Organ Segmentation in 3D Scans
Organ segmentation outlines livers and kidneys voxel by voxel so surgery and dose plans measure exact volumes.
Why Does This Exist?
Dose plans and resections need milliliters, not boxes. Medical image analysis frames U-Net; this page covers the multi-organ task. Input is a CT volume; output is one mask per organ with volumes. Livers span mL while adrenals span mL, so one loss weight never fits all.
Think of It Like This
Tracing organs in a butcher diagram for weights
A butcher traces each cut on translucent paper over the carcass, then weighs the paper shapes. Big loin cuts trace fast; tiny glands need loupes. Mislabeling fat as loin ruins the price.
Masks are tracings and volumes are weights. The analogy stops at deformability: breathing moves livers while paper stays flat.
How It Actually Works
A 3D U-Net maps to masks with Dice plus cross-entropy. Dice is per organ. Small organs sample at higher rates and weight -fold. Post-processing keeps the largest component per organ where anatomy says one.
Worked example
A liver mask predicts mL against truth with overlap , so Dice is . An adrenal predicts mL against with overlap , Dice . One voxel off swings adrenal Dice but liver , which is why small-organ review stays manual.
Watch Out For
Breathing motion doubling diaphragm edges
Liver domes blur cm across phases, so masks straddle two positions. The symptom is Dice that never passes on domes. Gate to one phase or average consistent phases before measuring.
The Quick Version
- Organ masks give volumes for surgery and dose, not just locations.
- Dice near on livers but near on tiny glands is normal.
- Small organs need oversampling and higher loss weight.
- Do not trust dome edges across breathing phases.
- Keep the largest component where anatomy expects one organ.