What Is a DICOM Segmentation (SEG) Object?

Which voxels belong to what, in a file of its own

A DICOM Segmentation, usually written SEG, is a DICOM object that stores which voxels belong to what: one or more labeled segments — an organ, a lesion, a target volume — each referencing the frames of the source series they were derived from. The segmentation travels as its own file, so the original images are never modified.

Every segment carries coded metadata: an anatomical category and type, and how it was produced — manually, semi-automatically or automatically, usually with the name of the algorithm — which is how a viewer can tell a generated segment from a hand-drawn one.

Three storage types exist. BINARY marks each voxel as in or out, packed one bit per voxel. FRACTIONAL stores a value per voxel, expressing either the probability that the voxel belongs to the segment or the fraction of it actually covered — which is what many AI models produce, and what lets a viewer draw a graded overlay rather than a hard edge. LABELMAP packs many segments into a single image by storing the segment number in the voxel value: compact for an object carrying dozens of structures, but, unlike the other two, with no room for segments that overlap.

Lining a segmentation up with images is not only a matter of the references it declares. A SEG names the series and the frames it was derived from, but it also shares a frame of reference with them — the same patient coordinate system — so a viewer can place the masks by spatial position when those references are missing or point at another reconstruction of the same acquisition. Its grid is defined independently of the source resolution, so drawing the overlay may involve resampling.

A segmentation is not an RT Structure Set: RT stores outlines as polygons in patient coordinates, SEG stores voxels, and converting between them loses something either way. Whichever it is, it is an assertion about the images rather than a correction of them, and it is worth no more than whatever produced it — which is why the algorithm is recorded alongside the segment.

In Weasis, see DICOM SEG viewer.