How to specify a lung-nodule CT dataset

Population and clinical pathway

Screening cohorts and incidental-nodule cohorts differ in prevalence, patient selection, acquisition and follow-up. The request should state whether nodules arise from low-dose screening, routine chest CT, oncology imaging or another pathway and whether the intended model must generalize across them.

Nodule size and cohort composition

Define measurement convention and clinically relevant size categories rather than requesting small, medium and large nodules without thresholds. Controls should be size- and acquisition-aware so malignancy performance is not driven by obvious size differences.

• Positive cases by nodule-size band and morphology
• Benign, malignant and indeterminate categories
• Size-matched benign and negative controls
• Single versus multiple nodules and prior-examination rules

Ground-truth hierarchy

Pathology-confirmed malignancy provides strong positive evidence when available. Benign truth may use pathology or a defined period of longitudinal stability. Report text alone may support discovery or weak labels but should not silently replace the required reference standard.

Acquisition, diversity and annotations

Specify contrast status, dose, slice thickness, reconstruction, field of view and required prior studies. Site and geographic diversity should be expressed as measurable requirements. Annotation scope should define nodule localization, segmentation or measurements, reviewer qualifications and adjudication.

Acceptance criteria

A pilot should be checked against patient-level counts, size and pathology distributions, technical constraints, source-to-label traceability, privacy, exclusions and delivery format before the final cohort is approved.