Conference Agenda

Clinical biomarker monitoring of vulnerable newborn patients relies on invasive and costly laboratory procedures, increasing the risk for undetected metabolic disbalances and delaying medical assistance. Addressing this critical need, we present a non-invasive, wearable paper-based sensor patch that captures multiple critical body functions via colorimetric analysis of body fluids (sweat and saliva). The sensor facilitates simultaneous measurement of four health biomarkers: temperature (T), pH, sodium (S), and glucose (G), on a miniaturized surface integrating twelve different colorimetric inks.

The sensing wearable patch is supported by a video-recording setup and an artificial intelligence (AI) imaging system with two main goals: (1) to enable accurate digital color-based parameter reading, accounting for the sensor deformation and applying color correction to compensate for the uncontrolled light conditions of the clinic, and (2) perform sensor tracking and segmentation on a simulated moving patient in the neonatal incubator (Fig. 1).

The sensor in-vitro colorimetric dye responses showed high precision and reproducibility in sensing ranges relevant for neonatal care (T: 32-41 [°C]; pH: 3-9; S: 2.92-29.20 [mg/mL], G: 0.039-0.625 [mg/mL]). We propose a U-Net denoising autoencoder and latent feature regression, for quantification of the colorimetric response and simultaneous color correction. The model achieved high precision for automated parameter measurement (T:0.455 [°C]; pH:0.416; S:0.857 [mg/mL], G:0.019 [mg/mL]; mean absolute error) (Fig. 2). High performance was also found for tracking and segmentation of the sensing patch (Mask-RCNN) in a testbench with baby dummies in a neonatal incubator (0.986 AP @ IoU=0.5) (Fig. 3).

The sensor’s optimized interface for biofluid sample collection and multi-biomarker measurement on skin, together with the high AI imaging system performance under clinically relevant conditions, demonstrate the feasibility of an AI-enabled colorimetric sensor that caters to the critical needs of non-invasive newborn health monitoring in the clinic. This novel medical technology offers a low-cost, easily applicable, miniaturized solution, delivering critical information at the point of care, thereby aligning with the emerging landscape of digital health technologies while expanding access to high-quality care in the neonate.

Multi-contrast MRI sequences acquire multiple images with different tissue contrasts in a single scan, offering rich insight into tissue microstructure. However, their clinical use is limited by long acquisition times. A common strategy to reduce scan duration is k-space undersampling, though this makes image reconstruction more challenging. Advanced reconstruction approaches, including compressed sensing and deep learning, have shown promise in mitigating these issues.

In this work, we leverage shared anatomical information across contrasts to further increase acceleration. We use complementary variable density Poisson disk sampling [1] that fully samples the k-space center of each contrast for reliable contrast encoding while applying complementary undersampling patterns across contrasts for the high-frequency components. To reconstruct the resulting highly undersampled data, we propose an implicit neural representation (INR) network that jointly reconstructs all contrast images, effectively exploiting the complementarity of the acquired data [2]. Our INR-based method is self-supervised and trained individually for each scan, eliminating the need for external training data, an important advantage in multi-contrast MRI, where large training datasets are often unavailable.

We demonstrate the effectiveness of our approach on multi-contrast MRI data acquired with a cartesian 3D MPnRAGE sequence [3] and show that our INR-based reconstruction outperforms both traditional and deep learning–based reconstruction techniques. With our approach, the scan time at 1.5 mm resolution can be reduced from originally 22 minutes down to only 3.25 minutes.

