Cardiac MRI segmentation is a solved-enough problem that the challenge leaderboards have converged: Dice around 0.90 for the LV cavity is the level a well-configured model reaches.
Two routes are in use. A self-configuring segmentation recipe trained on your own or challenge data is the accuracy reference. Vision-language models over image and report together add condition prediction and regression of measurements such as LVEF, and cost more per study.
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 cine cardiac MRI volume, typically a full short-axis stack. 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.
nnU-Net cardiac bundles (ACDC, M&Ms-2)
Vendor: DKFZ Heidelberg, with ACDC and M&Ms community models
What it does: segments the ventricles and myocardium at Dice 0.90–0.94 and derives the volumes and ejection fraction from them — the accuracy reference for this task, and it fine-tunes well on a modest local annotation set.
What it does: reads cardiac MRI alongside the report to regress ejection fraction and score 39 cardiovascular conditions, which goes further than segmentation alone at a higher cost per study.
Task-dependent; evaluated for LVEF regression and 39 cardiovascular conditions.
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 (studies/hour)
≈ 90
≈ 315
≈ 810
Choosing between them
If you need masks and volumes, the segmentation route is the right one and it fine-tunes well on a modest annotated set. If you need condition-level predictions from image and report together, the transformer models cover that at higher cost. Both are sized here for 3D volumes rather than single slices.
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 cardiac mri analysis. 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.