Skip to content

Sampling and reproducibility

A tissue tile is an image region selected after tissue detection. TumorQuantAI offers:

  • smoke: a seeded 1% subset for one first slide;
  • fast: a seeded 10% subset by default; and
  • full: 100% of detected tissue tiles.

Sampling is deterministic when the input fingerprint, tissue-selection settings, percentage, and random seed are unchanged. The summary records the seed plus sampled/total tile counts.

Deterministic does not mean representative. A 1% or 10% subset can miss spatial heterogeneity. Visual QC, scientific design, and sensitivity analyses remain the researcher's responsibility.

Raw fast-mode counts are detections in sampled tiles. They are not validated whole-slide estimates and TumorQuantAI does not multiply them by 100 / percent_slide. Keep fast and full in separate outputs and do not compare raw counts across different sampled areas as though their denominators match.

Resume reuses a cached task only when relevant inputs and settings still match. Changing sampling, seed, source MPP, model identity, or a source fingerprint should invalidate affected work.

Next: understand counts versus fractions.