Datasets:
license: cc-by-3.0
task_categories:
- image-segmentation
tags:
- medical
- mri
- t2-weighted
- prostate
- zonal-anatomy
- pi-rads
- tcia
- dicom-seg
pretty_name: PROSTATEx-Seg-Zones
size_categories:
- n<1K
dataset_info:
features:
- name: patient_id
dtype: string
- name: split
dtype: string
- name: image
dtype: image
- name: mask
dtype: image
- name: overlay
dtype: image
- name: preview_slice
dtype: int32
- name: num_slices
dtype: int32
- name: rows
dtype: int32
- name: cols
dtype: int32
- name: spacing_xyz
list: float32
- name: zones_present
list: string
- name: voxels_per_zone_json
dtype: string
splits:
- name: train
num_bytes: 27847888
num_examples: 98
download_size: 27850346
dataset_size: 27847888
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
PROSTATEx-Seg-Zones
Prostatic zonal anatomy segmentation on axial T2-weighted MRI. 98 patients, each with a four-class label map covering the peripheral zone, transition zone, prostatic urethra, and anterior fibromuscular stroma — the compartments a PI-RADS v2 sector map is built on.
The upstream TCIA product is an analysis result containing DICOM-SEG objects and no images. This mirror pairs each segmentation with its source T2 series from the base SPIE-AAPM-NCI PROSTATEx collection and ships both as NIfTI, so the merge does not have to be repeated downstream.
Dataset Details
| Field | Value |
|---|---|
| Modality | MRI, axial T2-weighted turbo spin-echo (t2_tse_tra) |
| Body part | Prostate |
| Task | Multi-class 3D segmentation (zonal anatomy) |
| Patients | 98 (one study, one T2 series, one segmentation each) |
| Classes | 4 + background, mutually disjoint |
| In-plane grid | 384x384 (82 cases), 320x320 (15), 640x640 (1) |
| Slices per volume | 19-25 |
| In-plane spacing | 0.5-0.5625 mm |
| Slice thickness | 3 mm (4.5 mm spacing typical) |
| Annotator | Single reader (ContentCreator=Reader1), reviewed by an expert urologist |
| Format | NIfTI (.nii.gz), converted from DICOM + DICOM-SEG |
| License | CC BY 3.0 Unported (commercial use permitted) |
Labels
| Value | Zone | SNOMED CT |
|---|---|---|
| 0 | background | — |
| 1 | Peripheral zone | 279706003 |
| 2 | Transition zone | 399384005 |
| 3 | Prostatic urethra | 71553001 |
| 4 | Anterior fibromuscular stroma | 717025007 |
There is no central zone — CZ is folded into the transition zone, consistent with PI-RADS v2 practice.
The four zones are strictly disjoint: verified across all 98 cases that no
voxel carries two labels and sum(per-zone counts) == |foreground|. Consume this
as a single label map; per-target binary fan-out is not needed and would be wrong.
All four zones are non-empty in all 98 cases. They are, however, very unbalanced — the urethra is ~2% and the anterior fibromuscular stroma ~3% of total foreground:
| Zone | Total voxels | Share of foreground |
|---|---|---|
| Transition zone | 5,532,407 | 65.2% |
| Peripheral zone | 2,508,037 | 29.6% |
| Anterior fibromuscular stroma | 261,059 | 3.1% |
| Prostatic urethra | 178,892 | 2.1% |
Expect low Dice on the last two. That is the class prior, not a defect.
Layout
volumes/ProstateX-0004/image.nii.gz # T2W volume (z, y, x)
volumes/ProstateX-0004/mask.nii.gz # uint8 label map, identical geometry
...
metadata.jsonl # one record per case
label_map.json # label -> name + SNOMED code
mask.nii.gz carries the same origin, spacing and direction as its image.nii.gz
(set via CopyInformation), so the pair is voxel-aligned with no resampling.
Each metadata.jsonl record holds patient_id, image, mask, split,
num_slices, rows, cols, spacing_xyz, voxels_per_zone, zones_present,
and the originating t2_series_uid / study_uid / seg_series_uid.
Splits
There is no official train/val/test split. All 98 cases are published as a
flat pool and are labelled train in metadata.jsonl. Any split is your own
construction — group on patient_id.
