MedOtter/PROSTATEx-Seg-HiRes

PROSTATEx-Seg-HiRes

Whole-gland binary prostate segmentation on axial T2-weighted MRI — 66 patients from the SPIE-AAPM-NCI PROSTATEx challenge cohort, annotated by tri-planar fusion.

Upstream: High Resolution Prostate Segmentations for the ProstateX-Challenge, TCIA analysis result doi:10.7937/TCIA.2019.DEG7ZG1U.

What "HiRes" means — read this first

"High resolution" describes the MASK's sampling, not the imaging. The images are ordinary 3 mm-thick T2 stacks, exactly like every other PROSTATEx product. The authors delineated the gland independently in the axial, sagittal and coronal T2 series, then fused the three anisotropic masks by averaging linearly-resampled signed distance transforms and thresholding at zero — producing a mask sampled at 0.5–0.6 mm isotropic (106–171 planes). The name is honest but easy to misread as a higher-resolution imaging variant. It is not.

What is in this mirror

Upstream ships segmentations only — 66 DICOM-SEG objects and no pixel data. The images come from the separate base SPIE-AAPM-NCI PROSTATEx collection (doi:10.7937/K9TCIA.2017.MURS5CL, also CC BY 3.0). This mirror performs that join once so downstream users never repeat it.

Volumes live at the repo root, not under data/. Dataset.push_to_hub() owns data/ and deletes any non-parquet file it finds there, so nesting the NIfTI under data/ would wipe them the next time the preview is refreshed.

The join key is ReferencedSeriesSequence, never SeriesDescription

Each SEG names its source series in ReferencedSeriesSequence[0].SeriesInstanceUID, and this mirror resolves all 66 that way. Matching on SeriesDescription would be wrong: several PROSTATEx studies carry two MR series both described t2_tse_tra, so a description match silently selects the wrong volume — and since image and mask would then simply disagree everywhere, it surfaces as a plausible-looking low score rather than a crash.

Why the mask was resampled, and what it cost

The upstream mask is on its own isotropic grid while the image is 19–25 slices at 3–5 mm. They are not the same grid — but ImageOrientationPatient is identical in all 66 cases (max angle between slice normals 0.027°) and the in-plane FOV matches exactly, so the pairing is a pure axis-aligned resample in the acquisition frame: no registration, no rotation, no cropping.

Linear interpolation of the binary followed by a 0.5 threshold (partial-volume majority) preserves gland volume to a factor of 0.9886–1.0104 (median 1.0013) across all 66 cases, and no mask is emptied. The per-case factor is recorded as volume_retained. The native mask is shipped alongside so nothing is lost.

images/ is never interpolated — evaluation runs on real acquired pixels.

Cohort

Splits: none

TCIA publishes a flat pool of 66 with no official partition, so everything is exposed as train. The paper used its own 19-test / 47-CV division, but it was never released. Loaders that request val/test should require an explicit single-split-fallback opt-in rather than silently self-evaluating on the whole cohort.

Ground truth: one tier, human ceiling 0.93

A medical student delineated each gland separately in all three planes; an expert urologist reviewed and corrected each in 3D Slicer with all three orthogonal scans visible; the fused isotropic mask was then reviewed once more. About 10% of cases needed manual rigid registration for inter-plane motion. TCIA distributes only the primary reader (ContentCreatorName = Reader1 on all 66), so there is no rater ambiguity here. The paper reports inter-observer whole-gland DSC 0.93 on a 20-case subset (apex 0.90, mid 0.96, base 0.89) — read model scores against that, not against 1.0.

⚠️ Overlap — never count this as an independent cohort

All 66 patients are a strict subset of the 98 in [`MedOtter/PROSTATEx-Seg-Zones`](https://huggingface.co/datasets/MedOtter/PROSTATEx-Seg-Zones), and both reference the identical source T2 series. This dataset contributes no new patients and no new images — only a different annotation of the same glands.

It is not derivable from Zones, though. Measured on all 66 shared patients after resampling both onto the common T2 grid, Dice between this whole-gland mask and the union of the Zones four-zone map is 0.9522 mean / 0.9553 median (range 0.8461–0.9918) — right at the 0.93 human inter-observer ceiling. The HiRes gland is systematically tighter: median 5.8% smaller, and in the low-agreement cases ~99% of the HiRes mask lies inside the Zones union, a containment relation rather than noise. That is the expected signature of tri-planar fusion versus axial-only zonal drawing. Each row carries its own dice_vs_zones_union.

Other overlaps:

• —PI-CAI — near-certain and not programmatically excludable. PI-CAI's public training set incorporates 328 of the 346 ProstateX cases; it re-anonymized to sequential IDs and publishes no mapping, so de-duplication would need image-level matching. The organizers advise against using both.
• —Base PROSTATEx / PROSTATEx-2 — subset by construction (same image archive).
• —PROMISE12 / MSD Task05 / NCI-ISBI 2013 — no ID overlap, but all draw partly on Radboud material, as PROSTATEx does, so a shared-institution coincidence cannot be excluded by ID alone.
• —PROSTATE-DIAGNOSIS / PROSTATE-MRI — different cohorts, no overlap.
• —⚠️ Not to be confused with [ProstateZones](https://zenodo.org/records/10718469) (Zenodo, Sci Data 2024) — a different, later zonal dataset despite the near-identical name.

patient_id preserves the canonical ProstateX-#### identifier (the challenge CSVs call it ProxID) so any PROSTATEx-derived set can be cross-referenced. in_prostatex_seg_zones is true on every row.

Usage

import json, nibabel as nib
from huggingface_hub import hf_hub_download

rows = [json.loads(l) for l in open(hf_hub_download(
    "MedOtter/PROSTATEx-Seg-HiRes", "train.jsonl", repo_type="dataset"))]
r = rows[0]
img = nib.load(hf_hub_download("MedOtter/PROSTATEx-Seg-HiRes", r["image"], repo_type="dataset")).get_fdata()
msk = nib.load(hf_hub_download("MedOtter/PROSTATEx-Seg-HiRes", r["mask"],  repo_type="dataset")).get_fdata()
print(r["patient_id"], img.shape, msk.shape, msk.max())   # binary {0,1}

License

CC BY 3.0 Unported — commercial use and redistribution permitted with attribution. Confirmed from three independent sources: the LICENSE file inside the TCIA download, the TCIA REST API (LicenseName on all 66/66 series), and the collection's Citations & Data Usage Policy.

Citation

@article{meyer2021anisotropic,
  title   = {Anisotropic 3D Multi-Stream CNN for Accurate Prostate Segmentation from Multi-Planar MRI},
  author  = {Meyer, Anneke and Chlebus, Grzegorz and Rak, Marko and Schindele, Daniel and
             Schostak, Martin and van Ginneken, Bram and Schenk, Andrea and Meine, Hans and
             Hahn, Horst K. and Schreiber, Andreas and Hansen, Christian},
  journal = {Computer Methods and Programs in Biomedicine},
  volume  = {200}, pages = {105821}, year = {2021},
  doi     = {10.1016/j.cmpb.2020.105821}
}

@misc{schindele2020hires,
  title     = {High Resolution Prostate Segmentations for the ProstateX-Challenge},
  author    = {Schindele, Daniel and Meyer, Anneke and Von Reibnitz, Donata F. and
               Kiesswetter, Verena and Schostak, Martin and Rak, Marko and Hansen, Christian},
  publisher = {The Cancer Imaging Archive}, year = {2020},
  doi       = {10.7937/TCIA.2019.DEG7ZG1U}
}