Our code is available at: https://github.com/compai-lab/2025-miccai-niessen

References

1. 3D Cartesian MRI with compressed sensing and variable view sharing using complementary poisson-disc sampling. Levine E, Daniel B, Vasanawala S, Hargreaves B, Saranathan M. 2017 May. pp. 77(5):1774-1785.
2. INR Meets Multi-contrast MRI Reconstruction. Niessen N, Pirkl CM, Solana AB, Eichhorn H, Spieker V, Huang W, Sprenger T, Menzel MI, Schnabel JA. s.l. : RIME and GRAIL workshops MICCAI 2025, pp. 23–33, Springer-Verlag, Berlin, Heidelberg, 2025, 2025.
3. Highly accelerated 3D MPnRAGE meets implicit neural representation (INR) reconstruction. Niessen, N., Solana, A. B., Sprenger, T., Pirkl, C. M., Eichhorn, H., Schulte, R. F., Menzel, M. I., and Schnabel, J. A. on behalf of the PREDICTOM consortium. s.l. : ISMRM Medicine conference proceedings, 2026.

Diffusion Spectrum Imaging (DSI) samples the MRI signal in three-dimensional q-space and enables the reconstruction of diffusion distributions in biological tissue. Dense sampling of q-space requires a large number of measurements and results in long acquisition times. Methods for accelerating DSI acquisition have therefore been proposed. Reducing the number of required measurements while preserving the diffusion signal is therefore a central objective in the acceleration of DSI. Recent work has explored machine learning approaches for DSI reconstruction from incomplete measurements. Since the DSI signal is complex-valued and contains both magnitude and phase information, reconstruction methods must account for the complex signal representation.

A complex-valued residual encoder–decoder architecture is used to reconstruct fully sampled q-space volumes from undersampled measurements. The network processes volumetric DSI data using complex-valued 3D convolutions and residual blocks. Training is based on a weighted combination of loss terms defined on complementary properties of the complex signal, including amplitude error, wrapped phase error, logarithmic magnitude error, global energy consistency, and total variation regularization.

Experiments were conducted on 10 000 synthetic diffusion datasets. The simulated signals follow the methodological approach of previous work and are based on multi-tensor diffusion models originally introduced in earlier studies. The underlying diffusion process is modeled using Brownian motion within a voxel according to the diffusion MRI signal model. Fully sampled q-space volumes were used as reference data, while degraded inputs were generated through controlled undersampling and by adding complex Gaussian noise with noise levels randomly sampled from predefined ranges (0.2–0.4 for low noise and 0.7–0.9 for high noise). Training was performed using the Adam optimizer.

Reconstruction remains stable down to a keep fraction of approximately 25%. For a keep fraction of 75%, the structural similarity index (SSIM) is greater than or equal to 0.98 while the global error remains below 5%. When the keep fraction is reduced to 50%, the SSIM remains greater than or equal to 0.96 and the global error stays below 8%. At a keep fraction of 25%, the SSIM is still greater than or equal to 0.95 with a global error below 10%. Even at a very low keep fraction of 10%, the SSIM remains around 0.88 while the global error is still below 10%.

Hematological diagnosis requires assessing cell population composition, such as blast-to-mature ratios or co-occurring dysplastic lineages. Current Multiple Instance Learning (MIL) pipelines discard these relationships: mean pooling treats cells as independent, and Attention-Based MIL (ABMIL) learns weights without modeling interactions. Graph Neural Networks (GNNs) can capture such structure, but cytology lacks a predefined graph topology; the graph must be inferred from learned features.