Provenance and integrity notes
Source images are a separate download. The Zone Segmentations product (149 MB) contains segmentations only. The corresponding 2,020 T2 DICOM (585 MB) come from the base PROSTATEx collection via TCIA's "Corresponding Original files" manifest. Both are CC BY 3.0.
T2 series were resolved via ReferencedSeriesSequence, never by
SeriesDescription. All 98 SEGs reference a series described t2_tse_tra, but
14 of the 98 studies contain two MR series carrying that identical description —
matching on the description silently selects the wrong volume in ~14% of cases.
Slice placement is by source SOP Instance UID. Each SEG frame was placed at
the index of its DerivationImageSequence source instance within the geometrically
sorted T2 stack. DICOM-SEG frame order is not anatomical order, and
ReferencedInstanceSequence is not guaranteed sorted.
Relationship to PROSTATEx-Seg-HiRes. A second TCIA analysis result
(10.7937/TCIA.2019.DEG7ZG1U) provides 66 cases of whole-gland binary
segmentation built by fusing axial/sagittal/coronal delineations. Those 66
patients are a strict subset of these 98, but the two are different products, not
two tiers of one annotation — HiRes has no zone labels, and its masks are not
voxel-aligned to any TCIA series (e.g. ProstateX-0020: 116 planes at 0.6 mm versus
the T2 series' 21 at 3 mm). For zonal work, this dataset is the reference.
⚠️ Patient overlap with PI-CAI
All 98 patients come from the SPIE-AAPM-NCI PROSTATEx collection (346 patients). The PI-CAI public training set incorporates 328 ProstateX cases, so effectively every patient here also appears in PI-CAI, and the PI-CAI organisers explicitly recommend not using both together.
This overlap is not programmatically excludable: PI-CAI re-anonymised to sequential identifiers and publishes no ProstateX↔PI-CAI mapping. De-duplication would require image-level matching (T2 geometry or voxel hashing).
The label content does not collide — PI-CAI annotates csPCa lesions and whole
gland, this dataset annotates zonal anatomy — so the two are complementary in
what they measure. But do not treat them as independent cohorts when aggregating
patient-level statistics. The patient_id field preserves the canonical
ProstateX-#### identifier for cross-referencing against any PROSTATEx-derived
set.
Other prostate datasets checked: PROSTATEx-2 shares this image archive (different task). Prostate158, PROMISE12, MSD Task05_Prostate, NCI-ISBI 2013, QIN-PROSTATE-Repeatability, and Prostate-MRI-US-Biopsy draw on different TCIA collections or institutions with no documented patient overlap.
Citation
@inproceedings{meyer2019prostatezones,
title = {Towards Patient-Individual {PI-RADS} v2 Sector Map: {CNN} for
Automatic Segmentation of Prostatic Zones from {T2}-Weighted {MRI}},
author = {Meyer, Anneke and Rak, Marko and Schindele, Daniel and
Blaschke, Simon and Schostak, Martin and Fedorov, Andriy and
Hansen, Christian},
booktitle = {IEEE International Symposium on Biomedical Imaging (ISBI)},
pages = {696--700},
year = {2019},
doi = {10.1109/ISBI.2019.8759572}
}
@misc{meyer2020prostatexzones,
title = {{PROSTATEx} Zone Segmentations},
author = {Meyer, Anneke and Schindele, Daniel and von Reibnitz, Daniel and
Rak, Marko and Schostak, Martin and Hansen, Christian},
year = {2020},
publisher = {The Cancer Imaging Archive},
doi = {10.7937/TCIA.NBB4-4655}
}
@misc{litjens2017prostatex,
title = {{SPIE-AAPM-NCI PROSTATEx} Challenges ({PROSTATEx})},
author = {Litjens, Geert and Debats, Oscar and Barentsz, Jelle and
Karssemeijer, Nico and Huisman, Henkjan},
year = {2017},
publisher = {The Cancer Imaging Archive},
doi = {10.7937/K9TCIA.2017.MURS5CL}
}
@article{clark2013tcia,
title = {The Cancer Imaging Archive ({TCIA}): Maintaining and Operating a
Public Information Repository},
author = {Clark, Kenneth and Vendt, Bruce and Smith, Kirk and others},
journal = {Journal of Digital Imaging},
volume = {26},
number = {6},
pages = {1045--1057},
year = {2013},
doi = {10.1007/s10278-013-9622-7}
}