Prostate-Anatomical-Edge-Cases

Stress-Testing Pelvic Autosegmentation Algorithms Using Anatomical Edge Cases — a TCIA collection of pelvic radiotherapy planning CT with manually contoured organs at risk, curated so that most cases contain anatomy known to break autosegmentation algorithms (Kanwar et al., Phys Imaging Radiat Oncol 2023).

Read before using — the name is misleading in two ways:

1. This is CT, not MRI. Despite "Prostate" in the name it is not a prostate mpMRI/zonal dataset: there is no T2W/DWI/ADC, no peripheral or transition zone, no urethra, and no lesion labels. It is RT simulation CT with pelvic organ-at-risk contours, and 4 of the 5 classes are not the prostate.
2. "Edge cases" means difficult anatomy, not difficult lesions. The cohort was selected for hip arthroplasty hardware, prostatic median-lobe hypertrophy, "droopy" seminal vesicles, Foley catheters, SpaceOAR hydrogel, brachytherapy seeds/calcifications, extracapsular surface irregularity, narrow rectum, in-field bowel and morbid obesity — features that degrade atlas-, model- and deep-learning-based contouring.

Dataset Details

Field	Value
Modality	CT (RT simulation) + DICOM RTSTRUCT contours
Body part	Pelvis — prostate and surrounding organs at risk
Task	3D multi-organ segmentation
Patients	131 (112 anatomical edge cases + 19 normal controls)
Series	262 = 131 CT + 131 RTSTRUCT (1:1 paired)
CT slices	23,359 (min 86 / median 167 / max 379 per case)
Scanner	Philips Brilliance Big Bore (100%); RTSTRUCT exported from Varian ARIA RTM
Format	DICOM (CT) + DICOM RTSTRUCT (contours)
Size	~17 GB
License	CC BY 4.0
DOI	10.7937/QSTF-ST65
Source	Oregon Health & Science University (single institution)

There is no train/val/test split — this is a stress-test cohort, not a challenge. Splitting is left to the consumer.

Classes

Five foreground structures, present in the StructureSetROISequence of every case, named per TG-263 and all typed ORGAN:

Label	ROI name
0	background
1	Prostate
2	Rectum
3	Bladder
4	Femur_Head_L
5	Femur_Head_R

Ground Truth — single manual tier

Every ROI in every case carries ROIGenerationAlgorithm = MANUAL. The contours were drawn by a single radiation oncologist, peer-reviewed, and used clinically for treatment planning — so there is exactly one annotation tier and no rater ambiguity.

The three autosegmentation outputs benchmarked in the paper (atlas-based/MIM, model-based/RayStation, deep-learning U-Net/RayStation v9B) were not deposited; all 262 series report ThirdPartyAnalysis = NO. The RTSTRUCTs here are the gold standard, not algorithm output.

Important Notes for Loaders

These were verified by parsing all 131 RTSTRUCT objects and are easy to get wrong:

• 15 of 131 cases declare an ROI with zero contours. The ROI appears in StructureSetROISequence but its ContourSequence is empty, so naive code silently emits an all-background mask for that class instead of skipping it. Only 116/131 have all five structures non-empty:

  Empty structure	n	Patient IDs (Prostate-AEC-…)
  Rectum	7	009, 116, 117, 119, 125, 127, 134
  Femur_Head_L	4	020, 037, 106, 111
  Femur_Head_R	3	010, 012, 027
  Prostate	1	072
  Bladder	0	— (always present)

• Prostate-AEC-101 carries a sixth ROI, BODY (external contour). Build masks by matching ROI name, never by ROI index or order.

• Patient IDs are non-contiguous: Prostate-AEC-001 … Prostate-AEC-134 with 021, 032 and 043 absent → 131 patients.

• ApprovalStatus = UNAPPROVED on all 131. This is a de-identification re-export artifact and contradicts the paper's "clinically approved" statement — do not use it as a quality filter.

• Prostate-AEC-072 is doubly suspect: empty Prostate plus a Rectum whose contours integrate to an anatomically implausible volume. Consider excluding it.

• Dates are shifted (StudyDate reads 1992–2003 for a 2011–2019 cohort) and PatientIdentityRemoved = YES.

