Skip to content

Annotation & Labeling

Step 4 – Annotation & Labeling: telling the computer what is what

Section titled “Step 4 – Annotation & Labeling: telling the computer what is what”

Models cannot guess what is tumor and what is normal unless you tell them. In this step you bring in human knowledge: drawing regions of interest, labeling tiles or patches, or assigning slide-level or case-level labels. These labels connect pixels to ground truth diagnoses, grades, or outcomes.

Technical name: Annotation & Labeling

Tell the computer what is what on the slide:

  • Draw boxes/regions (e.g., tumor, normal, necrosis).
  • Mark structures (glands, glomeruli, follicles).
  • Mark cells/events (mitoses, plasma cells, eosinophils).
  • “Can we mark all tumor area for training?”
  • “Can residents practice by labeling easier vs harder regions?”
  • “Can I export these outlines as a mask?”
  • ROI drawing for tumor, margin, necrosis.
  • Point annotations for individual cells/structures.
  • Assign labels (grade, pattern type) to regions/tiles.
  • Export polygons, masks, or labeled points.
  • QuPath — popular for region and cell‑level annotations; exports masks/labels.
  • ASAP — viewer plus manual annotation.
  • SlideRunner — focuses on point/cell annotation (e.g., mitoses, lymphocytes).
  • Cytomine / TissUUmaps — web platforms for collaborative annotation.

Like drawing on the slide with a pen and dictating notes—but in a form a computer can consume and learn from.

  • LAB-01/02: GeoJSON to masks — export QuPath annotations to GeoJSON, load them with Shapely, run point-in-polygon checks, and rasterize polygons into masks for AI training.