Research

Efficient synthesis, volumetric and temporal modeling, and evaluation of generative models for medical imaging.

I study how to generate medical images efficiently—and how to evaluate what these models actually learn.

Efficient synthesis

How can we generate high-resolution medical images with practical computational resources?

My work explores compact latent representations and diffusion models for 3D CT synthesis. Alongside image quality, I study the computational and environmental cost of training and generation.

MedLoRD Latent Pollution Model

Volume & time

How can a generative model capture anatomy across both space and time?

I investigate diffusion transformers for controllable volumetric synthesis and joint 4D cardiac MRI generation, connecting spatial structure with temporal dynamics.

CardioDiT VolDiT

CardioDiT framework: slice encoding, joint 4D diffusion, and decoding into cardiac MRI volumes. — CardioDiT framework. Seyfarth et al., 2027, Fig. 1. Select the figure to view full size.

VolDiT framework: volumetric patch tokens and diffusion-transformer blocks with spatial control. — VolDiT framework. Seyfarth et al., 2027, Fig. 1. Select the figure to view full size.

Evaluation beyond realism

What makes synthetic data useful beyond looking convincing?

I study diversity, anatomical variability, and patient-data memorization. These questions help distinguish perceptual quality from the properties that matter when synthetic data is used in research.

WAD-Div Memorization study Coronary diversity

WAD-Div plots measure intrinsic dataset diversity for chest X-ray and lung CT as the number of unique samples increases. — Measuring intrinsic diversity with WAD-Div. Seyfarth, Dar & Engelhardt, 2026, Fig. 2.