The air cast
Take the CT volume of this head (GE Revolution Maxima, PNS protocol, 0.59×0.59×0.63 mm voxels), keep only the connected pocket of air that starts inside the nasal cavity, and render that pocket as a solid object — as if a cast had been poured into the airway and the head dissolved away. The scan was ordered for a breathing problem with a suspected deviated septum, so the one question this picture is built around is: where is this airway narrow, and how do the two sides compare?
Colour = side of an anatomy-derived midline plane: patient right patient left
One full turn (60 frames; same content as air_cast_spin.mp4). The big lateral wings are the maxillary sinuses, the lobed top is the frontal sinus, the crinkly middle column is the nasal cavity and ethmoid labyrinth, and the smooth stalk at the back is the nasopharynx ending at the inferior cut plane.
A cast is the negative of the anatomy: where the nasal passage is narrow the cast is thin, and where a turbinate bulges into the airway the cast is dented. The two sides of the middle column are visibly different — that observation is the artifact's whole job; what it means clinically is a question for a radiologist, not for this page.
Numbers
Side volumes include the sinuses, not just the breathing passages, and the midline is an estimated plane (through the midpoint of the two maxillary-sinus air centroids) — so treat the split as a comparison aid, not a measurement of "nasal airflow capacity". No normal range is quoted anywhere on this page because a threshold picture of one scan cannot support one.
How much is the threshold's opinion?
| air threshold | total | right | left | vs −500 |
|---|---|---|---|---|
| −400 HU | 100.38 cm³ | 54.77 | 45.61 | +3.2% |
| −500 HU (used) | 97.25 cm³ | 52.89 | 44.36 | — |
| −600 HU | 93.77 cm³ | 50.96 | 42.81 | −3.6% |
Moving the threshold ±100 HU slides a one-voxel skin over the whole surface (±~3.5%) but changes no structure: nothing connects or disconnects, and the right–left difference barely moves. The shape is robust; the absolute cm³ carry a few-percent threshold uncertainty on top of scanner partial-volume effects.
Take the object
Both files are the full-resolution marching-cubes surface of the gaussian-smoothed mask (σ = 0.8 voxel). Smoothing costs −2.6% volume vs the raw voxels and can pinch sub-millimetre passages closed in the mesh, which is why all volumes above are voxel counts, not mesh volumes.
MinIP — the airway as a projection
The study folder ships mip_frontal.png, a
maximum-intensity projection: each pixel keeps the densest voxel along
its ray, so bone wins everywhere and the airway is invisible. A
minimum-intensity projection is the same reduction with
min instead of max: the airway, the one reliably
darkest thing inside a head, wins every ray it touches. Room air would win
every ray of all, so the volume is first masked to the head and outside voxels
are set to +60 HU; display is inverted so bright = air.
The bright side-clusters are mastoid / middle-ear air cells — genuinely air, deliberately not part of the cast above (they are separate air spaces, not connected to the nose at any tested threshold).
For contrast: a ~4 mm average slab through the mid nasal cavity and the midline. An average behaves like one thick CT slice — air is dark again, off-slab structures vanish instead of projecting — which is why MinIP and a slab of the same data read so differently: one is a search for the darkest voxel over 10+ cm of ray, the other a local mean over 4 mm.
Proof the mask didn't leak
The one silent failure mode of airway segmentation on this volume: a global "< −500 HU" selects 74.8% of the field of view, because everything around the head is also air (plus out-of-FOV padding at −3065 HU). The overlay shows the seeded component stayed inside the head: room air (dark red) untouched, mastoid air uncoloured, and the yellow line marks the inferior cut. The colour seam is the estimated midline plane.
Method & honesty notes
- Segmentation: air = HU < −500 inside a head mask (tissue > −400 HU, largest component, holes filled per axial slice), then the single 6-connected component seeded at the darkest voxel inside the nasal cavity. Threshold stated on every image because the surface is the threshold.
- Cuts: inferiorly a flat plane at z = −8 mm (just below the soft-palate tip; keeps vestibule + nasopharynx, drops the oropharyngeal column). At the nostrils the cast ends where nostril air meets room air at the skin line. The flat floor on the cast is the cut, not anatomy.
- Included by honesty: a thin connected sheet of oral-vestibule air (lips/teeth) — real connected air, < ~2 cm³, ignore it when comparing sides.
- Orphaned air: a 0.83 cm³ cell in right sphenoid / posterior-ethmoid territory (plus a 0.07 cm³ neighbour) never connects at −400/−500/−600 and is excluded. Whether it is a separate cell or sealed by mucosa, this pipeline cannot tell.
- Midline: plane through the maxillary air centroids. It tracks the bony midline well, but near a deviated septum "geometric midline" and "whose air is whose" are different questions — that mismatch is exactly what the colour imbalance is designed to show. The frontal sinus is one cavity; its teal/coral split is arbitrary.
- Rendering: orthographic first-hit ray casting of the voxel mask, Lambertian shading with normals from the same smoothed field the mesh comes from. Brightness differences are lighting and depth fog, never density.
- Reformats are computed from the axial volume; the shipped coronal/sagittal series have uneven slice spacing and were not used.
- Full details and failure modes:
SOFT_SPOTS.md; reproducible pipeline:make_air_cast.py.