Ilan LenznerAI developer · design engineer Work in progress
Research prototype · imaging + AI

Don't read the scan.
Reach into it.

INSIDE is imaging software you work with a pen. The slices become a body in 3D, AI segments the organ you point at, and finds one when you ask for it by name.

The film is rendered frame by frame from the real page on my PC's GPU, with my own hand and pen on the Cintiq cut in between. One thing is sped up: a real lift takes about 45 seconds, the film shows it in one. Sound on.

01What it does

A scan you can hold, cut and question.

An axial section of the chest

Move through it with a pen

Left and right turn the body, down cuts into it, pressure peels it from the skin in. A slider tips it to a view from above, so the cut becomes an axial section.

The heart lifted out and cut open

Point at an organ, lift it out

Hold the pen's button to freeze the view, press twice on the organ. SAM 2 traces it through all 409 sections, Claude names it and says where it really ends.

Typing liver into Find

Or ask for it by name

Type or tap "liver". Claude looks through the sections, picks one and a point inside the organ, and the same tracing takes it from there.

The same body as CT

Anatomy and CT on the same spot

The body's own CT, lined up with the sections. A second slider blends the cut from tissue to CT; bone comes straight out of it as a skeleton.

02The lift

From the whole body to one organ, in a second.

The moment the organ comes out is the product's whole idea, so it gets its own animation: a scan line turns the body into a fine mesh, the mesh falls apart into cubes from the outside in, and only the organ is left.

01Point
02Scan to mesh
03Into cubes
04Only the organ
05Whole, or one chamber

03On the Cintiq

No mouse. A pen on the body.

It runs on a Wacom Cintiq 16, a screen you draw on. The pen's tilt, pressure and side button are all part of it, so it feels closer to holding a scalpel than to clicking through a viewer.

Turning an axial section with the pen The cut following the pen Turning the liver with the pen The cut torso turning under the pen

04Every layer

The same body, five ways.

Skeleton and organs together
Skeleton and organs, back together after the exploded view.
A CT slice
Where the film starts: one grey CT slice.
The body building from one slice
The body building itself from that slice.
The skeleton
Bone, by density, from the CT.
The exploded view
Exploded: every lifted organ pulled out from the middle.

05How it works

The browser draws. The GPU next to it does the seeing.

The page marches rays through the whole volume in WebGL2, 610 by 345 by 409 voxels at about a millimetre each. A small server on the same PC runs the segmentation and asks Claude what it is looking at.

tilt · pressurebuttona clicka montagethe organ's mask Pen Wacom Cintiq 16 The page WebGL2 raymarch of the volume cut · peel · CT blend · explode PC · RTX 3060 SAM 2.1 traces it through 409 sections every level: chamber to organ Claude names each level, says where it ends, finds organs by name

06What fought back

Four problems I did not see coming.

01

SAM doesn't know where a heart ends.

Traced section by section, the heart's outline kept going up into the aorta and the great vessels. A size rule couldn't tell where to stop.

What I did

Claude looks at a montage of every eighth section, names what each outline really holds and gives the range. It also names the levels: the left ventricle, the right side, the whole heart.

02

The newer model saw nothing.

SAM 3 finds things by name, which sounded perfect. On this body it found the heart in 0 of 150 sections, and called a slice "meat" with 85% confidence.

What I did

Stayed on SAM 2, where a click says where, and let Claude do the naming and the finding.

03

The CT came in a 1990s format.

The frozen CT of the same body is Unix-compressed GE Genesis files, and it doesn't line up with the photographed sections: other pixel size, other centre, front and back flipped.

What I did

Wrote the decoder, then lined each CT slice up with the body's outline in the photographs. Bone above 200 HU became the skeleton.

04

A double press that never arrived.

Hold the button to freeze the view, press twice to lift. On the Cintiq the view kept jumping: Windows slips mouse events in while the pen's button is down.

What I did

Logged the real presses, measured the rhythm (a second between them), kept the view still for that second and ignored mouse events while a pen is near.

07Where it stands

Honest status, as of today.

Works now

  • The head and the torso of one body, in 3D, driven by the pen.
  • Lifting any organ the pen points at, with its levels.
  • Finding an organ by name.
  • Resting the pen: Claude names what is under the tip.
  • CT blend, skeleton from the CT, exploded view.

Not yet

  • A patient's scan. Everything here is one public body; a patient has CT or MRI in grey, without the photographed colour.
  • Speed: a lift takes about 45 seconds.
  • MRI, and cleaner organ ends (they are cut flat where the range ends).
  • It is a research prototype, not a medical device.

08Under the hood

Built on a real body and a few good models.

Visible Human Project, NLM: 409 colour sections, 1 mmFrozen CT of the same body WebGL2 volume raymarcherSAM 2.1 base+ video predictorClaude Opus 5.5 Python server on an RTX 3060Wacom Cintiq 16Own .Z / GE Genesis decoder Film rendered frame by frame from the pageMusic generated with ACE-Step, on the same PC