Twenty-five TCGA colorectal diagnostic slides are low-resolution derivatives whose stated pixel spacing in the Imaging Data Commons is about half the true value

Parallax Pathology · Technical note

Russell Parrish · analysis carried out with Claude (Anthropic) · Version 2, 4 November 2026

1. Background

TCGA slides are distributed by the GDC as Aperio SVS files and by the IDC as DICOM whole-slide images [1, 2]. Most pipelines read pixel spacing from file metadata to tile slides at a fixed physical scale. A file without a spacing tag is usually dropped or assigned a default; a file with a wrong tag is tiled at the wrong physical scale without any warning. Site-related differences in TCGA slide images are well documented [3]; this note concerns one specific, checkable case.

2. Data and methods

Header census. DX slides were listed through the GDC API. TIFF headers were read by HTTP byte ranges (about 1 GB of header bytes in total): all 459 TCGA-COAD DX slides, and for TCGA-READ every DX file under 20 MB plus two per tissue source site (41 slides). For each slide we recorded dimensions, pyramid levels, compression and the Aperio ImageDescription tags (MPP, AppMag, scanner, date).

IDC. DICOM series for TCGA sites AA and AG were queried with idc-index 0.13.4 (IDC v25, the latest index it provides at the time of writing), recording PixelSpacing and TotalPixelMatrix dimensions.

Scale bearings (17 COAD derivatives).

1. Scanner-field grid (direct ruler). In both the regular Indivumed scans (MIRAX-converted SVS, tagged 0.2325 µm/px) and the derivatives, area outside the scanned region is pure white (255, 255, 255) while scanned glass is not (near 253 in regular scans, near 212 in derivatives). The outline of the scanned region therefore steps on the scanner's camera-field grid (Figure 3). We cleaned the mask (5-px opening and closing), formed edge-position profiles along each axis and took the autocorrelation maximum. Validation: on 8 regular slides at known scale, a wide search (30–160 px at a 16-times-reduced level) recovered a field of 290 × 208 µm (median); 6 of 8 slides per axis agree, the others lock onto a double field. On the derivatives (level 0, read in full) the search was bounded to lags corresponding to 0.6–1.9 µm/px. The derivative field aspect ratio (median 1.35) matches the regular one (1.39).
2. Tissue area. The scale that gives each derivative a median-sized section, against the 156 regular Indivumed sections measured on their thumbnails at known scale.
3. Structure scale. Half-width of the grey-level autocorrelation, calibrated on regular level-0 blocks reduced by known factors; held-out reductions by 5 and 7 were recovered within about ±40%.
4. Tile grid of Zenodo 3784345, an implied value rather than a measurement: derivative tiles were cut on a 200-px grid, regular Indivumed tiles on 1,101 px.

Downstream sets. File listings of Zenodo 3784345 [4] and the ZIP central directories of the CRC_DX archives of Zenodo 2530835 [5], read by HTTP range. Tile coordinates were read from locally held archives of Zenodo 3784345.

No clinical or outcome data were used.

3. Results

3.1 Three production routes in TCGA-COAD

TCGA-READ adds 8 small derivatives at site AG (Indivumed's rectal code), for 25 in total; all 8 lack MPP and magnification tags. The 17 COAD derivatives are written by Aperio Image Library v12.1.3. Frozen-section (TS and BS) slides exist for all 25 patients (51 for the COAD patients, 20 for the READ patients); the COAD ones are full-size scans tagged 0.5015 µm/px. GDC records a creation date of 2021-10-13 for 351 of the 459 COAD DX files across all three routes (the other 108 are dated 2018-05-21). GDC Data Release 31.0 (29 October 2021) states that "All TCGA slide images that were removed earlier this year have been restored" [6]; the shared date is consistent with that restoration and carries no information about the derivatives' origin. No GDC release note we read mentions these files.

3.2 IDC states 0.5 µm/px for all 25

Each of the 25 derivatives appears in IDC v25 with the same pixel matrix as the GDC file and PixelSpacing 0.0005 mm. No other AA or AG DX series in IDC carries this value; the 228 others carry 0.00023 mm. The source SVS files have no spacing tag, so the 0.5 value was assigned during or after conversion.

