SpacingBetweenSlices = 0.45, which disagrees with the true 0.600 mm slice increment. Geometry here was derived from ImagePositionPatient, matching the released masks. Do not use that tag.

Targets

train.jsonl has 144 rows = 72 studies × 2 targets.

lesions — paired with vmi40

Instance segmentation: one integer per lesion. 4,071 lesions total, 8–138 per study (median 51); no study is empty. For a semantic binary mask, take mask > 0 on the raw array.

⚠️ Instance IDs are not contiguous. Myel_060 has 52 lesions but a max ID of 75; Myel_072 has 64 with a max ID of 99. Assuming 1..N, or min–max normalising the mask, silently drops or distorts labels.

Lesions are small and sparse: 0.019 % of voxels on average (0.0008–0.079 %), and 18.9 % of instances are under 100 voxels. Single-sample debug runs will often see no foreground.

Definition used by the annotators: hypodense, no sclerotic rim, >5 mm largest axial dimension.

spine — paired with convct

Semantic segmentation of individual vertebrae, 1–24 = C1…L5 in VerSe/TotalSegmentator order, plus 25 = an extra/transitional L6.

⚠️ Label sets are per-study and non-contiguous. 15–25 vertebrae are present (mean 23.9) depending on field of view. Label 25 occurs in 11 studies, but the official Notes column flags only 3 of them as "Extra L6 vertebra" — do not drive label handling off that column. Myel_017 and Myel_030/Myel_077 are missing label 19. Use labels_present in train.jsonl.

Splits and patient grouping

No official train/val/test split exists — all 144 rows are split: "train".

Five patients were scanned twice, published as Myel_012_a / Myel_012_b (also 018, 023, 043, 052). So 72 studies but only 67 patients.

⚠️ Group on `patient_id` (Myel_012), never on study_id. The DICOM PatientID tag also carries the unsuffixed form, while the TSVs and SeriesDescription carry the suffixed one.

Annotation provenance

One gold-standard tier; the stages below are how it was built, not competing versions.

• —Vertebrae — nnU-Net trained on VerSe2020, applied to conventional CT, then all errors manually corrected and 3-D inspected.
• —Lesions — 3-round iterative semi-automatic annotation on VMI 40 keV (seeded at 10 scans, retrained at 19 and 43), corrected and verified by two board-certified radiologists (>5 yrs) working on disjoint subsets — so no inter-rater agreement is computable and no rater choice exists.

SegmentAlgorithmType reads AUTOMATIC in the DICOM SEG objects despite the manual correction (a dcmqi default); do not filter or downweight on it.

⚠️ Evaluation note for the `spine` target. The vertebra masks were bootstrapped from a VerSe2020-trained model and inherit VerSe label naming. Benchmarking a VerSe-trained model against them partly measures it against ground truth derived from its own training distribution. The lesions target is unaffected. There is no patient overlap with VerSe, SPIDER, or any other public collection.

Files

dataset/{study}/{study}_convct.nii.gz              conventional CT, int16 HU
                {study}_vmi40.nii.gz               VMI 40 keV, int16 HU
                {study}_spine_segmentation.nii.gz  uint16, 1..25
                {study}_lesions_segmentation.nii.gz uint16, instance IDs
metadata/*.tsv                                     the four official TCIA tables
train.jsonl                                        144 rows, one per (study, target)

train.jsonl fields: study_id, patient_id, study_tag, target, recon, image, mask, shape (z,y,x), spacing_mm (x,y,z), fg_voxels, n_labels, max_label, labels_present (spine only), metal_implant, notes, patient_age, patient_sex, iss_stage, split.

Five studies carry metal_implant: true (spinal instrumentation → streak artifact).

Citation

Nohel, M., Válek, V., Rohan, T., Štork, M., Jakubíček, R., Chmelík, J., Dostál, M. (2026). Segmentation of spinal multiple myeloma lesions in dual-energy CT (Spinal-Multiple-Myeloma-SEG) (Version 1) [Dataset]. The Cancer Imaging Archive. https://doi.org/10.7937/k4qv-hh78

A peer-reviewed descriptor for Version 1 was submitted to Scientific Data and is not yet citable. A v2 extension adding trabecular-bone labels for L1–L5 on 71 of the 72 exams is described in arXiv:2608.01482 and released separately on Zenodo (not included here).