Reported AUC across body parts spans 0.85–0.96. That range is the point: the same model is clinically useful on a wrist and not on a rib series, and no open model in this area carries regulatory clearance.
Both groups here come from challenge lineages, which means the code is available and the quality varies by repository. Part of our work is selecting and re-validating rather than taking a leaderboard position at face value.
Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.
Models in this group take a single radiograph. Each table gives three hardware tiers — Minimum, the smallest configuration on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.
Fracture detection CNNs
Vendor: Community (RSNA and MURA challenge lineage)
What it does: flags fractures with a localisation heatmap, at AUC 0.85–0.96 depending on the body part. Wrist and long bone are strong, ribs and spine weak — so we scope a deployment to where it holds rather than selling a general fracture detector.
AUC 0.85–0.96 by body part; wrist and long bone strongest, ribs and spine weakest (independent). No regulatory clearance.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
6
12
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (images/hour)
≈ 12,000
≈ 42,000
≈ 108,000
Deep bone age models
Vendor: Community, BoneXpert-style open reimplementations (RSNA Bone Age challenge)
What it does: estimates skeletal age from a hand radiograph to within about four to six months of radiologist consensus — a well-behaved task and a straightforward time saving.
Mean absolute error about 4.2–6 months against radiologist consensus (independent).
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
6
12
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (images/hour)
≈ 12,000
≈ 42,000
≈ 108,000
Choosing between them
Scope to the body part first, then benchmark — a single AUC figure across all radiographs hides more than it shows. Paediatric caution applies throughout radiology here: most training data is adult, and paediatric-specific models are a documented gap.
Accuracy figures above are those the producers and independent evaluations report, on their own test sets. They are a shortlist tool, not a prediction of what you will see. At the start of a project we run a short proof of concept on a sample of your own data, which replaces them with real figures — so the cost and the schedule for the full engagement are known before anything is committed.
Send a representative sample, your expected volume and your latency target for musculoskeletal imaging. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.