Workshop - BICV

Magnetic resonance imaging allows us to look inside the living human body without ionizing radiation, but conventional MRI primarily reveals anatomy and function rather than the underlying chemistry. Magnetic resonance spectroscopy (MRS) goes one step further, allowing us to non-invasive detect molecules involved in brain metabolism and neurotransmission. However, extending MRS to image neurochemistry across entire brain circuits remains technically challenging, particularly at ultra-high magnetic fields.

In this talk, I will introduce the basic physical principles behind MRI and magnetic resonance spectroscopy (MRS), from nuclear spins and radiofrequency excitation to spatial encoding and chemical shift. I will then show how these principles lead to some of the major engineering challenges in ultra-high-field spectroscopic imaging. Using my current research at the University of Oxford as a case study, I will discuss the development of new radiofrequency and parallel-transmission methods for 7 Tesla MRSI, with the goal of simultaneously measuring neurochemical changes across interconnected brain regions during motor learning.

Finally, I will briefly discuss where artificial intelligence and data-driven methods may complement this type of physics-based research, from synthetic data and analysis to pulse-design optimization. More broadly, I will argue that computational tools are most powerful when they are applied to a well-defined scientific problem grounded in an understanding of the underlying physics.

Foundation models are increasingly being applied to neuroimaging, offering rich representations that can transfer across tasks and datasets. However, models developed for natural images or language do not directly account for the three-dimensional organization and specific semantics of brain imaging. Understanding how these representations can be adapted to neuroimaging, and what information emerges in their embedding spaces, is therefore an important challenge.

This talk will present recent work exploring these questions from two complementary perspectives. First, we investigate how a frozen 2D vision foundation model, DINOv3, can be adapted to 3D structural MRI through lightweight spatial learning. Weak positional encoding selectively improves the representation of subcortical anatomy and the anatomical organization of the embedding space, while stronger spatial constraints can disrupt the information inherited from the pretrained model. In parallel, we explore how representations extracted from brain MRI can interact with language models. Concept-based analyses of visual tokens reveal emerging alignment between imaging representations and task-relevant semantic concepts, providing a way to inspect how information is reorganized in a shared visual–language space. Together, these studies illustrate how probing anatomical and semantic organization can help move beyond predictive performance toward more interpretable foundation models for neuroimaging.

Visual representation learning in biomedical imaging faces two compounding challenges that scale-based approaches address only partially: annotation scarcity and interpretability demands. Current methods, including self-supervised and contrastive learning, significantly reduce annotation requirements but still rely on large unlabeled corpora to converge on generalizable feature spaces, and produce representations that remain structurally opaque by design.

We propose that these limitations reflect an architectural choice of representing visual knowledge as statistical regularities over continuous parameter spaces rather than as explicit structural descriptions. As an alternative, we present a hierarchical constraint inference framework in which visual representations are constructed as symbolic automata encoding primitive structural patterns, and higher-order automata (HOA) encoding their compositional relationships. In this approach, rather than learning from data distributions, the system infers constraints directly from pattern structure, constructing complete representations from minimal examples without a training phase. The framework produces fully interpretable, hierarchically organized representations stored as inspectable graph structures. Primitive patterns are encoded as elementary automata capturing spatial and relational constraints. Composite patterns are encoded as HOAs that explicitly reference their component automata, creating a multi-level hierarchy that reflects the natural compositional organization of visual structure. Similarity detection clusters structurally related automata, enabling generalization across variations.

We present proof-of-concept results demonstrating complete hierarchical representation formation from minimal examples, with quantifiable graph metrics tracking structural generalization across multiple levels of abstraction. We discuss how the core principle of constraint inferences over compositional hierarchies translates to pixel-domain biomedical imagery, where annotation is scarce, structural regularity is strong, and interpretability is clinically required, and identify open research questions at the intersection of symbolic structural learning and biomedical computer vision.

Psychiatric disorders are highly complex and heterogeneous conditions that likely originate from cumulative, small-scale changes across genetic and physiological systems. The significant variability in biological and clinical profiles between individuals has complicated efforts to develop reliable disease biomarkers, suggesting the need to transition from rigid diagnostic categories to personalised clinical profiling.

Recent advances in open science, neuroimaging and AI offer a unique opportunity to implement this paradigm shift. One promising approach is normative modelling, which characterises the range of relevant variation in a reference population and quantifies deviations from this normative range, offering a personalised index of atypicality.

This lecture will discuss the potential and risks of using AI to study and manage complex psychiatric disorders, focusing particularly on normative modelling and anomaly detection frameworks as powerful means of capturing patients' unique neurobiological profiles. We will present recent studies that integrate multi-site clinical and neuroimaging data to advance the search for biomarkers of complex diseases such as major depressive disorder and bipolar disorder. This will promote the refinement of AI-powered diagnostics and prognostics in psychiatry.

Recent advances in high-resolution imaging and artificial intelligence are transforming the way biomedical tissues can be characterized across multiple spatial scales. Here, we present an integrated framework combining synchrotron phase-contrast micro-computed tomography with AI-driven image analysis to investigate tissue microarchitecture with unprecedented detail. Using human trabecular bone as a case study, we demonstrate how convolutional neural networks and large-scale morphometric analysis enable the automated quantification of millions of osteocyte lacunae from sub-micron-resolution datasets. The approach reveals disease-associated microstructural remodeling signatures in osteoporosis and post-COVID-19 bone that are not captured by conventional densitometric measures alone. Beyond bone research, the presented workflow highlights the broader potential of combining high-resolution imaging and AI for biomarker discovery, mechanobiology, and precision medicine applications in biomedical engineering.

Artificial Intelligence is rapidly transforming nuclear medicine across the entire clinical workflow. This talk provides an overview of current applications and emerging directions, from image acquisition and reconstruction, including dose reduction, denoising, faster scans, and CT-free PET, to image processing and quantitative analysis, with applications in segmentation, harmonization, and synthetic image generation. We will also discuss the role of AI in diagnosis and clinical decision support, as well as its growing potential in theranostics and personalized medicine, including patient-specific dosimetry, treatment planning, and digital twins. Through representative examples, the talk will highlight both the opportunities and the challenges of translating AI into clinical practice, outlining the transition toward more integrated, quantitative, and personalized nuclear medicine.