The headline evidence is a radiologist study in which four-fold accelerated reconstructions were rated diagnostically interchangeable with fully sampled scans. That is a strong result, and it is specific to the anatomy and acceleration factor tested.
Reconstruction is compute-heavy and latency-sensitive, since it sits between the scanner and the reading queue. The sizing below assumes reconstruction as an inline step rather than a batch job.
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 undersampled k-space data from the scanner. 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.
fastMRI baselines (VarNet, E2E-VarNet)
Vendor: NYU / Meta AI (fastMRI)
What it does: reconstructs a diagnostic image from a quarter of the usual acquisition — radiologists rated four-fold accelerated reconstructions as diagnostically interchangeable with fully sampled scans, which turns scanner time back into throughput.
Radiologists rated 4x-accelerated reconstructions as diagnostically interchangeable with fully sampled scans in the fastMRI+ study (independent).
Requirement
Minimum
Medium
High
GPU type
RTX 3090
A100 80 GB
H100 80 GB
VRAM
24 GB
80 GB
80 GB
vCPUs
12
24
48
RAM
64 GB
128 GB
256 GB
Server
1× RTX 3090 24 GB
1× A100 SXM 80 GB
2× H100 SXM 80 GB
Rate (volumes/hour)
≈ 60
≈ 210
≈ 540
Choosing between them
The decision is the acceleration factor, not the model: how much time you want back against how much reconstruction uncertainty your readers will accept. We benchmark on your own k-space at several factors and let your radiologists set the line.
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 mri reconstruction and acceleration. 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.