Segmentation as Measurement Infrastructure

A mask defines the region being measured

Segmentation is often introduced as one prediction task among many. In a region-based quantitative workflow it has a more basic role: the mask defines which samples count as the region of interest. A volume, mean attenuation, diameter, or texture summary is not calculated over a tumour in the abstract; it is calculated over the voxels selected by an operational region definition. The mask is therefore an input to the measurement, not merely a result displayed beside it.

This does not mean that every image model needs a segmentation. It means that whenever a downstream quantity depends on a region, the provenance and geometry of that region are part of the quantity’s meaning. A boundary shift can add or remove voxels from an intensity summary, change the physical volume, and change which voxel pairs enter a texture calculation. The direction and magnitude of the change depend on the boundary, the image, the grid, and the feature; the important point is that the dependency is real even when the mask error is hard to see in a viewer.

The mask becomes part of a longer measurement chain

The chain can be written compactly as image -> mask -> measurement -> phenotype -> prediction. Each arrow adds an opportunity for a different question to be answered. The segmentation model may be evaluated against an annotation, the measurement may be evaluated for agreement or stability, and a prediction model may be evaluated for discrimination, calibration, or utility. A strong result at one arrow does not certify the next arrow.

Automation makes this distinction more consequential, not less. When masks are expensive, a study may limit itself to a small number of carefully reviewed regions. When an automated system makes many masks available, the same mask factory can feed volumetry, radiomics, registration, and prediction. The scale of downstream use increases the value of a reliable mask, but it also creates more opportunities for an unnoticed boundary or reference error to propagate. This is a qualitative task-dependent risk, not a claim that errors must increase or that one ordering of risks is universal. Visual inspection can expose an anatomically implausible contour, while quantitative checks can expose a systematic or small boundary difference. Neither one replaces the other, and even together they do not prove downstream validity without evaluating the declared downstream measurement and its uncertainty.

The use of the mask determines the evaluation question

For a mask used only to display an anatomical structure, overlap and surface agreement may answer much of the immediate question. For a mask used to compute a volume, the implied volume and its agreement with the intended reference also matter. For a mask used to calculate a texture feature, the feature’s stability under plausible boundary variation and its measurement definition matter. For a mask used as input to a classifier, the complete downstream procedure needs its own validation boundary.

The next section develops the first part of that evaluation without reducing it to one score. Dice is useful as an overlap summary, but its meaning depends on structure size and it does not describe every spatial or downstream error.