We present HematoGraph, a framework that jointly learns a latent cell graph for slide-level classification. Our main contribution is the first application of Continuous Differentiable Graph Models (cDGM) to hematology: cDGM parameterizes a soft adjacency with learned temperature and threshold, yielding a fully differentiable graph optimized end-to-end. We pair cDGM with Adaptive Convolutions on Graphs (ACM) message passing and graph-aware Differentiable Pooling (DiffPool) aggregation, where a Graph Convolutional Network (GCN) based assignment groups neighboring cells into clusters before pooling. We benchmark against heuristic k-Nearest Neighbors (k-NN), discrete DGM with Gumbel top-k sampling, Snowflake variants, and Poincaré ball embeddings, combined with GCN, Graph Attention Networks (GAT), and five aggregators: mean, ABMIL, DiffPool, virtual node, and Graph Multiset Transformer (GMT). We ablate components on a 4-class subtyping task—Acute Myeloid Leukemia (AML), Myelodysplastic Syndromes (MDS), Myeloproliferative Neoplasms (MPN), and Normal—and evaluate the configuration across 8 tasks. Ablations validate each choice: cDGM yields the best topology (86.0% vs. 82.4% without graph), ACM outperforms GCN (83.6%) and GAT (78.8%), and DiffPool (86.0%) surpasses ABMIL (84.9%) and GMT (81.3%). Across all tasks, cDGM+ACM+DiffPool significantly outperforms mean pooling (+3.5 pp, p=0.025, winning 7/8 tasks) and ABMIL on 6/8 tasks (+1.5 pp), with gains up to +20.6 pp on AML/Acute Lymphoblastic Leukemia (ALL) and +5.9 pp on Dresden morphology. On AML Hehr, ABMIL retains a +5.7 pp edge, suggesting attention suits tasks where signals lie in rare cells rather than communities. Graph-aware hierarchical pooling via cDGM+ACM+DiffPool provides a principled alternative to standard MIL for cytology. DiffPool is the only aggregator systematically benefiting from graph structure (p<0.05 for 7/9 graph learners); ABMIL and GMT degrade when graphs are introduced, indicating structure helps only when the aggregator can exploit topology

Instance segmentation in crowded scenes requires both high object capacity and precise boundary modeling. Recent promptable foundation models offer strong generalization, but their automatic mask generation can be unreliable outside the training domain and their computational cost scales unfavorably with very large numbers of instances. We present CPNv2, a modernized contour-based instance segmentation framework that combines dense proposal prediction with learned, sub-pixel accurate refinement. CPNv2 introduces centroid-based Hungarian target assignment to reduce ambiguous supervision on low-resolution output grids, iterative score refinement for stable confidence estimation, differentiable feature sampling via bilinear interpolation for continuous contour updates, and deep contour refinement with a contour refinement transformer trained with contour denoising. Across our experiments, CPNv2 consistently improves over the previously established CPN architecture and achieves strong results across diverse scientific imaging datasets in terms of F1@50 and PQ@50 scores. At the same time, it preserves a key strength of contour-based segmentation: among the evaluated methods in our experiments, contour-based models show the highest efficiency in terms of latency, memory, and output size. This makes CPNv2 particularly attractive for high-throughput settings, where large images with many instances must be processed efficiently without sacrificing segmentation quality. When benchmarking recent SAM variants we find that even under ground-truth prompting, they do not consistently match the strongest task-specific models on our datasets. Code and trained models will be made publicly available.

High-quality labels are of utmost importance in biomedicine and instrumental for tasks such as disease understanding, drug discovery, and medical diagnosis. Human-in-the-loop approaches and AI-assisted annotations are currently widely accepted as the safest approach for data curation, offering both speed and human supervision, while lever-aging the power of AI. Indeed, foundation models are often used to ob-tain initial labels which are then manually corrected by a human expert and subsequently used for downstream tasks. In this work, we uncover a rarely discussed risk of such semi-automated approaches and present a case study to demonstrate this. Focusing on microscopy imaging seg-mentation, where annotation quality is critical and costly, we collected and annotated a Transmission Electron Microscopy dataset in a semi-automated approach, using BATS, a software we built for AI-assisted segmentation labeling in microscopy which encourages data centric prac-tices. We show how batch effects can be derived by a mix of human and modelannotationsaffectingthedownstreamanalysis.Finally, wedemon-strate a solution which mitigates these batch effects and discuss how future researchers can adopt responsible and accurate data annotation pipelines.

