Possible one-row misalignment in metadata.xlsx structured reports from PanTS_00002294

#4
by kevin-wang-usyd - opened

Thank you for releasing PanTS. An audit of the structured report column against the supplied segmentation masks found a systematic case/report mismatch in the following metadata snapshot. Could you help confirm the intended correspondence and whether a corrected version is available?

File and scope

  • Dataset: BodyMaps/PanTSMini
  • File: metadata.xlsx at revision 3b1cd611, 1,334,813 bytes
  • SHA-256: 4bcbf14a31b1ca9441af051a104702e7a832391f47d3af5f493048ec813283e3 (matches the official LFS oid)
  • Mask source: the official PanTSMini_Label.tar.gz
  • The comparisons below use report IDs 1–8999 and candidate mask IDs within 1–9000. Here, “report ID” means the case ID associated with the spreadsheet entry, not its physical Excel row number.

Observed pattern

Two separately implemented checks—local boundary sampling and a fresh audit of the official training masks—support the following dominant pattern:

  • Reports stored under IDs 1–2293 correspond to the same-ID masks.
  • Reports stored under IDs 2294–8999 correspond to the next-ID masks.
  • For example, the report stored under PanTS_00002294 appears to describe PanTS_00002295.

Some boundary examples, using spleen volume in cc:

Case ID holding the report Reported volume Same-ID mask Next-ID mask
PanTS_00002290 110.2 110.223 616.704
PanTS_00002294 188.4 98.988 189.260
PanTS_00002296 413.1 238.822 413.313

The full-range volume comparison gives:

Report IDs Reports with comparable organ volumes Same-ID agreement Next-ID agreement
1–2293 2142 2132 / 2142 (99.53%) —
2294–8999 6256 80 / 6256 (1.28%) 6237 / 6256 (99.70%)

Volumes were calculated from decoded foreground voxel counts and voxel volume, accounting for NIfTI scaling. The comparison uses spleen, liver, pancreas, and left/right kidney volumes where reported. Spatial units unspecified in the NIfTI header were interpreted as millimetres.

“Agreement” requires every comparable organ in a report to satisfy:

abs(report - mask) <= max(0.050001 cc, 0.05 * max(report, mask))

The 601 reports without comparable organ volumes are excluded from the denominator, not counted as passing. Under the piecewise correspondence above, 8369/8398 comparable reports pass this criterion; 29 retain discrepancies. Only 1262/8398 pass the stricter 0.050001 cc rounding-only criterion, so this is not a claim of exact numerical identity.

Lesion presence provides additional support: of the 864 reports (IDs 1–8999) that list a pancreatic lesion, 578 have an empty same-ID lesion mask. Under the piecewise correspondence above—same ID for reports 1–2293 and next ID for reports 2294–8999—all 864 correspond to a non-empty lesion mask.

Boundary cases and clarification requested

This is a supported candidate correspondence rule, not proof that every report has been uniquely identified or that shifting entries fully corrects their contents. In particular:

  • A reliable report source for mask PanTS_00002294 has not been established. Assigning report 2293 to it would not be justified by this pattern.
  • The source of the report stored under PanTS_00009000 has not been verified; it is excluded from the figures above.
  • At the worksheet-structure level, the final report cell (O9902, associated with PanTS_00009901) is absent. This is compatible with a missing entry but does not establish where or how it was lost; it also does not verify the correspondence of the intervening test-range reports.

Could you please clarify:

  1. Is this a known alignment issue, or is there a version/preprocessing difference that explains the pattern?
  2. Is an authoritative case-to-report mapping or corrected metadata.xlsx available, including guidance for PanTS_00002294?
  3. If confirmed, could a versioned correction or dataset-card notice identify the affected snapshot and recommended handling?
  4. The comparisons presented here do not establish report-to-mask correspondence for the official test range (PanTS_00009001–PanTS_00009901). Could you check whether those reports are affected as well?

Separately, the audit found a tumor? = 1 / empty-lesion-mask inconsistency and some reported lesion sizes exceeding the corresponding mask's 3D bounding-box diagonal even under the candidate correspondence. These can be documented separately; we have not established whether they share a cause with the alignment issue.

I can provide a compact comparison CSV and the audit scripts if useful. Thank you for taking a look.

Sign up or log in to comment