Module 05 / Cross-module
Two scans of the same person never line up. That misalignment is the measurement.
The problem
Scan someone in January and again in August and you cannot simply subtract one from the other. They stood differently. They breathed differently. They are a different shape. Every voxel has moved.
The usual response is to treat that as noise — align the two volumes as best you can, throw the alignment away, and compare summary numbers. One muscle volume against another.
That discards almost everything. Two people can arrive at the same total with completely different distributions, and a single number cannot tell them apart.
The idea
Keep the alignment. It is the answer.
To make August match January you have to push every point in the volume some distance in some direction. That field of displacements — one vector per voxel — is what registration throws away.
Take its local volume-change factor at each voxel and you get one interpretable value at every point in the body:
↑
Above 1
tissue is locally larger than it was.
–
Exactly 1
unchanged.
↓
Below 1
locally smaller.
Not one number per region. A number per voxel. A map of where you changed, and by how much.
Two references
Change over time
Point it at the same person’s earlier scan.
Deviation from typical
Point it at a population reference — how this body differs from the distribution, region by region.
The only difference is what you compare against.
Where this comes from
This is not new mathematics, and that is deliberate. The technique descends from optical flow, generalised to three dimensions, where it became deformable image registration.
1981
Optical flow — estimating motion between two video frames.
20 years
A standard measurement in neuroimaging.
2–4%
Annual tissue loss detected in Alzheimer’s research, below what manual measurement resolves.
What has not existed is a clinical-grade, multi-organ implementation packaged as a measurement product. That is the gap.
At scale
Any programme that scans the same people repeatedly accumulates something more valuable than a series of scans. It accumulates the transformations between them.
One scan describes a body. Two scans describe a trajectory.
And a trajectory is measurable long before any single scan looks abnormal — which is the entire argument for scanning people who feel fine.
Status
Registration and validation shipped. Per-region feature extraction across the full pipeline is in progress.
Research use only. Not a diagnostic.
Scanning the same people more than once?
That is the case MorphX was built for. I would like to hear about it.