Introduction to Generative Modelling with Flows and Diffusions: From Theory to Application in Unsupervised Anomaly Detection in Neuroimaging
Wednesday | April 8, 2026 | 8:30 – 17:00
This tutorial introduces generative modelling with flows and diffusions, two key frameworks for learning data distributions in high-dimensional spaces. Flows, defined by deterministic ordinary differential equations (ODEs), and diffusions, defined by stochastic differential equations (SDEs), describe data generation as probability transport from random noise to realistic images through a time-dependent vector field. The first part builds a rigorous yet intuitive understanding of these models, explaining how they relate and how their training objectives are derived. The second part is a hands-on session. Participants will run a diffusion model in PyTorch, explore its main components (U-Net, time embeddings, variance schedule), and apply it to unsupervised tumor detection in brain MRI using the AnoDDPM framework for pseudo-healthy image reconstruction. The tutorial combines mathematical insight with practical implementation, giving biomedical imaging researchers both the theoretical background and coding experience needed to apply modern generative models in neuroimaging.
Attendees will leave with a principled workflow for auditing, triaging, and remediation of data issues, a map of tools with trade-offs in accuracy, compute, and scalability, as well as reproducible materials to apply immediately in their projects.
Francesca Galassi
Univ Rennes, CNRS, Inria, Inserm, IRISA UMR 6074, EMPENN – ERL U 1228, F-35000 Rennes, France
Hugues Roy
Sorbonne Université, Institut du Cerveau – Paris Brain Institute – ICM, CNRS, Inria, Inserm, AP-HP, Hôpital de la Pitié Salpêtrière, F-75013, Paris, France, 187 rue du Chevaleret
Maëlys Solal
Sorbonne Université, Institut du Cerveau – Paris Brain Institute – ICM, CNRS, Inria, Inserm, AP-HP, Hôpital de la Pitié Salpêtrière, F-75013, Paris, France, 187 rue du Chevaleret
Alireza BAGHAI-WADJI
University of Cape Town
Customized Multiresolution Analysis and Learning Algorithms in Biomedical Imaging
Wednesday | April 8, 2026 | 8:30 – 17:00
The tutorial aims to provide a comprehensive experience of different aspects of image formation in magnetic resonance imaging. Participants learn about This tutorial dissects the structural complexity of multiresolution analysis (MRA), delves into the foundations of feedback loops embedded in learning algorithms (LAs), and discusses current challenges in cutting-edge biomedical imaging (BMI). It pursues an amalgamation of classical computational signal and data processing and quantum algorithms (QAs). It proposes a collaborative approach to customizing algorithm design for software development in clinical applications within the ISBI community. The foundations of wavelets, frames, curvelets, edgelets, and related topics in BMI are clearly presented and illustrated graphically. To incorporate biology-inspired LAs into the design of MRA, feedback loops in standard and advanced LAs are scrutinized and presented diagrammatically. The anatomies of error-correction, memory-based, Hebbian, competitive, Boltzmann, credit assignment, and stochastic LAs are analyzed. Applications in BMI, image codification and compression, and the growing role of AI are discussed. The tutorial is self-contained and packed with clear arguments. Upon completion of the tutorial, the participants are invited to discuss open problems, initiate collaborative ties, and join an existing voluntary mentorship program. The customized design of biology-driven MRA tools should appeal to aspiring students, early-career scientists, professionals seeking continued education, and established researchers alike. A well-designed book-style manuscript will be made available.
Daniel Christopher Hoinkiss
Fraunhofer MEVIS
From Sequence Diagrams to Medical Images – A Comprehensive and Interactive Tutorial on Image Formation in Magnetic Resonance Imaging
Wednesday | April 8, 2026 | 8:30 – 12:00
The tutorial aims to provide a comprehensive experience of different aspects of image formation in magnetic resonance imaging. Participants learn about different types of MR sequences and their respective adjustments using the interactive gammaSTAR framework. Through a built-in MR simulator and image reconstruction framework, the effects of changing various protocol parameters on generated images are directly experienced and potential sources of image artefacts are revealed. The interactive session is accompanied by in-depth slides describing the process of image reconstruction from the simulated raw data.
Learning with Covariance Matrices: Foundations and Applications to Network Neuroscience
Wednesday | April 8, 2026 | 8:30 – 12:00
Covariance matrices are ubiquitous in biomedical signal processing, particularly in modalities involving spatially distributed data, such as magnetic resonance imaging (MRI). In network neuroscience, anatomical covariance matrices and functional connectomes model structural and functional interdependencies across brain regions. Principal component analysis (PCA), which relies on the covariance spectrum, has been a mainstay in biomedical data analusis. However, PCA-based methods face challenges including limited reproducibility and constrained applicability to a fixed set of features, which restrict their generalizability. In contrast, modern deep learning (DL) models—especially those specialized for graph-based data—offer viable tools to overcome these limitations. This tutorial will introduce a novel family of DL models called coVariance Neural Networks (VNNs), which operate directly on covariance matrices. By treating covariance matrices as weighted graphs, VNNs integrate insights from PCA, graph signal processing and graph neural networks to provide principled learning architectures. The attendees will be introduced to the foundational theory behind VNNs drawn from fundamental notions in signal processing and appreciate their practical benefits in neuroimaging and biomedical signal processing, focused on enhanced reproducibility, generalizability, and explainability. The tutorial will be of broader interest to researchers seeking principled integration of statistical and deep learning methods in biomedical signal processing.
Saurabh Sihag
University at Albany
Gonzalo Mateos
University of Rochester
Elvin Isufi
Delft University of Technology
Alejandro Ribeiro
University of Pennsylvania
Fabian Gröger
University of Basel
Modern Data Cleaning
Wednesday | April 8, 2026 | 13:30 – 17:00
This tutorial presents modern, hands-on strategies for cleaning and curating biomedical image datasets. We review the phenomenology of data quality issues (e.g., near duplicates, off-topic/outlier samples, label errors), connect them to their impact on evaluation validity and clinical translation, and position data cleaning as complementary to robust learning with noise. We then cover practical detection methods and open-source libraries, with guided notebooks and a bring-your-own-data session.
Attendees will leave with a principled workflow for auditing, triaging, and remediation of data issues, a map of tools with trade-offs in accuracy, compute, and scalability, as well as reproducible materials to apply immediately in their projects.
Marta Varela
City St George’s University of London & Imperial College London
BioMedPINNs: Successfully using Physics-Informed Neural Networks in Biomedical Applications
Wednesday | April 8, 2026 | 13:30 – 17:00
Physics-Informed Neural Networks (PINNs) combine the strengths of machine learning with the interpretability of mathematical models, offering powerful new ways to model complex biological systems. In biomedical imaging, where data are often scarce and governed by well-understood physics, PINNs can bridge the gap between data-driven AI and mathematical models of physics and physiology.
This tutorial will provide a practical and accessible introduction to PINNs, tailored to imaging researchers. Through a mix of lectures and hands-on coding sessions, participants will learn how PINNs can be used to solve imaging challenges such as MRI perfusion quantification and myocardial fibre orientation mapping. The session will also cover advanced extensions to improve PINN performance and situate PINNs alongside other emerging methods, such as Implicit Neural Representations.
By the end of the tutorial, attendees will not only understand the principles behind PINNs but will also leave with ready-to-use code, practical skills, and a clear roadmap to apply these methods to their own biomedical imaging problems.