MedSpatial

The problem

The scans exist. The labels don’t.

Patient-specific surgery runs on 3D anatomy: where the fragments of a fracture belong, where a screw can safely go, what shape an implant or an appliance must take.

Teaching software any of that normally takes hundreds of cases labelled by experts, and most specialties will never have them. Hospitals, meanwhile, hold enormous numbers of scans that no one has labelled for any task.

MedSpatial began in cleft care, designing infant molding appliances from a few dozen intraoral scans. A few dozen scans cannot teach a model what 3D anatomy is from scratch. They should only have to teach what is specific to cleft care.

Unlabelled scans number in the tens of thousands. Expert-labelled cases, in the hundreds.

3D cases in example collections, log scale.

Scans without task labels: can teach general anatomyCases with expert labels: what each clinical task needs

Show the numbers

Collection	Cases
Public chest CT with radiology reports	47,149
Public abdominal CT with reports	15,314
Public whole-body CT, bones outlined	1,228
Public pelvis CT, bones outlined	1,106
Real pelvic fractures with surgeon-verified reductions	170
Pelves with 57 landmarks placed by experts	99
Infant cleft scans available to MoldStudio	a few dozen

The idea

One model of anatomy, shared by every task.

A surgeon does not relearn the pelvis for each new procedure. MedSpatial gives software the same head start: learn the shapes and relationships of anatomy once, from many unlabelled scans, then adapt to each clinical task with a small number of expert cases.

Layer 1 · Perception

What is where?

Existing segmentation tools outline bones, organs and vessels in CT, MRI and surface scans. MedSpatial builds on them.

CT
MRI
Surface scans

Layer 2 · MedSpatial

How is it shaped, and how does it fit together?

A learned model of anatomical shape: which point on one patient matches which point on another, where each piece belongs, and how sure the model is.

shape
correspondence
relationships
uncertainty

Layer 3 · Engineering

What should we build?

Design tools turn a clinical goal into reduction plans, implants, guides and appliances, each checked by a deterministic validator.

plans
implants
guides
appliances

current anatomy+desired state+available intervention→validated design

How it learns

Practise on bones nobody labelled. Then learn the task.

The model first teaches itself what bone surfaces look like by filling in pieces we hide from it. Only then does it see a handful of expert-labelled cases for the task at hand.

Step 1 · Scans
Ordinary CT scans
Existing tools already outline each bone. The main pool is 1,228 public whole-body CTs (TotalSegmentator).
Step 2 · Surfaces
Bones become points
Each bone's surface becomes a cloud of points measured in millimetres. This person has 62 bones and about 185,000 surface points.
Step 3 · Practice, no labels
Fill in what is hidden
The surface is cut into small patches and 60% are hidden (teal). The model learns by rebuilding them. Nobody has to label anything.
Step 4 · The task
A few labelled cases
The practised model is then fine-tuned on 10 or 50 expert-labelled cases: here, naming each rib; elsewhere, putting fracture fragments back in place.

The same skeleton with the ribs, costal cartilage, hip bones and sacrum in amber and every other bone in bone colour.

The hardest test so far · Route 14

Practise on everything except the ribs and pelvis. Then ask about ribs and pelvis.

To check that the model learns anatomy in general rather than memorising one bone, a model called X practised on 914 people with every rib, costal cartilage, hip bone and the sacrum left out. X still helped on both tasks it had never seen: labelling ribs and reducing pelvic fractures.

What X practised onNever shown to X

Renders of one training subject’s bone surfaces (TotalSegmentator s0591), not model output.

Progress

From nothing working to real patients in four weeks.

The first application is pelvic fracture reduction: putting each broken piece back where it belongs. It is hard, it is measurable, and a public set of real cases with surgeon-verified answers exists to score against.

Fragment error on 170 real pelvic fractures, best result to date

Mean distance between each fragment and its surgeon-verified position, in millimetres. Lower is better. Each point averages two independent training runs.

MedSpatial, best to dateClassical methods, no learningLeaving every fragment where it lies

37%placed within 5 mm and 10°, untouched

42%first learned model, 18 Oct

48%trained three times longer, 19 Oct

51%with anatomy pretraining, 25 Oct

Show the numbers

Date	Method	Mean error, mm	Placed within 5 mm, 10°
16 Oct	Leave every fragment where it lies	10.26	37.2%
16 Oct	Copy the healthy side onto the broken one	11.30	32.4%
16 Oct	Fit fragments to an average pelvis	14.29	13.9%
18 Oct	First learned model, runs 1 / 2	8.92 / 8.81	42.1% / 41.6%
19 Oct	Trained three times longer	7.76 / 7.84	49.4% / 46.1%
25 Oct	With anatomy pretraining, trained longer	7.31 / 7.38	51.3% / 51.4%

On 26 October the 170 real cases were split: 82 for working and tuning, 88 sealed and scored once per finished model. This chart stops there. Results on the sealed half, down to 6.6 mm, are under Evidence.