3.3 The true spacing is coarser than stated

The bearings differ from each other by up to about 40%, but their medians all lie between 0.9 and 1.3 µm/px, 1.8 to 2.6 times coarser than the stated 0.5, and every interquartile range excludes 0.5. Per-slide values are noisy: single slides reach 0.62 (field grid) and 0.54 (structure scale), and no bearing resolves the scale of an individual slide. The field-grid bearing is the most direct, a physical ruler present in each image, and gives the finest value; the tissue-area bearing would read too coarse if these sections are smaller than typical. A spacing of 0.93 µm/px would equal a 4-times reduction of the 0.2325 µm/px MIRAX scans; this is consistent with the field-grid median but not established.

3.4 Reach into public tile sets

Zenodo 3784345 (598 patients; "512x512 px at 0.5 µm/px" per its record [4]) contains all 25 derivative patients; two of their archives (TCGA-AA-3488, TCGA-AG-3732) are 1,008 bytes and hold no tiles. For the 15 COAD derivative patients with non-empty archives held locally, every tile origin lies within the derivative's 10,000-px frame and the tile side is 200 px, so the tiles were cut from these files and enlarged to 512 px. At 0.9–1.3 µm/px a 200-px crop spans 180–260 µm against the nominal 256 µm, giving about 0.35–0.51 µm per tile pixel instead of 0.5, and no detail finer than the source pixel. Zenodo 2530835 (360 CRC_DX patients) contains none of the 25.

3.5 Consequences observed in one analysis

In exploratory work on the Zenodo 3784345 tiles, the derivative patients separated from all others on simple image measures and on a tissue classifier. The figures below are from the authors' own analysis and are given as an example of how the files behave, not as a general estimate.

Classifier: frozen DINOv2 ViT-S/14 features with a logistic probe trained on the NCT-CRC-HE-100K-NONORM tiles [7]; mean predicted share over quadrants of 30 tiles per patient.

4. Limitations

• The field-grid bearing assumes the derivatives were scanned with the same camera-field size as the regular Indivumed slides. The matching aspect ratio supports this but does not prove it, and the derivatives' grey glass differs in tone from the regular scans'. Per-slide autocorrelation peaks are weak, and on regular slides one axis in four locked onto a double field.
• The 8 TCGA-READ derivatives were identified and checked in IDC but not scale-measured.
• We did not establish how or when the derivatives were made, or whether full-resolution originals exist.
• Only GDC release notes, papers and general web search were checked for prior reports; IDC and GDC user forums were not searched directly.

5. Practical checks for users

• The 25 slides can be identified without the list below: DX slide, site AA or AG, file under 20 MB, long edge exactly 10,000 px, no MPP tag (GDC), or PixelSpacing 0.0005 mm on an AA/AG series (IDC).
• Pipelines that drop untagged slides already exclude them from GDC data; pipelines reading IDC DICOM include them at the stated 0.5 µm/px.
• In Zenodo 3784345, their tiles carry the record's stated 0.5 µm/px.
• Frozen-section slides are available for the same patients.
• Any cohort split by TCGA site AA or AG is also split by scanner route: in TCGA-COAD, site AA is entirely MIRAX-converted or derivative.

Data and code availability

Scripts, outputs and SHA-256 hashes are held by the authors (folders TCGA_PRODUCTION_ROUTES_20261104 and DERIVATIVE_NOTE_20261104) and are available on request. Only public metadata and image pixels were used.

References

1. Genomic Data Commons Data Portal. https://portal.gdc.cancer.gov/
2. Fedorov A, et al. National Cancer Institute Imaging Data Commons: toward transparency, reproducibility, and scalability in imaging artificial intelligence. RadioGraphics 2023. https://imaging.datacommons.cancer.gov/
3. Howard FM, et al. The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nature Communications 2021;12:4423.
4. Kather JN. Image tiles of TCGA-CRC-DX histological whole slide images, non-normalized, tumor only. Zenodo 2020. doi:10.5281/zenodo.3784345
5. Kather JN. Histological images for MSI vs. MSS classification in gastrointestinal cancer, FFPE samples. Zenodo 2019. doi:10.5281/zenodo.2530835
6. GDC Data Release Notes, Data Release 31.0, 29 October 2021. https://docs.gdc.cancer.gov/Data/Release_Notes/Data_Release_Notes/
7. Kather JN, Halama N, Marx A. 100,000 histological images of human colorectal cancer and healthy tissue. Zenodo 2018. doi:10.5281/zenodo.1214456

Appendix: the 25 slides

Per-slide field-grid values are individually uncertain (Section 4); IDC states 0.5 for every row.