• Rasterisation: each RTSTRUCT references exactly one CT series via ReferencedFrameOfReferenceSequence, so contour→labelmap conversion with rt_utils (or manual polygon filling against the CT ImagePositionPatient grid) is unambiguous.

Cohort labels (edge case vs control) are NOT in this release

The paper's Data Availability statement says "edge case labels and basic demographic data have been deposited on TCIA", but they were not. TCIA exposes only the manifest and the NBIA digest; the RTSTRUCT headers carry no edge-case tag, no ClinicalTrialSubjectID and no PatientComments.

The 112-vs-19 edge/control assignment and the 8 anatomical-variant classes (prostate hypertrophy 52, droopy seminal vesicles 37, hip arthroplasty 11, surface irregularity 9, calcifications/seeds 8, Foley catheter 4, SpaceOAR 4, narrow rectum 2, in-field bowel 1, morbid obesity 1 — overlapping, so multi-label) exist only in the paper's Supplementary Table 2. That table also numbers cases 1–131 contiguously while TCIA uses 001–134 with gaps, so the join is not the identity map. This mirror deliberately ships only what TCIA publishes rather than a derived mapping.

Cross-dataset Overlap

None known. Every widely-used prostate segmentation set is MRI from a different institution, whereas this is CT from OHSU: PROMISE12, PI-CAI, QIN-PROSTATE, Prostate-3T, PROSTATEx, NCI-ISBI 2013, MSD Task05_Prostate, ProstateX-Seg-HiRes, Prostate-MRI-US-Biopsy and Prostate Fused-MRI-Pathology all share neither modality nor cohort. TCIA reports "No related Collections found", and there is no cross-reference ID field — Prostate-AEC-### IDs are collection-local.

Structure

images/<PatientID>/<SeriesInstanceUID>/*.dcm          # CT (131 series, 23,359 slices)
segmentations/<PatientID>/<SeriesInstanceUID>/*.dcm    # RTSTRUCT (131 objects, 1 per patient)
series_to_patient.json                                 # series-level metadata (all 262)
LICENSE.txt

series_to_patient.json keys each SeriesInstanceUID to: PatientID, Collection, StudyInstanceUID, Modality, SeriesDescription, SeriesNumber, BodyPartExamined, Manufacturer, ManufacturerModelName, ImageCount, FileSize, License, DOI, ThirdPartyAnalysis, and the relative path — so CT↔RTSTRUCT pairing needs no TCIA round-trip.

Source & Citation

• TCIA collection: https://www.cancerimagingarchive.net/collection/prostate-anatomical-edge-cases/
• DOI: 10.7937/QSTF-ST65
• Official, author-deposited (Thompson, Kanwar, Merz, Cohen, Fisher, Rana, Claunch, Hung); fully public, no registration required.

@article{kanwar2023edgecases,
  author  = {Kanwar, Aasheesh and Merz, Brandon and Claunch, Cheryl and
             Rana, Shushan and Hung, Arthur and Thompson, Reid F.},
  title   = {Stress-testing pelvic autosegmentation algorithms using anatomical
             edge cases},
  journal = {Physics and Imaging in Radiation Oncology},
  volume  = {25},
  pages   = {100413},
  year    = {2023},
  doi     = {10.1016/j.phro.2023.100413}
}

@misc{thompson2023paec,
  author    = {Thompson, R. F. and Kanwar, A. and Merz, B. and Cohen, E. and
               Fisher, H. and Rana, S. and Claunch, C. and Hung, A.},
  title     = {Stress-Testing Pelvic Autosegmentation Algorithms Using Anatomical
               Edge Cases [Data set]},
  year      = {2023},
  publisher = {The Cancer Imaging Archive},
  doi       = {10.7937/QSTF-ST65}
}

@article{clark2013tcia,
  author  = {Clark, Kenneth and Vendt, Bruce and Smith, Kirk and others},
  title   = {The Cancer Imaging Archive (TCIA): Maintaining and Operating a
             Public Information Repository},
  journal = {Journal of Digital Imaging},
  volume  = {26},
  number  = {6},
  pages   = {1045--1057},
  year    = {2013},
  doi     = {10.1007/s10278-013-9622-7}
}