MT-BT Asklepion Labs · Research preview
MR is the better eye for soft-tissue brain damage — but at 2 a.m. in an emergency room, the scan that's already warm is the CT. MT-BT learns from MR and runs on CT alone.
You'd assume an AI has to see, in the field, the same kind of scan it saw while learning. Ours doesn't.
That asymmetry is the whole bet. The expertise locked inside an MR — and the radiologist who reads it — shouldn't belong only to the hospitals and the hours of day that can afford both. If a routine CT can be taught to carry some of that signal forward, the gap between the best-equipped hospital and the one down the road gets a little narrower.
For a single patient, an MR and a CT are two photographs of the same brain from slightly different cameras. Line them up precisely and anything an expert marks on the MR can be carried onto the CT — pixel for pixel, in the same place.
The two scans are brought into a shared coordinate space — correcting first for rotation and position, then scale, then the subtle local stretching two scanners of the same head always disagree on.
The injured region is marked on the MR, where damage is most visible, then projected onto the aligned CT. The CT inherits an expert-grade label it could never have produced on its own.
Across enough CT-plus-inherited-label pairs, a segmentation model learns the pattern directly on CT. From then on, it works from a CT and nothing else.
Scrub through the slices of a representative patient. Switch the base image between CT and diffusion MR, and toggle the two overlays — the clinician label transferred from MR, and the model's prediction made from CT alone. The Dice score is their overlap on the current slice.
Trained on transferred MR labels, it flags the kind of injury experts have marked before — the model behind the live case above.
A different system studies normal anatomy until it can reconstruct a healthy brain from memory. Hand it a real scan; wherever reconstruction and reality disagree, you get a map of unfamiliarity. It works the way you spot a stranger in a family photo — you don't need to know who they are to know they don't belong.
The most common way these systems lie to their creators is data leakage — slices of the same patient landing in both training and testing, so the model is graded on faces it has memorised. We split strictly by patient. The numbers we earn that way are smaller and real.
Labels are produced or reviewed by specialists, and transferred outlines are checked before they ever reach the model. The system highlights regions for a qualified clinician — it does not decide.
Weak alignment, a too-small lesion, scanner artefact, low contrast. That catalogue — not the average score — is what tells us which part of the pipeline to fix next.
The output is a damage-likelihood map with reviewable confidence, designed to be interrogated by the person reading the image — not a verdict handed down.
MT-BT is a wedge into a larger thesis: the best medical signal is often locked in data hospitals already have. We partner with investors, hospitals, and clinical teams who want to widen who that signal reaches. If that's you, let's talk.