28–30 Sep
Started on synthetic shapes
About seventy small experiments on spheres, cones and tori. Nothing carried over, and an outside review showed the test could not tell good models from bad. We moved to real anatomy.
Ruled out
2–3 Oct
A benchmark built from real bone
Partial views of ribs, vertebrae, hips and skulls from public CT. A compact learned model beat the standard hand-built method on people it had never seen.
Result
14–15 Oct
First transfer test: no
What that model learned did not help it place pelvic landmarks, frozen or fine-tuned. Recorded as a clean negative, and it changed what we trained next.
Ruled out
16 Oct
Sealed test, opened once: confirmed
The bone-alignment lead held on 37 people locked away from the first day, under a pass mark written before the run.
Result
16 Oct
A yardstick on 170 real fractures
Every classical method did worse than leaving the fragments where they were. That set the bar any learned method had to clear.
Built
17–19 Oct
First learned reduction beats doing nothing
Trained only on simulated fractures, scored on real ones: 8.9 mm, then 7.8 mm with longer training.
Result
19 Oct
Landmarks with zero labels
6.8 mm across 57 pelvic landmarks, ahead of a network trained on every labelled patient.
Result
20–23 Oct
Teaching the model the map of the pelvis
Given most of a pelvis, it learned where each point on the surface belongs, to about 5 mm, from shape alone, and gave the first label-saving result: 0.6 mm better with four labelled patients. Small isolated pieces stayed hard, and three ways of handing that knowledge to the fracture task did not help.
Mixed
24–25 Oct
The founding idea, shown on real patients
Pretraining on unlabelled pelvis scans made fracture reduction more accurate, under a rule written before the run, in both independent runs. With longer training: 7.3 mm, 51% of fragments placed.
Result
26 Oct
The real cases split in two
82 fractures for working and tuning, 88 sealed and scored once per finished model. The same day, a combined training objective did not improve fractures (Route 6).
Built
27 Oct–2 Nov
Bigger pelvis models stopped helping
Across four model sizes the gain per step shrank from 0.37 to 0.01 mm on the practice task, and the largest did not reduce fractures better (Routes 7, 7b).
Ruled out
28 Oct
A second anatomy: labelling ribs
Practice on 782 unlabelled rib cages lifted rib labelling on a sealed test of 160 cases: 0.89 overlap from 10 labelled cases, against 0.74 from scratch (Route 8).
Result
30 Oct–2 Nov
A more realistic fracture simulator
Training on more lifelike simulated fractures made the real ones worse: 8.3 and 9.4 mm against 7.7 mm on the working half (Routes 11, 11b).
Ruled out
31 Oct
One model for every bone, first attempt
A whole-skeleton model that also learned surface maps fell well short of the rib specialist: 0.78 overlap against 0.89 (Route 9).
Ruled out
2 Nov
Bone identity, learned with no labels at all
Keeping only the fill-in-the-hidden-patches practice, the whole-skeleton model tells bone types apart 88% of the time, up from 67%, level with the rib specialist (Route 13).
Result
3 Nov
A refinement pass, confirmed on the sealed half
Teaching the model to refine its own answer beat the earlier best on the 88 sealed cases in both runs: 7.3 and 7.1 mm against 7.8 and 7.9 (Route 10b). Refining for longer missed its pass mark (Route 15).
Result
5 Nov
It helps anatomy it never saw
A model that practised without any rib or pelvis still beat training from scratch on both: rib labelling and fracture reduction (Route 14).
Result
5 Nov
Fine-tuning on real cases: best model so far
The refinement model, fine-tuned on the 82 working cases, reached 6.7 and 6.5 mm on the 88 sealed ones, 0.59 mm better in both runs (Route 17).
Result
6 Nov
Not just more training
Given three times the training, a model from scratch still trailed the pretrained model trained just as long by 0.7 mm, and only tied one trained a third as long (Route 18).
Result
10 Nov
More people, or a bigger model?
Practising on more people helped fractures but not ribs. A model about eight times larger helped neither, and no general model reached the rib specialist (Route 19).
Mixed
10 Nov
Bones only, or bones plus soft tissue?
Four new encoders trained, their rib runs scored once on the sealed test, fracture runs under way in the cloud. No verdict yet (Route 21).
Running
10 Nov
A harder practice task, and gentler fine-tuning
Rebuilding whole hidden bones as well as patches (Route 20), and lower fine-tuning rates for the large model (Route 22). Both running on the laptop.
Running
Next
A pelvis top-up for the general model
Approved, to start after Route 22: a short extra round of practice on pelvis scans, to see whether the general model can catch the pelvis specialist (Route 23).
Approved

