Arterial Spin Labeling (ASL) is a non-invasive magnetic resonance imaging technique that enables quantitative assessment of cerebral perfusion by using magnetically labeled arterial blood water as the endogenous tracer. Despite its clinical potential, ASL remains challenging due to low signal-to-noise ratio and the presence of macrovascular signal contributions, which can bias perfusion estimates if not properly modeled. This work proposes a physics-informed neural network (PINN) framework for modeling the macrovascular component of time-encoded pseudo-continuous ASL (te-pCASL) signals. The approach combines a first network producing an implicit representation of the ASL signal with a second network that estimates voxel-wise haemodynamic parameters within a gamma-kernel-based macrovascular model. A physics-based constraint is introduced by enforcing consistency between the temporal derivative of the predicted signal and the analytical derivative of the macrovascular model, enabling direct parameter estimation while preserving physiological interpretability. The proposed framework is evaluated through a two-step process. First, a slice-wise training paradigm is employed to analyze model behavior and optimize key parameters and hyperparameters. Subsequently, the framework is extended to a whole-volume setting, incorporating three-dimensional spatial encoding and voxel-wise loss weighting to improve spatial coherence. The model is tested on in vivo time-encoded pCASL data acquired from healthy subjects. Results demonstrate that the proposed framework accurately reproduces the temporal dynamics of the macrovascular signal and benefits significantly from volumetric training, which improves the spatial consistency of estimated parameter maps, but does not completely address the identifiability limitations. However, challenges remain in the reliable estimation of physiological parameters, particularly bolus arrival time, due to parameter coupling and identifiability limitations. Overall, this study highlights the potential of physics-informed learning for integrating physiological models with data-driven approaches in medical imaging. The proposed framework represents a promising direction for improving macrovascular modeling in ASL, while also identifying key challenges that must be addressed to achieve robust and quantitatively reliable parameter estimation.
La Arterial Spin Labeling (ASL) è una tecnica di risonanza magnetica non invasiva che consente la valutazione quantitativa della perfusione cerebrale utilizzando l’acqua del sangue arterioso marcata magneticamente come tracciante endogeno. Nonostante il suo potenziale clinico, l’ASL presenta ancora diverse sfide, tra cui un basso rapporto segnale-rumore e la presenza di contributi di segnale macrovascolare, che possono introdurre bias nelle stime di perfusione se non opportunamente modellati. Questo lavoro propone un framework basato su Physics-Informed Neural Networks (PINN) per la modellazione della componente macrovascolare dei segnali di time-encoded pseudo-continuous ASL (te-pCASL). L’approccio combina una prima rete che produce una rappresentazione implicita del segnale ASL con una seconda rete che stima parametri emodinamici voxel-wise all’interno di un modello macrovascolare basato su kernel gamma. Viene introdotto un vincolo fisico imponendo la coerenza tra la derivata temporale del segnale predetto e la derivata analitica del modello macrovascolare, permettendo una stima diretta dei parametri mantenendo al contempo l’interpretabilità fisiologica. Il framework proposto viene valutato attraverso un processo in due fasi. In primo luogo, viene utilizzato un approccio di training slice-wise per analizzare il comportamento del modello e ottimizzare parametri e iperparametri chiave. Successivamente, il metodo viene esteso a un contesto volumetrico completo, incorporando una codifica spaziale tridimensionale e una pesatura voxel-wise della loss per migliorare la coerenza spaziale. Il modello viene testato su dati in vivo di te-pCASL acquisiti da soggetti sani. I risultati dimostrano che il framework proposto è in grado di riprodurre accuratamente la dinamica temporale del segnale macrovascolare e che beneficia significativamente dal training volumetrico, il quale migliora la coerenza spaziale delle mappe dei parametri stimati, pur non risolvendo completamente le limitazioni di identificabilità. Rimangono tuttavia difficoltà nella stima affidabile dei parametri fisiologici, in particolare del tempo di arrivo del bolo, a causa dell’accoppiamento tra parametri e dei limiti di identificabilità. Nel complesso, questo studio evidenzia il potenziale dell’apprendimento physics-informed per integrare modelli fisiologici con approcci data-driven nell’imaging medico. Il framework proposto rappresenta una direzione promettente per migliorare la modellazione macrovascolare nell’ASL, pur mettendo in luce le principali sfide che devono essere affrontate per ottenere stime dei parametri robuste e quantitativamente affidabili.
A Physics-Informed Neural Network Approach to Model Macrovascular Artifacts in Time-Encoded Arterial Spin Labeling
COCCA, NICOLA
2025/2026
Abstract
Arterial Spin Labeling (ASL) is a non-invasive magnetic resonance imaging technique that enables quantitative assessment of cerebral perfusion by using magnetically labeled arterial blood water as the endogenous tracer. Despite its clinical potential, ASL remains challenging due to low signal-to-noise ratio and the presence of macrovascular signal contributions, which can bias perfusion estimates if not properly modeled. This work proposes a physics-informed neural network (PINN) framework for modeling the macrovascular component of time-encoded pseudo-continuous ASL (te-pCASL) signals. The approach combines a first network producing an implicit representation of the ASL signal with a second network that estimates voxel-wise haemodynamic parameters within a gamma-kernel-based macrovascular model. A physics-based constraint is introduced by enforcing consistency between the temporal derivative of the predicted signal and the analytical derivative of the macrovascular model, enabling direct parameter estimation while preserving physiological interpretability. The proposed framework is evaluated through a two-step process. First, a slice-wise training paradigm is employed to analyze model behavior and optimize key parameters and hyperparameters. Subsequently, the framework is extended to a whole-volume setting, incorporating three-dimensional spatial encoding and voxel-wise loss weighting to improve spatial coherence. The model is tested on in vivo time-encoded pCASL data acquired from healthy subjects. Results demonstrate that the proposed framework accurately reproduces the temporal dynamics of the macrovascular signal and benefits significantly from volumetric training, which improves the spatial consistency of estimated parameter maps, but does not completely address the identifiability limitations. However, challenges remain in the reliable estimation of physiological parameters, particularly bolus arrival time, due to parameter coupling and identifiability limitations. Overall, this study highlights the potential of physics-informed learning for integrating physiological models with data-driven approaches in medical imaging. The proposed framework represents a promising direction for improving macrovascular modeling in ASL, while also identifying key challenges that must be addressed to achieve robust and quantitatively reliable parameter estimation.| File | Dimensione | Formato | |
|---|---|---|---|
|
Cocca_Nicola.pdf
embargo fino al 20/04/2027
Dimensione
9.2 MB
Formato
Adobe PDF
|
9.2 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/107592