The light-table test: when a scan has to hold together
Rain ticks on the skylight as a conservator lays clear plastic sheets on a light table. Each sheet has a faint grey smudge. Pressed into a stack, the smudges should line up as one solid shape. One sheet slips a touch, and the shape turns into a nervous blur.
A CT scan works the same way. It’s a 3D picture made from lots of thin slices, like those sheets. With a common, dangerous pancreas cancer, the lump can barely stand out from normal tissue. There also aren’t many carefully marked scans to learn from.
People tried making extra slices one by one, to pad out the pile. A single slice could look fine, but when you stack them, the lump jumps or changes shape between neighbours. On the light table, it fails: the stack won’t make one believable shadow.
So the team changed the job. They first made real scans easier to compare: same 3D grid, brightness tuned for soft tissue, and harsh metal streaks softened by swapping the spikes for a typical pancreas brightness from that scan. Then they cut out small, fixed-size cubes to focus on texture.
Here’s the key move: the maker stopped drawing each “sheet” alone. It looks at the whole cube, squeezes it down to catch the big shape, then builds it back up with little shortcuts that carry clues forward. Sheets are CT slices, smudges are tissue patterns, and the shortcuts are alignment marks. Takeaway: build the whole stack together so it stays coherent.
One snag: the tumour cube and the pancreas cube were made separately, so they had to be merged. Like cutting a dark stain from one stack and fitting it into another. A simple paste left a hard edge, smoothing helped, but a final blend that also nudged nearby texture made the border look like it belonged.
They checked more than single slices. They also judged the whole cube at once, like lifting the stack and turning it to see if the shadow stays steady. That 3D check used a fixed shape-and-texture finder to make a 512-number fingerprint, though it wasn’t built for medical images so it’s only an early yardstick.
When the extra cubes were used for practice, a separate checker got better at sorting tumorous from healthy, helped by a more balanced pile of examples. Limits stayed: scan prep is still heavy work, the tumour was inserted in a fixed spot, and some cubes picked up extra noise. Still, the stacks that were built to line up finally passed the light-table test.