CardioDiT

Latent Diffusion Transformers
for 4D Cardiac MRI Synthesis

MICCAI 2026

Marvin Seyfarth · Sarah Kaye Müller · Arman Ghanaat · Isabelle Ayx · Fabian FastenrathPhilipp Wild · Alexander Hertel · Theano Papavassiliu · Salman Ul Hassan Dar† · Sandy Engelhardt†

Joint last authorship

University Hospital Heidelberg Institute for Artificial Intelligence in Cardiovascular Medicine

Paper Code Cite

A heart in four dimensions

Synthetic short-axis cine MRI · synchronized across depth

Explore cardiac motion

Explore cardiac motion

Inspect a synthetic 3D volume throughout the cardiac cycle.

Loads on request · all 32 frames of one synthetic sequence.

Coarse through-plane sampling; slice views retain the source detail. Powered by NiiVue.

How CardioDiT works

Anatomy and motion, modeled together

Joint 3D+t synthesis in a compact latent space.

CardioDiT framework: encode 2D plus time slices, stack their latents into a full 4D volume, jointly denoise with a diffusion transformer, then decode.
CardioDiT framework · Figure 1View full figure

  1. Encode the sequence

    Compress each 2D+t slice with a spatiotemporal autoencoder.

  2. Model space and time

    Model stacked slice latents jointly with a 4D diffusion transformer.

  3. Decode the motion

    Decode the latent volume into a complete cardiac cine MRI sequence.

CardioDiT learns anatomy and cardiac motion jointly across a complete 3D+t volume. Four-dimensional patches preserve relationships across space and time.

See the evidence

Generation quality, measured

Results from the MICCAI 2026 paper · 50 generated volumes per model.

Public cohortsACDC + M&Ms-2 · CineMA features
Model FID ↓ Precision ↑ Recall ↑
2D+t LDM 100.1 0.89 0.17
3D+t U-Net 221.5 0.59 0.64
CardioDiT 71.9 0.97 0.70
Private cohortCineMA features
Model FID ↓ Precision ↑ Recall ↑
2D+t LDM 32.9 0.58 0.36
3D+t U-Net 51.3 0.43 0.23
CardioDiT 21.2 0.82 0.54

Each cohort uses its real validation set as reference.

Cardiac function across the cycle

Ejection fraction · mean ± standard deviation (%)

Public cohorts

CardioDiT
49.3 ± 15.1
Real
47.6 ± 17.1

Private cohort

CardioDiT
42.3 ± 6.3
Real
44.6 ± 7.3

Fixed spatial sizes and 32-frame standardization. Full structural and function analyses are reported in the paper.

Full results and ablations
Ongoing research

Beyond the original model

Ongoing experiments in spatial and temporal modeling.

F07 development example · central short-axis slice300,000 updates · 100 sampling steps · EMA weights

  1. Flexible geometry

    Axis-aware 4D rotary position encodings and support for variable spatial shapes.

  2. Native temporal lengths

    Development toward sequences with variable frame counts.

  3. Alternative objectives

    Experiments with flow matching and self-conditioning.

The example is from a fixed-shape F07 run. These experiments are separate from the paper benchmarks; comparative gains are not yet established.

Resources & citation

Build on CardioDiT

Paper, implementation, and citation.

Research resources

Pretrained weights are listed as forthcoming in the public repository.

Cite this work

@inproceedings{seyfarth2026cardiodit,
  title = {CardioDiT: Latent Diffusion Transformers for 4D Cardiac MRI
           Synthesis},
  author = {Seyfarth, Marvin and M{\"u}ller, Sarah Kaye and Ghanaat, Arman
            and Ayx, Isabelle and Fastenrath, Fabian and Wild, Philipp and
            Hertel, Alexander and Papavassiliu, Theano and Dar, Salman Ul
            Hassan and Engelhardt, Sandy},
  booktitle = {Medical Image Computing and Computer Assisted Intervention --
               MICCAI 2026},
  year = {2027},
  publisher = {Springer Nature Switzerland},
  series = {Lecture Notes in Computer Science},
  volume = {16890},
  pages = {69--79},
  doi = {10.1007/978-3-032-38189-7_7}
}
Marvin Seyfarth · University Hospital HeidelbergBack to top