Acute ischaemic stroke remains the predominant cause of disability and a major contributor to mortality globally. The damage caused by ischemia, triggered by vascular occlusion, progresses rapidly, with brain tissues beginning necrosis within minutes. This necessitates urgent clinical decision-making. This is particularly important for reperfusion therapies such as intravenous thrombolysis and mechanical thrombectomy. The efficacy of these treatments is time-sensitive and associated with risk of intracranial hemorrhage. In addition, treatment efficacy varies considerably across patients and depends upon several factors. This underscores the need for individualized risk stratification.

In this project, we aim to develop a deep learning framework for predicting 90-day functional outcomes using multimodal data acquired at hospital admission. We are integrating heterogeneous data sources across 25 different hospitals within Germany (German Stroke Registry data), including clinical scores, patient history, and MRI images. Traditional centralized learning approaches face limitations related to data privacy, small sample sizes, and institutional barriers. Swarm Learning provides a privacy-preserving, decentralized alternative. However, a decentralized pipeline for multimodal MRI-based model development is currently lacking. To ensure robustness and generalizability, we will implement model training in a decentralized swarm learning manner. Our objective is to validate a multimodal deep learning model for individualized 90-day outcome prediction in a decentralized AI infrastructure. We aim to advance precision medicine for acute stroke care and establish a foundation for globally collaborative, privacy-preserving AI development.

Graph Neural Networks are powerful machine learning techniques for processing sparse data with irregular geometries, as encountered in high-energy physics detectors. However, deploying such models within hardware triggers remains challenging due to stringent real-time constraints in terms of both latency and throughput. State-of-the-art hardware triggers in collider experiments impose hard latency deadlines on the order of 1 to 10 microseconds, necessitating the development of custom machine learning accelerators based on Field Programmable Gate Arrays.

This work presents a deployment methodology for mapping Graph Neural Networks onto such platforms. By implementing commonly used neural network operators as reusable architecture templates, and leveraging model quantization and pruning, our approach achieves microsecond inference latencies. We demonstrate the methodology by deploying a Graph Neural Network based clustering algorithm for the Electromagnetic Calorimeter of the Belle II experiment. Through hardware-algorithm co-design, we achieve an end-to-end system latency of 1.050 microseconds, while preserving clustering quality, and meeting the real-time constraints required for hardware triggers. We validate our approach through cycle-accurate simulation and direct deployment on hardware, achieving complete agreement between simulation and measured results. Furthermore, we investigate the use of heterogeneous System-on-Chip architectures, such as AMD Versal platforms, as a path to deploy even larger neural network models.

To conclude, this work establishes a deployment methodology for graph-based machine learning inference under extreme real-time constraints.

Atomistic simulations are at the forefront of materials discovery, yet their complex setup and debugging often require specialized expertise, limiting the widespread adoption of Integrated Computational Materials Engineering (ICME). To bridge this critical know-do gap, we introduce GENIUS, a novel AI-driven framework that autonomously designs and executes simulation protocols. GENIUS seamlessly integrates a smart knowledge graph tailored for Density Functional Theory (DFT) calculations with a hierarchical architecture of advanced Large Language Models (LLMs), supervised by a robust finite-state error-recovery machine. Focusing initially on DFT, GENIUS translates human-generated prompts into validated input files, achieving successful execution on a diverse set of 295 benchmarks. A key strength of GENIUS is its autonomous error handling, which repairs errors in 76% of failed runs, significantly boosting reliability. Compared to LLM-only baselines, GENIUS halves inference costs and virtually eliminates the 'hallucinations' that can lead to incorrect results. By intelligently automating protocol generation, validation, and repair, GENIUS democratizes access to electronic-structure simulations, enabling researchers to focus on scientific discovery rather than computational complexities. This framework empowers large-scale materials screening, accelerates ICME design loops, and promotes innovation across academia and industry by bridging the gap between experimental work and simulations, democratizing advanced simulation methods for users lacking extensive computational experience.

Foundation models have progressed by scaling parameters and training data, driving rapidly increasing computational demands. This trend is especially pronounced in scientific imaging, where datasets can span terabytes to petabytes. Exascale systems provide the compute to train models at this scale. However, using these machines efficiently is non-trivial. At extreme scale, bottlenecks shift from GPU throughput to end-to-end workflow behavior, including startup overheads, storage access, communication, and synchronization.