Evidence

What the results show.

Six findings carry the case, newest first. Each was tested against a control chosen to catch wishful thinking, and each is shown here with its numbers and its spread across independent runs.

On fractures kept out of every decision, pretraining still wins, and it is not just extra training.

0.7 mm

less error than the same model trained from scratch for just as long, on 88 sealed real fractures (0.75 and 0.67 mm in two runs)

Since 26 October, 88 of the 170 real fractures have been sealed. No model is tuned on them, and each finished model is scored on them once.

Training from scratch three times longer helps, but it only ties a pretrained model trained a third as long. Fine-tuning on the 82 working cases then gives the best result so far: 6.6 mm.

How it was testedQuestions and pass marks written before each run (Routes 17 and 18). Two independent training runs per model. Both gains in the caption have 95% intervals above zero in both runs.

Mean fragment error on the 88 sealed real fractures

Millimetres, lower is better. Two marks per row: two independent training runs. The number is their average.

Trained from scratchWith anatomy pretraining

Pretrained ahead of scratch at 60,000 updates: 0.75 / 0.67 mm. Fine-tuning on real cases then adds 0.59 / 0.59 mm. Scratch at 60,000 against pretrained at 20,000: −0.14 / −0.13 mm, intervals across zero, a tie.

Show the numbers

Model	Read-out	Runs 1 / 2, mm
From scratch, 20,000 training updates	one pass	9.03 / 9.12
From scratch, 60,000 updates	one pass	7.96 / 8.00
Pretrained, 20,000 updates	one pass	7.82 / 7.87
Pretrained, 60,000 updates	one pass	7.21 / 7.33
Pretrained with a refinement pass, 20,000 updates	three passes	7.31 / 7.10
… then fine-tuned on the 82 working cases	three passes	6.72 / 6.51

Paired comparison, same 88 cases	Run 1 [95%]	Run 2 [95%]
Pretrained ahead of scratch, both 60,000	0.75 [0.43, 1.05]	0.67 [0.35, 0.97]
Scratch at 60,000 against pretrained at 20,000	−0.14 [−0.37, 0.10]	−0.13 [−0.34, 0.10]
Fine-tuning on real cases adds	0.59 [0.18, 1.08]	0.59 [0.20, 1.04]

Each model is read out the way its pre-registration named: one pass, or three refinement passes for the refinement models. Compare rows within the same read-out where it matters.

Pretraining helps anatomy the model has never seen.

+0.12

rib-labelling overlap from 10 labelled cases, for a model that never saw a rib: 0.86 against 0.74 from scratch

Rib labelling on a sealed test of 160 cases

Label Dice: how well each predicted rib overlaps the expert’s, 1.0 is perfect. Higher is better.

Pelvic fracture reduction on all 170 real cases

Mean fragment error, mm, lower is better. Working and sealed halves pooled, as the pre-registration set out.

From scratchX: never saw a rib or a pelvisSpecialist pretrained on that anatomy

How much of the specialist’s gain the general model X recovers

X’s gain over scratch as a share of the specialist’s gain over scratch, with 95% intervals. Ribs: four fine-tuning runs. Fractures: two pairings of runs.

Show the numbers

Task	From scratch	X	Specialist	X’s share [95%]
Ribs, 10 cases, run 1 / 2	0.739 / 0.740	0.860 / 0.854	0.893 / 0.896	79% [71, 86] / 73% [65, 81]
Ribs, 50 cases, run 1 / 2	0.882 / 0.880	0.936 / 0.942	0.946 / 0.932	84% [70, 97] / 121% [101, 154]
Fractures, mm, pairing 1 / 2	9.12 / 9.24	8.52 / 8.36	7.76 / 7.96	44% [35, 53] / 69% [59, 78]

How it was testedPre-registered (Route 14). X practised on 914 people with every rib, costal cartilage, hip bone and the sacrum left out. Two runs each; all four rib gains and both fracture gains over scratch have 95% intervals above zero. The specialists still do better: X recovers about three quarters of the rib specialist’s gain and half to two thirds of the pelvis specialist’s.

