Toward Brain Digital Twins: integration of biophysical features in virtual brain models of healthy and pathological brain

Computational neuroscience aims to understand how intelligent behaviour emerges from the complex biophysical processes that govern simpler brain functions in animals, from rodents to humans. Artificial intelligence (AI), in turn, studies and develops intelligent agents, i.e., systems designed to perform tasks that require aspects of cognition. These two fields are increasingly interconnected: a deeper understanding of the neural mechanisms underlying natural intelligence can guide the design of better AI models. Natural intelligence emerges from distributed brain activity and causal interactions between brain regions during task-directed behaviour. Therefore, multiscale brain models are needed to capture these interactions and investigate how intelligent behaviour, i.e., emergent functions, arises from complex biophysical processes, thus linking high-level cognitive functions to the underlying neural mechanisms. Within this framework, Brain Digital Twins (i.e., digital counterparts of individual brains) can help bridge the gap between natural intelligence and AI, as they are designed to reproduce and predict brain dynamics. A key enabling concept to model brain activity is the use of generative models (e.g., The Virtual Brain), which, given a subject-specific structural scaffold based on axonal connectivity and a set of interpretable parameters, simulate whole-brain dynamics and generate plausible neuronal activity patterns that can be directly compared with empirical functional data. As a step toward personalised and physiologically meaningful Brain Digital Twins, this thesis investigates how subject-specific biophysical features extracted from MRI can be integrated into virtual brain models. In particular, diffusion MRI and myelin-sensitive MRI provide non-invasive proxies of tissue microstructure and white-matter organisation, which can be used to better constrain inter-regional coupling and signal propagation in whole-brain simulations. Accordingly, I combined AI-based image enhancement with microstructure-informed features to develop a multiscale framework for simulating complex brain dynamics in both healthy and pathological brains, with the overarching goal of improving the physiological accuracy of current virtual brain models. Overall, this thesis provides a methodological pathway towards more physiologically grounded and personalised whole-brain simulations. By combining AI-based super-resolution of diffusion MRI with subject-specific microstructural properties, the proposed framework improves the realism of large-scale virtual brain models. These contributions support the development of more accurate Brain Digital Twins and contribute to the long-term convergence between computational neuroscience and AI. This thesis is divided into seven chapters. Chapters 1, 2, and 3 provide the methodological background for the work, spanning MRI, AI-based image enhancement, and large-scale brain modelling. Chapter 1 introduces the MRI techniques used in this thesis, describing their physical principles, acquisition protocols, and the modelling approaches adopted to extract quantitative features. Chapter 2 introduces Image Quality Transfer, a random forest–based framework that learns a mapping from low- to high-quality images and applies this mapping to enhance lower-quality data. Importantly, since Image Quality Transfer is one of the two main approaches for the methodological baseline of this thesis, a brief review of the state of the art is provided to contextualise the proposed work. Chapter 3 introduces multiscale brain modelling, including the mesoscopic neural mass model used here (i.e., the Wong–Wang model) and The Virtual Brain, the neuroinformatics platform employed in this thesis to simulate large-scale brain dynamics. Chapters 4 and 5 present the main original contributions of this thesis. Chapter 4 presents the first main contribution of this thesis: the spatial resolution of diffusion MRI data is enhanced using Image Quality Transfer, a newly developed machine learning–based framework. First, to improve accuracy and extend the applicability of the method, I adapted the Image Quality Transfer framework to enhance the resolution of diffusion-weighted images acquired with ultra-high b-values and multiple diffusion times. Once Image Quality Transfer was trained and validated, it was applied to reconstruct diffusion-weighted images at higher isotropic resolution. Then, the quantitative and histological accuracy of the reconstructions was demonstrated by using the higher-resolution diffusion-weighted images for advanced biophysical modelling and comparing them with the BigBrain atlas, a high-resolution silver-staining histological reference of the human brain. Second, Image Quality Transfer was evaluated on clinical data acquired using standard-performance scanners, and the quantitative accuracy of the reconstructed data was assessed using diffusion signal–based and biophysical modelling approaches. Overall, this section evaluates the clinical applicability of Image Quality Transfer, representing an important step toward the translation of Image Quality Transfer super-resolution methods from research environments to routine clinical practice. Chapter 5 presents the second main contribution of this thesis: brain models are built using The Virtual Brain, integrating subject-specific, biophysically meaningful features derived from multi-shell diffusion and myelin-sensitive MRI as a priori parameters, with the aim of advancing personalised virtual brain modelling and Brain Digital Twin technologies. First, I linked Image Quality Transfer to macroscale activity simulation by using higher-resolution diffusion-weighted images to provide the structural backbone constraining large-scale brain dynamics. I evaluated whether higher-resolution tractograms influenced The Virtual Brain simulations, particularly in terms of the predictive accuracy of virtual brain models. Second, I enhanced the simulation framework by integrating more biologically meaningful measures of structural connectivity together with subject- and tract-specific conduction velocities derived from MRI metrics, thereby overcoming key simplifications of standard virtual brain modelling. Then, the impact of these microstructure-informed connectomes on simulated brain dynamics was assessed by testing whether The Virtual Brain ability to reproduce subject-specific brain activity improved in comparison to the standard framework. This approach supports the need for more accurate subject-specific simulations to advance Brain Digital Twin technologies. Chapter 6 presents an overall discussion of the thesis highlighting the possible improvements and future perspectives, and summarizes the main conclusions of this study. Chapter 7 summarizes additional research activities undertaken during the PhD program, which complement the core work of the thesis.