Inspected packages/vxdata-schemas generated Create models and payload definitions for DataSource, Patient, PathologySpecimen, PathologyAssessment, HistoScan, HistoMap, PatientStatus, and Measurement.
Compared recent histopathology integrations including PANDA, ACROBAT, BRACS, and Frankfurt to identify current parentage, identifiers, provenance, and patient-registration conventions.
Inspected PESO Zenodo record 5137717 and its test mapping, plus the LEOPARD challenge data description and public training labels.
Verified a public LEOPARD tissue TIFF uses binary values 0 and 1.
Reported exact constructors and enums without editing the monorepo. Flagged that no typed censored time-to-event outcome resource currently exists, and that PatientStatus other_data is only a temporary lossless fallback for LEOPARD biochemical-recurrence follow-up.
Lessons Learned: Pitfalls
The existing histo workstream SOP index does not describe dataset-ingestion schema choices; the related infrastructure vxData SOP would have been the more direct starting point.
Prostatectomy WSI does not establish block_format=WHOLE_MOUNT; source descriptions that say several physical slides are packed into one TIFF make that inference especially unsafe.
PESO’s cancer/non-cancer labels apply to regions, not specimen-level clinical significance, so they must not be mapped to PathologyAssessment.is_cspca.
Lessons Learned: Improvements
Add a concise current-schema table to the vxData ingestion SOP, including exact Create class names, required fields, and enum casing.
Document a standard approach for pseudo-patients when public histology datasets expose one slide/case identity but no patient mapping.
Add a dedicated typed censored time-to-event outcome payload with endpoint name, event indicator, duration, unit, and censoring semantics before integrating survival/challenge datasets such as LEOPARD.
