Skip to content
Imaging Evidence Path

Segmentation as infrastructure

Tonight · ~15 min · read · energy: low · setup: none

Segmentation looks like one AI task among many. In quantitative imaging it is more than that: it is the infrastructure almost every other task stands on. This short lesson explains why so much rides on the mask, connecting Chapter 2 (what a mask is) and Chapter 3 (what the mask enables) to the AI that produces it.

The anchor: measurements need a region

A tumour volume, a mean attenuation, a texture feature, a diameter — every one of them needs a region (the ROI) to be measured over. That region is a segmentation. So the segmentation is not a deliverable sitting beside the measurement; it is the input that defines the measurement. A wrong mask makes every downstream number wrong, silently — which is why segmentation quality propagates into radiomics, volumetry, and any image-based prediction.

flowchart LR
    IMG["image"] --> SEG["segmentation model"]
    SEG --> MASK["mask (ROI)"]
    MASK --> V["volume / diameter"]
    MASK --> R["radiomic features"]
    MASK --> P["prediction / prognosis"]
    MASK --> Q["quantitative phenotype"]

Why segmentation got central

Several forces pushed segmentation to the centre of imaging AI:

  • Measurement chains depend on it. The radiomics pipeline (Ch. 3) is segmentation → measurement → phenotype → prediction. The whole chain inherits the mask’s errors.
  • Tools like TotalSegmentator made masks cheap. A model that segments ~100+ structures in one pass turns “produce a mask” from a manual bottleneck into an automatic step — which makes the mask the substrate everything else reads from.
  • Errors propagate invisibly. Because downstream features compress the mask, a boundary error that is hard to see by eye becomes a biased volume or a shifted texture feature that is impossible to trace back without Visual QC (Ch. 2).

The consequence for how you judge a segmentation

Because the mask is infrastructure, judging it by one overlap number is exactly wrong. The right question is the downstream one: does the tumour volume the mask implies agree with a human’s? Does the texture feature survive a one-voxel boundary perturbation (Ch. 3)? That is the subject of the next lesson — why Dice is not enough — and it is why TRACE-CT’s Visual QC (Ch. 2) checks masks against source anatomy, not just against an overlap score.

Stop and think — then reveal

A one-voxel boundary error changes a tumour’s mean attenuation by a tiny amount and its Dice by almost nothing. Which downstream number is still corrupted, and why does Dice fail to catch it?

The volume (and any volume-derived measurement) is corrupted, because moving a boundary voxel in or out of the ROI changes Σ voxels — and on a small tumour a one-voxel shell is a meaningful fraction of the volume. Dice barely moves because Dice is a gross-overlap summary: on a tumour where almost all voxels already agree, a one-voxel boundary slip changes the intersection/sum ratio by a whisker. So the very property that makes Dice look reassuring (insensitivity to small boundary slips) is the property that lets a real volumetry error hide behind a “good” Dice. This is the bridge to why Dice is not enough.

What to retain

  1. Segmentation is infrastructure: the mask defines the region every measurement is taken over, so its errors propagate into every downstream number.
  2. The radiomics chain (segmentation → measurement → phenotype → prediction) inherits the mask’s quality.
  3. Tools like TotalSegmentator made masks cheap, which made them central — and made mask-quality errors more consequential, not less.
  4. Judge a mask by its downstream effect (volume, feature stability), not by one overlap number — the subject of the next lesson.

Next: why Dice is not enough to judge that infrastructure.