When a rough box finally starts telling the truth
In a dim repair room, a crushed paper bird lantern is lifted from a box. The box shows where the bird used to sit, but not its thin neck, curved wings, or hollow ribs. A scan with only a box round an organ gives the same sort of clue: place, not shape. That is the whole problem.
If the repairer trusts the box, the bird turns into a puffy lump on the table. Fine edges disappear, empty spaces get filled, and bits of packing cloth can pass for paper in the same flat grey light. Box-only organ finding stumbles in much the same way, especially in small scan cut-outs.
So the first new help is a shape memory. The software keeps a bead-like version of an organ, with the outside and the inner supports, then pulls its rough guess closer to that form. It only does this when the scan chunk seems to hold the whole organ. No point judging half a bird against a full one.
The second new help is about likeness, not darkness. It starts with rough maybes inside the box and rough noes outside, then learns which nearby patches truly belong together. In the lantern room, that is like sorting paper by weave and stiffness, not by how dark it looks under one weak bulb.
At first those two hints seem different, but they back each other up. On kidney scans, the full set-up did clearly better than using only one hint. Take away the likeness cue and the fit drops. Take away the shape cue and it falls much further. Even the inner structure of the template matters.
That is the useful bit. People can still give quick box marks instead of tracing every edge, but the result gets much closer to the real outline in organs like the liver, kidney, and hippocampus. The box has not become clever on its own. It just stopped working alone.