More people in practice help fractures. A bigger model helps neither task, so far.

0.18–0.28 mm

less fracture error from practising on 914 people instead of 100. Rib labelling did not move.

Which lever helps which task

Each panel changes one thing. Left: how many people the model practised on. Right: how large the model is. Rib labelling from 10 cases, on the sealed rib test; fracture error on all 170 real cases.

Ribs · more people

Label Dice, higher is better

Ribs · bigger model

Label Dice, higher is better

Fractures · more people

Mean fragment error, mm, lower is better

Fractures · bigger model

Mean fragment error, mm, lower is better

Two fine-tuning runs per modelX2: the 914-person recipe, practised a second timeSpecialist pretrained on ribs or pelves, range of two runs

Show the numbers

Model	People	Size	Ribs, runs 1 / 2	Fractures, mm
D100	100	4.9 M	0.862 / 0.857	8.70 / 8.64
D300	300	4.9 M	0.846 / 0.872	not run
X1	914	4.9 M	0.860 / 0.854	8.52 / 8.36
X2, second practice run	914	4.9 M	0.851 / 0.850	8.45 / 8.37
M	914	14.6 M	0.863 / 0.854	not run
L	914	38.5 M	0.864 / 0.871	8.90 / 8.45
Rib specialist R	782 rib cages		0.893 / 0.896
Pelvis specialist P	pelvis scans			7.76 / 7.96

Paired gains [95%], runs 1 / 2. Fractures, 914 people against 100: 0.18 [0.07, 0.30] / 0.28 [0.15, 0.41] mm. Fractures, L against X1: −0.38 [−0.51, −0.25] / −0.09 [−0.22, 0.05] mm. Ribs, X1 against D100: −0.002 / −0.003, both intervals across zero. Ribs, L against X1: +0.004 [−0.004, 0.012] / +0.017 [0.009, 0.026], ahead in one run only. Sizes count the encoder alone.

How it was testedPre-registered (Route 19), verdict 10 November. Every model practised with the same recipe and was fine-tuned twice per task. X2 repeats X1’s recipe to show how much two identical practice runs differ: 0.009 on ribs, 0.07 mm on fractures. The 300-person and mid-size models were not run on fractures, by design.

In progress: does practising on soft tissue as well as bone help?

Pending

no verdict yet. It is written once the fracture runs are scored, and shown here only then.

Route 21 trains two kinds of encoder on the same people: bones only, or bones plus organs, vessels and back muscles. Two practice runs of each. All four are then fine-tuned on ribs and on fractures.

A lung-lobe task was planned too. It was withdrawn on 9 November, before any encoder was trained, when a pre-set check found it could not tell encoders apart well enough. Its 100 sealed lung scans stay unopened.

Four encoders trained and checked · 10 Nov
16 rib runs fine-tuned, then scored once on the sealed rib test · 10 Nov
Fracture runs in the cloud · started 10 Nov
Verdict on ribs and fractures together

First shown on 170 real cases, 24–25 October: learning the pelvis first makes fracture reduction more accurate.

0.4–1.0 mm

less error than the identical model trained from scratch

Same model, with and without anatomy pretraining

Mean fragment error on 170 real fractures, mm, lower is better. The gain grows with more task data.

Trained from scratchWith anatomy pretraining

Show the numbers

Training fractures, run	From scratch	Pretrained	Paired gain [95%]
1,000, run 1	9.59	9.15	0.43 [0.24, 0.64]
1,000, run 2	9.69	9.03	0.66 [0.45, 0.87]
16,000, run 1	9.12	8.13	0.99 [0.78, 1.21]

A second 16,000-fracture comparison is left out: its from-scratch model never started learning.

Most of the gain arrives with the first 100 scans

Mean fragment error by the number of unlabelled pelvis scans used for pretraining, mm.

The model already benefits from 100 unlabelled scans. Since then, a model practising on the whole skeleton has shown that going from 100 to 914 people does keep helping fractures, by 0.18–0.28 mm (Route 19, above).

Show the numbers

Unlabelled scans	Mean error, mm
None, trained from scratch	9.59
100	9.02
250	8.94
756	8.89

How it was testedPass mark written before the run. The pretrained model is compared with an identical model started from scratch, with its settings tuned for the scratch model. Two independent training runs, scored once on 170 real cases.

