Liver segmentation on CT means drawing the organ’s boundary — a mask of the liver, not the eight Couinaud segments inside it. Manual contouring on every slice is slow and operator-dependent. Automated methods (thresholding, region growing, CNNs) do that contour so a volume, a radiation plan, or a resection remnant has a number attached to it.
If you meant Couinaud segments I–VIII on CT → liver segments on CT. If you meant how to train a U-Net on Task03 → automatic liver segmentation — Part 1. Generic methods (any organ) → medical image segmentation. nnU-Net → nnU-Net for medical image segmentation.
What the mask is
A liver mask is a voxel-wise label: liver / not-liver (sometimes liver / tumor / background). It is the input to volumetry, to a dose grid, and to a 3D model. It is not “this lesion is in segment VI.” That assignment uses hepatic and portal veins. Keep the two jobs apart.
Manual contouring on MDCT is still the reference in many trials. It does not scale. Two readers will disagree at the dome, the bare area, and next to a fat-infiltrated edge. Automated contouring exists to make that contour repeatable, not to invent a new anatomy.
MDCT made the mask usable
Early single-slice CT gave you a handful of thick cuts. You could see the liver. You could not build a honest volume. Multi-detector CT (MDCT) acquires a continuous slab: many rows per rotation, near-isotropic voxels, a breath-hold abdomen. That volume is what a segmentation algorithm actually eats. The scan mode is volumetric CT. This page starts after the volume exists — we are talking about the contour, not the gantry.
Contrast still matters. Iodine in the late arterial phase lights hypervascular lesions; the portal-venous phase lights parenchyma and the organ edge against kidney, spleen, and bowel. A mask trained only on non-contrast will leak. Pharmacokinetics (rate, volume, kidney function) change that edge. That is a protocol choice, not a Couinaud lecture.
How the contour is drawn
- Thresholding. Liver HU vs fat / air. Fast. Fails where the liver touches the heart or a contrast-filled vessel of the same density.
- Region growing. A seed inside the liver expands until the edge. Needs a clean seed and a stop rule. Leaks through a thin fissure.
- Classical ML. Hand-built features, a classifier per voxel. Better than a raw threshold. Still brittle across scanners.
- CNNs / U-Net. Learned from labeled volumes. The current default for an organ mask. Does not magically output Couinaud I–VIII unless you trained it for that.
Dice, Hausdorff, and volume error are how you score a mask against a reader. We are not publishing a fake “Moderate / Good / High” table. If you want numbers, run the labels. The MONAI series is the place we actually trained one: Part 1.
Where the mask is used
- Resection / transplant volumetry. Whole-liver cc, then (if someone also labeled Couinaud) a remnant. The remnant arithmetic is on liver segments on CT. The blob that gives you “liver cc” is this page.
- Radiation therapy. A planning contour so the dose misses spare parenchyma.
- Interventional radiology. A 3D liver so an ablation or TACE path has an organ, not a guess.
Photon-counting and dual-energy change tissue contrast; they do not replace the need for a mask. Multimodal (CT + MR) is a registration problem first. None of that is a reason to paste a Couinaud table here.
PYCAD ships organ-mask models and the viewers that display them. The how-to for this liver (MONAI, Task03, the errors, the clone) is the four-part series. Case studies.