5th Dutch Inverse Problems Meeting
The meeting took place on 30-31 October 2025 at the UPark Hotel, Enschede. The program consisted of a masterclass on Deep Learning for (medical) image processing by Jelmer Wolterink, a tour of the imaging labs on campus (thursday) and scientific talks (friday), including a plenary talk by Benedikt Wirth and an industry-talk by Alexandru Onose (ASML).
Program
Thursday
10:30 - 11:30: Masterclass part I - Fundamentals by Jelmer Wolterink
This session will introduce the principles of deep learning in the context of medical imaging. We will cover the key components of modern neural networks, training strategies, and performance evaluation, with a focus on challenges specific to inverse problems in imaging. Examples will highlight how domain knowledge, such as physics-based modeling and prior information, can be integrated into network architectures to improve interpretability, robustness, and generalization. By the end of the session, participants will have a clear understanding of the building blocks needed to develop and critically assess deep learning methods for medical image analysis.
11:30 - 12:30: Masterclass part II - Applications by Jelmer Wolterink
In this session, we will explore challenges in applying deep learning to medical imaging in real-world settings. We will present two case studies: reconstruction of (photo)acoustic images, and segmentation of anatomical structures in 3D MRI and CT volumes. Beyond these applications, we will examine the broader methodological tension between two approaches: (1) large-scale, pre-trained foundation models that promise versatility but require extensive computational resources, and (2) methods that exploit the underlying geometry and physics of the imaging process to achieve data-efficient training. We will discuss how these strategies can complement each other, and what trade-offs practitioners face when bringing AI solutions from the lab into clinical practice.
Friday
09:30 - 10:30: On MAP estimates and source conditions for drift identification in SDEs by Benedikt Wirth (U. Münster)
Parameter identification in stochastic differential equations (SDEs) is a nonlinear inverse problem whose forward operator and log-likelihood can typically be explicitly computed, thereby allowing maximum a posteriori estimates to reconstruct the parameters. To analyse the convergence and convergence rates of these reconstructions, one usually requires source conditions. We will review some of these and discuss the example of drift identification from observing stochastically moving particles. This is joint work with Daniel Tenbrinck, Nikolas Uesseler, Philipp Wacker.
11:00 - 11:30: Reducing acquisition time and radiation damage: data-driven subsampling for spectro-microscopy by Maike Meier (RUG)
Spectro-microscopy is an experimental technique which can be used to observe spatial variations in chemical state. While this is a powerful method, the technique is often limited by factors such as long acquisition times and radiation damage. In this talk, I will present two measurement strategies that exploit structural properties of the data to significantly shorten experiment times and doses applied, in some cases reducing the amount of data required for a measurement to less than 5% of a conventional measurement. The strategies are based on taking only a small subset of all the measurements (e.g. sparse acquisition or subsampling), and then computationally inpainting all unobserved measurements. The methods are data-driven, using spectral and spatial importance subsampling distributions to identify important measurements, and are rooted in modern numerical linear algebra tools, such as the CUR decomposition.
11:30 - 12:00: From Light to Structure: Tackling Inverse Problems in Nanoscale Imaging by Wiebke Albrecht (AMOLF)
At the nanoscale, a material’s functionality is intimately tied to its structure, but directly imaging such structures in realistic environments remains a major challenge. In this talk, I will present how we address this challenge in my group by turning optical spectra into structural fingerprints, essentially solving the inverse problem of inferring 3D nanoparticle morphology from measured light scattering. This requires bridging electromagnetic simulations, electron microscopy, and machine learning to disentangle the “uniqueness problem,” where different shapes can yield similar spectra. A second frontier is dynamic imaging: using in situ transmission electron microscopy with integrated laser excitation, we directly visualize how light drives atomic rearrangements in real time. A central challenge across these directions is to extract quantitative structural information with atomic resolution and ultrafast temporal sensitivity from the limited and noisy signals that such experiments inevitably provide.
12:00 - 12:30: Physics-Trained Neural Networks for Inverse Problem Solving in Geophysical Data Processing by Jing Sun (TU Delft)
Inverse problems are central to geophysics, where the ultimate goal is to estimate subsurface properties from indirect measurements. This process begins with data processing, which provides the foundation for reliable results. Rather than relying solely on data-driven learning, primarily supervised approaches that require training pairs with ground-truth data, which are usually unavailable in real-world geophysical acquisitions, we embed established physical theories into the learning objective of a deep neural network (DNN). This strategy allows the DNN to estimate solutions without the need for ground-truth data while ensuring consistency with the governing physics. We demonstrate the approach with examples from seismic surface-related multiple elimination and potential-field downward continuation.
14:00 - 14:30: Optimal experimental design with k-space data: application to inverse hemodynamics by Miriam Löcke (RUG)
Parameter estimation from MRI measurements relies on high-resolution data, which requires long acquisition times. While subsampling is thus a standard practice, it remains unclear how to choose sampling patterns for reliable parameter estimation. We are therefore using Optimal Experimental Design (OED) techniques to derive sampling patterns specifically designed for the estimation of blood flow parameters from frequency space MRI measurements. We show the resulting masks as well as the results of the parameter estimation based on synthetic data with these masks compared to conventional masks, and demonstrate a 10x speed-up while maintaining the same accuracy. This is a joint work with Ahmed Attia and Dariusz Ucińsky.
15:00 - 15:30: Inverse problems in the semiconductor industry by Alexandru Onose (ASML)
Modern microprocessors contain billions of transistors packed tightly onto silicon wafers, requiring high manufacturing precision across multiple layers. A specially designed reticle mask is needed to guide the exposure of a photosensitive material (generically called resist) to generate the desired features. There are multiple inverse problems that need to be solved, both for the design of the reticle as well as for associated metrology tasks. In this context, I aim to present a high-level overview of the manufacturing process and highlight some of the inverse problems that arise. I will present the inverse lithography task in more details and discuss some of the challenges that result from the non-convexity and the large-scale nature of the problem.
15:30 - 16:00: Cortisol Dynamics : Identifiability analysis of the model and comparison of drug response by Sayali Bhatkar (RUG)
Cortisol is an important glucocorticoid hormone that plays a critical role in relulating metabolism, immune response, and stress adaptation. In cases of adrenal insufficiency, replicating endogenous cortisol dynamics in patients remains a significant therapeutic challenge. Hence, it is desirable to develop physiology based models to shed light on the the behaviour of drugs/cortisol in human body. We perform identifiability analysis of the model developed by Dorin et al. (2022) that includes the interaction of cortisol with CBG and albumin and also a peripheral compartment to account for the uptake of cortisol from the plasma. We reduce the model based on the unidentifiabilities we encounter. The reduced model is structurally identifiable and can be solved analytically.