OTLS 3D Pathology Datasets

The Liu-lab open-top light-sheet releases — 3.8 TB of raw 3D prostate biopsy volumes with recurrence outcomes on TCIA, and 191 GB of labelled 2D levels for prostate and Barrett’s oesophagus on Zenodo — the only public 3D pathology with outcomes, and a modality no instrument in this department produces.

The Liu-lab open-top light-sheet releases — 3.8 TB of raw 3D prostate biopsy volumes with recurrence outcomes on TCIA, and 191 GB of labelled 2D levels for prostate and Barrett’s oesophagus on Zenodo — the only public 3D pathology with outcomes, and a modality no instrument in this department produces.

What it is

Two public releases from the same research programme (J. T. C. Liu’s lab, University of Washington / Stanford), both cited in the data-availability statement of Gao 2026 — TRICARE (deep-learning triage of 3D pathology) and covering overlapping material:

TCIA PCA_BX_3DPATHOLOGY Zenodo 10.5281/zenodo.20052262
What Raw 3D volumes of the prostate development cohort — H&E-analogue OTLS images, synthetic cytokeratin-8 immunofluorescence, and 3D gland segmentation masks (the ITAS3D work) The 2D levels and binary labels used in the TRICARE experiments: prostate development and test cohorts, oesophagus development and test cohorts
Size / format 3.8 TB; HDF5, XML, TIFF 191 GB in 28 files (zip/7z parts plus four CSV label sheets)
Subjects 50 patients, with biochemical-recurrence outcomes Prostate: 54 UW + 29 UPenn patients; oesophagus: 24 + 5 UW patients
Access Public download, CC BY 4.0, citation required; DOI 10.7937/44MA-GX21 Open, CC BY-NC 4.0
Released with Xie et al. 2021 (Cancer Research), Glaser et al. 2019 (Nat Commun) — not filed here Gao et al. 2026 (Nat Biomed Eng)

The imaging: whole FFPE biopsies, dewaxed, bulk-stained with TO-PRO-3 and eosin, cleared in ethyl cinnamate, imaged intact on custom open-top light-sheet microscopes (0.21–0.44 µm per voxel across three instrument generations), and false-coloured to an H&E look. The Zenodo labels are normal_tissue / tumor_tissue, standing for low-grade versus GG≥2 prostate cancer and benign versus dysplasia/cancer in oesophagus — the label sheets say so explicitly.

Why it matters for my work

Honestly: as a reference point, not as data to use. The department has brightfield slide scanners; nothing here produces cleared-tissue light-sheet volumes, and the group’s tooling (QuPath Tooling, the tiling and MIL pipelines) assumes a slide. So the direct-reuse case is weak.

Three indirect reasons it is still worth a page:

  1. It is the only public 3D-pathology dataset with outcomes, and the TCIA half is CC BY. If a question about volumetric sampling — how much a single section under-represents a biopsy — ever needs a number rather than an argument, this is where it would be measured.
  2. The Zenodo release is a ready-made depth-ordered MIL benchmark: labelled levels with neighbours, patient IDs, and the paper’s own splits. A student wanting to reproduce the 2.5D aggregation-order result on Weakly Supervised Learning could do so without touching the raw volumes.
  3. The false-coloured H&E-analogue images are a stress test for any H&E-trained model. CONCH worked on them with a one-layer adapter; whether the encoders in Foundation Model Cache would is unknown and cheap to find out on a few Zenodo levels.

How it connects

Gao 2026 — TRICARE (deep-learning triage of 3D pathology) — the paper both releases serve, and the source of every cohort figure above.

TCGA — the other large public slide resource here, and the contrast: TCGA is FFPE sections on standard scanners; this is uncut blocks on a research instrument, and the two are not interchangeable for any purpose.

Weakly Supervised Learning — what the Zenodo levels are shaped for.

Foundation Models in Pathology — the modality-shift question these images pose to any H&E-pretrained encoder.

Whole Slide Imaging — the page whose substrate this modality does without.

Liu 2026 — ULTRA (ultrarapid cleared stimulated Raman 3D histology) — the other route to 3D histology filed here, and the reason these releases stay unusual. It images unfixed, unstained, unsectioned glioma by stimulated Raman scattering rather than light sheet, and its abstract shows no public data at all; the light-sheet programme is still the only source of open 3D pathology in this wiki.

Open questions

  • The TCIA collection’s 50 patients and the paper’s 54-patient development cohort do not match exactly; whether TCIA is a subset, an earlier cut, or counts differently is not stated in either source. [unverified]
  • Whether the raw oesophagus volumes exist anywhere public. Zenodo holds levels only; TCIA holds prostate only; the paper offers “additional 3D pathology datasets” from the corresponding author on request.
  • The Zenodo licence (CC BY-NC) is more restrictive than TCIA’s (CC BY) for overlapping material — recorded, not investigated.
  • Nothing here has been downloaded; file contents are as the records describe them.