We study these effects and derive practical scaling lessons from a real training pipeline. This work was carried out on the JUPITER system at Jülich Supercomputing Centre within the JUPITER Research and Early Access Program (JUREAP) and the GCS Exascale Pioneer project brainfm. We adapt the execution environment and I/O path to reduce indirect I/O and metadata pressure caused by containers, runtime-generated artifacts, and logging. In parallel, we evaluate model- and loss-level choices that reduce synchronization and collective communication. To quantify data access performance, we compare HDF5 and Zarr for highly concurrent random access across file layouts and backends. We demonstrate the approach by training a neuroscience vision foundation model with contrastive learning on terabyte-scale microscopic images of histological human brain sections (CytoNet, https://arxiv.org/abs/2511.01870). We scale the workflow up to 16,384 NVIDIA GH200 superchips across 4,096 compute nodes.

Across large runs, indirect I/O emerges as a primary scalability limiter, driven by container image access, startup scripts, bytecode generation, temporary-directory traffic, and uncontrolled logging. Staging container images into node-local memory and redirecting runtime-generated files away from shared storage reduces filesystem metadata storms and improves startup robustness. On the algorithmic side, synchronization-heavy components constrain scaling, motivating architecture choices that avoid batch-level collectives (e.g., batch normalization) and a contrastive-loss implementation that reduces redundant per-rank compute while limiting collective communication. For highly concurrent data access, we find that Zarr with the TensorStore backend provides the lowest and most stable access times.

We distill these findings into practical guidelines that link workflow engineering, model design, and storage choices for training scientific foundation models at extreme scale.

In this talk, we first present a novel approach and benchmark for long-horizon robotic manipulation. Our method integrates the high-level task planning capabilities of Large Language Models (LLMs) with the precise object grounding of Vision-Language Models (VLMs). Given a detailed grounded plan, a 3D low-level motion planner executes actions conditioned on natural language. While our approach demonstrates excellent performance in real-world settings, robust manipulation also requires reliable error handling. We introduce a framework for detecting planning and execution failures, highlighting the critical role of high-quality training data. In challenging long-horizon tasks, this failure detection and recovery mechanism significantly improves overall system performance.

In the second part of the talk, we explore learning vision-based policies for multi-fingered robot hands using human video demonstrations. Our method employs reinforcement learning with trajectory-guided rewards and unified visual policy training. Experiments in both simulation and real-world environments demonstrate that our approach outperforms state-of-the-art methods across three complex dexterous manipulation tasks. Because training these models requires high-fidelity visual data, we propose a method to generate temporally dense and consistent captions and object groundings. We conclude by showing that a model trained on a large-scale automatically annotated dataset achieves state-of-the-art results for this task.

Introduction of research at HMGU, as Program Spokesperson

Osteoarthritis is one of the leading causes of disability and pain worldwide, with over 300 million people affected. Currently no curative treatments are available. A detailed understanding of disease aetiopathology and novel drug targets are therefore urgently needed.

In this talk, I will give an overview of how we have used translational genomics approaches to enhance our understanding of the genetic aetiology of osteoarthritis, shed novel biological insights, and provide a stepping stone for translating genetic associations into osteoarthritis drug development.

Optoacoustic imaging has the unique ability to visualize optical contrast deep in biological tissue dynamically and at high resolution without using ionizing radiation. Clinical translation of the modality into clinical workflows depends not only on advances in instrumentation, but also on computational methods that can cope with limited data, complex image formation, and application-specific biological variability. While AI in imaging is often identified with large-scale deep learning, optoacoustic applications frequently require a broader computational perspective in which learning-based methods are combined with biophysical models and probabilistic inference.

