Code cannot create clinical truth by itself
Software can organize reports, propose labels and manage annotation files. It cannot independently determine that a tumour, stroke or fibrosis stage is correct. Clinical claims require a defined reference source and, where necessary, qualified clinician review.
Choose the evidence for the task
A model intended to segment anatomy needs spatial annotations. A study-level classifier may need only adjudicated labels. A quantitative biomarker may require measurements, laboratory evidence, pathology or outcomes. The annotation plan follows the intended use rather than adding masks to every case.
- Label definition and inclusion/exclusion rules
- Annotator qualification and blinding
- Single-reader, double-reader or adjudicated workflow
- Inter-reader agreement and difficult-case review
- Versioned annotation format and source-image mapping
Missing and uncertain labels
Absence of a diagnosis in a report does not automatically mean a normal negative. Unknown, unassessed and technically limited cases are represented explicitly so they are not mislabeled as controls.
Questions, answered directly.
Can labels be generated automatically?+
Automated extraction can support triage, but clinically meaningful labels still need a validation standard appropriate to the task.
Are reports equivalent to segmentations?+
No. Reports provide narrative or study-level evidence; segmentations provide spatial boundaries.
Do all cases need two readers?+
Not always. Reader count and adjudication should reflect task risk, label ambiguity and buyer acceptance criteria.