Segmenting structures of interest is a crucial first step in any quantitative analysis pipeline. In our group, robust and adaptable segmentation approaches are developed for different application domains. While the primary focus lies on material-science and medical imaging data, the methods are designed to be transferable and easily applied to new datasets. 

To overcome the limited availability of annotated images, synthetic training data are generated for supervised segmentation tasks. Binary structures are produced using a GAN and subsequently refined with a CycleGAN to add realistic noise and texture. In this way, highly realistic training data are created to support the development of robust segmentation models.

Super-resolution methods are used to increase the resolution of volumetric or two-dimensional image data. In volumetric MRI, for example, low through-plane resolution often limits downstream analysis. To address this issue, a supervised VAE–GAN framework is used to generate realistic high-resolution training data. This enables accurate super-resolution reconstruction and leads to more robust volumetric MRI segmentation.

Event detection is an essential task when working with temporal image sequences. In dynamic contrast-enhanced MRI (DCE-MRI) of the lung, for example, breathing motion must be identified and removed to enable meaningful downstream analysis. To automate this step, learning-based models are employed that reliably detect respiratory motion, even under varying image contrast conditions.

A broad portfolio of analysis methods is available to reliably quantify geometric properties in digital material twins. Metrics such as porosity, wall thickness, and pore radius can be extracted with high accuracy, providing a detailed description of the underlying microstructure. These quantitative descriptors form an essential basis for establishing structure–property relationships and supporting data-driven materials analysis.

Data-driven analysis methods are used to extract higher-dimensional descriptors from microstructures. Statistical metrics such as two-point correlation functions combined with principal component analysis, as well as neural network–based approaches, enable the identification of latent structural features. These representations support the clustering and comparison of microstructures and provide deeper insights into underlying structure–property relationships.

To accelerate the simulation of wetting behaviour in porous membranes, pore network models are extracted from digital membrane twins. These models provide a simplified yet informative representation of the pore space, enabling faster simulations and allowing larger material domains to be analyzed efficiently while preserving the essential characteristics of the underlying microstructure.

To quantify perfusion in dynamic contrast-enhanced MRI (DCE-MRI) of the lungs, an automated pipeline is developed. Based on a mathematical model, the residue function is obtained from 3D+t lung measurements by deconvolving the sub-traction image with the arterial input function. This residue function is subse-quently used to compute quantitative perfusion maps, including parameters such as pulmonary blood flow (PBF) and the percentage of perfusion defects (QDP).

Diffusion models are trained on segmented and reconstructed digital material twins to generate synthetic, highly realistic microstructure representations. By conditioning the generation process, the properties of the generated microstructures can be specified in a targeted and controlled manner.

Geometric structure generation methods are employed to synthesize a wide range of microstructures, including membranes, foams, particle packings for geological and battery-related applications, as well as triply periodic minimal surfaces. Algorithmic approaches enable targeted control over structural features, allowing key properties to be adjusted directly through model parameters.