Our group develops such methods for optoacoustic systems on multiple scales spanning chemically specific microscopy, mesoscopic high-resolution imaging of tissues, and macroscopic systems capable of imaging at depths of several centimeters. We combine models of system physics, tissue biology, and disease effects with Bayesian variational inference and use AI selectively where it offers clear benefit, including algorithm acceleration and generative prior modeling. In multispectral optoacoustic tomography (MSOT), we enable high image quality in real time through deep-learning-accelerated reconstruction and denoising, and we advance quantitative optoacoustics through hierarchical Bayesian models and variational inference. In mid-infrared optoacoustic microscopy (MiROM), we reduce acquisition time through sparse measurements, while compensating information loss using spatial correlation priors and uncertainty quantification.

These computational tools act as enabling technologies for translational optoacoustics. They deliver high-quality images on systems, support quantitative and uncertainty-aware readouts, and move advanced sensing and imaging workflows toward clinically relevant timescales. More broadly, our work highlights a practical view of AI in biomedical imaging: not as a purely data-driven replacement for modeling, but as a targeted component within biophysics-informed computational pipelines that bring optoacoustic imaging and sensing closer to clinical impact.

Understanding how complex biological systems give rise to functional health and disease remains a major challenge, particularly in tissues where multiple interacting factors shape emergent behavior. In the human airway, mucociliary clearance is a key defense mechanism whose dysfunction underlies many respiratory diseases, yet its regulation across molecular, cellular, and mechanical scales is poorly understood.

Here, we combine organotypic human airway models with systematic perturbation screens and multimodal readouts, including immunofluorescence staining, functional clearance assays, and single-cell RNA sequencing. Using these data, we are in the process of developing an AI-based predictive framework (CellFlow) that maps how combinations of developmental, inflammatory, and physiological factors regulate epithelial differentiation and clearance function.

This approach will enable prediction and experimental validation of previously untested perturbations, revealing regulatory logic that cannot be inferred from single-factor studies. More broadly, our work aims to demonstrate how AI-integrated experimental systems can transform biological models into predictive platforms for studying disease mechanisms and guiding therapeutic strategies.

The virtual cell to model and predict cellular behavior and perturbation responses is a defining ambition of AI-biology. Two powerful paradigms are driving progress: sequence-based models predict molecular properties bottom-up, whereas transcriptomic approaches represent cell states and forecast trajectories. Yet between both is a layer that neither reliably reaches: the physical, conditional molecular networks that execute molecular programs.

Despite the transformative impact of AlphaFold on structural biology, accurately predicting binary protein-protein interactions remains an unsolved problem. Drawing on recent work I will show where current structure-based AI models systematically fail in interaction prediction, and why large-scale experimental mapping remains indispensable, not as a stopgap, but as the empirical foundation for a biochemistry that models cannot yet replace.

This experimental foundation becomes concrete in the context of microbe-host biology. The gut microbiome is known to influence the trajectory of numerous complex diseases, yet the molecular mechanisms remain largely obscure. We showed that commensal bacteria deploy a molecular machine only known from pathogens to inject bacterial proteins into host cells. We mapped the interactions of bacterial and human proteins and integrated these experimental data with structural modeling and network analysis. While AlphaFold modeling identified specific functional hypotheses that are currently being followed up, network analysis revealed that effectors target host proteins linked to autoimmune and metabolic disease, and that effector abundance is specifically elevated in Crohn's disease.

Genetic variation represents another highly consequential route by which molecular networks are perturbed in disease. Translating genetic associations into mechanisms requires knowing how risk variants are organized within these networks. To link GWAS signal to mechanisms we developed a graph representation learning framework integrating network topology with multi-modal genetic evidence to predict core disease genes — validated against Mendelian disease genes and enriched for druggable, untargeted proteins and setting the foundation for further mechanistic understanding.

Achieving the virtual cell vision will require knowledge of molecular wiring in cells and understanding how that wiring is remodeled by genetic variation or hijacked by viruses and bacteria.