Landmarks placed on a new pelvis with zero labelled examples.

6.8 mm

across 57 landmarks, with no labelled training data

Our pipeline places 57 anatomical landmarks on a pelvis it has never seen without a single labelled example, and lands closer on average than a network trained on every labelled patient.

That makes it a candidate for a first clinician tool: pre-placing landmarks for a specialist to confirm, which turns hours of annotation into minutes of review.

How it was testedCross-validated over 79 patients, each scored as an unseen case. Two settings were tuned on 20 of them. Another 20 patients stay sealed.

Mean landmark error across 79 patients

Millimetres, lower is better. The amber line marks how far two human experts differ.

Show the numbers

Method	Labels used	Mean error, mm
MedSpatial pipeline	0	6.78 [6.58, 6.99]
Network trained on every labelled patient	all	7.19
Best-matching templates from the labelled set	all	8.22
Average landmark positions	all	9.05
Two human experts, difference	n/a	4.56 ± 2.04

A compact learned model aligns bone scans better than the classic method.

+6 to +12 pts

more bone pairs aligned, on 37 people never seen in training

Lining up two partial scans of the same bone is the basic act of 3D medical geometry: it underpins navigation, follow-up comparison and every registration step in planning.

A small MedSpatial model, trained in a few GPU-hours, aligns more bones than the standard hand-built method from 2009. The lead is largest on ribs and hips.

How it was testedA test set sealed from the first day and opened once, with the pass mark set in advance. Each alignment was repeated five times and averaged.

Bone pairs aligned within 5 mm and 15°

922 pairs of partial bone scans from 37 people, each person weighted equally.

Show the numbers

Method	Aligned	Gain [95%]
Standard hand-built method	56.0%
MedSpatial compact model, run A	62.9%	+6.9 [+4.1, +9.6]
MedSpatial compact model, run B	62.2%	+6.2 [+3.3, +8.9]
Run A, trained twice as long	67.5%	+11.5 [+8.8, +14.2]

How we work

Built to be believed.

Every experiment states its question, its controls and its pass mark before it runs. The result is recorded whether it is good news or not.

119experiment protocols and pre-registrations written before the runs they describe

84results reports, negative results included. Each pre-registration also records its own result.

5sealed test sets, never used for tuning. Each finished model is scored once on the set for its task. Two have never been opened.

1laptop runs every model. Training used that laptop and free cloud notebooks.

Controls that catch wishful thinking
Every pretrained model is compared with the identical model started from scratch, so a gain from the architecture is never mistaken for a gain from learning.
Failures are kept
More than a dozen routes that did not work are documented as carefully as the ones that did. Each one narrowed down what the next experiment should try.
Local first
Models run on a single machine with no cloud AI service in the loop, which suits hospitals that keep patient data on site.

What it could become

One foundation. Many procedures.

Many procedures share one structure: current anatomy, plus a goal, plus a device, gives a design. If one model of anatomy helps several of them, it stops being a project and becomes infrastructure.

Where a shared anatomy model could reach

Inner ring: shown in research. Middle ring: next, with public data identified. Outer ring: the long-term vision. Select one to see it as a design problem.

Shown in research

Pelvic fracture planning

Current anatomyFractured pelvis

GoalRestored pelvic ring

InterventionPlates and screws

DesignA fragment-by-fragment reduction plan

6.6 mm mean fragment error on 88 sealed real fractures, after fine-tuning on 82 others. Pretrained models lead training from scratch by about 0.7 mm at equal length.

The proving ground for the approach, on public research data.

From a clinical goal to a validated design.

The long-term system pairs the anatomy model with design tools and checks. An agent proposes a step, a geometry engine builds it, and a deterministic validator accepts or rejects it. Nothing counts as done until a validator passes it, and every operation is versioned like code.

InputClinical goal

“Restore the pelvic ring” or “an appliance that guides the cleft closed over six weeks”.

InputPatient spatial model

Bones, vessels, soft tissue and landmarks of this patient, from MedSpatial.

OutputValidated, versioned design

A plan, implant, guide or appliance, with the full history of how it was made and which checks it passed.

Operation log · git for medical geometryIllustrative example

op 001	orient_anatomy	pelvis to standard frame	pass
op 002	identify_landmarks	57 placed, 3 flagged for review	pass
op 003	offset_surface	plate wall 1.5 mm	fail · below 2.0 mm minimum
op 004	offset_surface	plate wall 2.0 mm	pass
op 005	check_clearance	screw paths against joint surface	pass

The road from here.

Now · research

Prove the foundation

Label-free practice on the whole skeleton
Two anatomies: pelvic fractures and ribs
Testing what practice needs: more people, larger models, soft tissue

Next · first tools

Give clinicians time back

Landmark pre-placement for specialists to confirm
More anatomy: spine, teeth, skull
A general model that matches the specialists

Later · the patient model

See the whole patient

Bone, vessels, soft tissue and teeth in one spatial model
Change over time: growth, healing, treatment
Uncertainty a surgeon can see

Vision · engineering

Design what fits

From clinical goal to validated, manufacturable design
Cleft appliances, implants, guides and braces
Every operation checked and traceable

Where we stand

Early, measured, and specific about the gaps.

Research stage. Nothing here is a clinical claim, and MedSpatial is not a medical device.
Not yet the most accurate. The best published fracture systems still have lower error. They are scored on a different, hidden test set and train on the real cases we keep for testing, so the numbers do not compare directly.
Two anatomies so far. Pretraining has helped pelvic fractures and rib labelling, even when the model never saw either. A general model still trails the specialists pretrained on ribs or pelves.
Bigger is not better yet. A model about eight times larger improved neither task. Practising on more people helped fractures but not ribs.
Small fragments are the frontier. The model improves every size of fragment, but the smallest pieces still carry the most error.

Error by fragment size, 25 October model

Mean error per fragment across the 823 moving fragments in all 170 real cases, mm, lower is better. Each row is a quarter of all fragments.

Left where it liesAfter MedSpatial reduction

Show the numbers

Fragment size	Left in place	Reduced	Placed within 5 mm, 10°
Smallest quarter, under 17 cm³	13.61	11.04	17%
17 to 45 cm³	10.13	7.37	44%
45 to 236 cm³	10.56	7.12	54%
Largest quarter, over 236 cm³	8.62	5.96	71%

Work with us

Build the anatomy layer of medicine with us.

Clinical partners
Surgeons and departments in orthopaedics, craniofacial and dental care who want help with 3D planning and can tell us what matters at the table.
Imaging and data partners
Hospitals that keep data on site. MedSpatial runs locally, with no cloud AI service in the loop.
Engineering collaborators
People who design implants, guides and appliances and know exactly where geometry is the bottleneck.

To start a conversation, reach out through the person who shared this page with you.

Learn anatomy once. Plan for every patient.

The scans exist. The labels don’t.

One model of anatomy, shared by every task.

What is where?

How is it shaped, and how does it fit together?

What should we build?

Practise on bones nobody labelled. Then learn the task.

Ordinary CT scans

Bones become points

Fill in what is hidden

A few labelled cases

Practise on everything except the ribs and pelvis. Then ask about ribs and pelvis.

From nothing working to real patients in four weeks.

Started on synthetic shapes

A benchmark built from real bone

First transfer test: no

Sealed test, opened once: confirmed

A yardstick on 170 real fractures

First learned reduction beats doing nothing

Landmarks with zero labels

Teaching the model the map of the pelvis

The founding idea, shown on real patients

The real cases split in two

Bigger pelvis models stopped helping

A second anatomy: labelling ribs

A more realistic fracture simulator

One model for every bone, first attempt

Bone identity, learned with no labels at all

A refinement pass, confirmed on the sealed half

It helps anatomy it never saw

Fine-tuning on real cases: best model so far

Not just more training

More people, or a bigger model?

Bones only, or bones plus soft tissue?

A harder practice task, and gentler fine-tuning

A pelvis top-up for the general model

What the results show.

On fractures kept out of every decision, pretraining still wins, and it is not just extra training.

Pretraining helps anatomy the model has never seen.

More people in practice help fractures. A bigger model helps neither task, so far.

In progress: does practising on soft tissue as well as bone help?

First shown on 170 real cases, 24–25 October: learning the pelvis first makes fracture reduction more accurate.

Landmarks placed on a new pelvis with zero labelled examples.

A compact learned model aligns bone scans better than the classic method.

Built to be believed.

Controls that catch wishful thinking

Failures are kept

Local first

One foundation. Many procedures.

Pelvic fracture planning

From a clinical goal to a validated design.

The road from here.

Prove the foundation

Give clinicians time back

See the whole patient

Design what fits

Early, measured, and specific about the gaps.

Build the anatomy layer of medicine with us.

Clinical partners

Imaging and data partners

Engineering collaborators