Physics-Informed Neural Networks for Quantitative MRI: Bloch-Torrey-Inspired Parameter Estimation

Sammanfattning

Magnetic resonance imaging (MRI) is widely used in both clinical and research settings. However, most protocols are qualitative, relying on contrast mechanisms that depend on scanner settings and acquisition parameters. This provides relative measurements of tissue properties, making them subject to variability and limiting the comparability. Quantitative MRI (qMRI) tries to address these limitations by instead estimating tissue-specific parameters such as the longitudinal (T1) and transverse (T2) relaxation times and the equilibrium magnetisation (M0), providing reproducible and quantitative measurements. Common qMRI methods, such as least-squares (LSQ) fitting, estimate these parameters by fitting the acquisitions to signal models. Although effective, it requires multiple acquisitions and are sensitive to noise, limiting its clinical practicality. This project investigates a Bloch–Torrey–inspired physics-informed neural network (PINN) that integrates a diffusion-like physical constraint while learning directly from the MRI signal evolution. The model was evaluated on simulated brain data generated from BrainWeb anatomical models with varying sampling densities and noise levels, and further assessed using in vivo mouse brain MRI data. The performance of the proposed PINN was compared with LSQ fitting. On the simulated data, the PINN achieved comparable or improved performance relative to the LSQ fitting under reduced sampling and high noise conditions, demonstrating the potential advantages of the method. However, on the in vivo data, the LSQ estimations remained more accurate overall, while the PINN exhibited slow convergence and limited improvement in parameter estimations. The results indicate that a PINN can be a promising method for qMRI but it requires further optimisation to reach its full potential. Overall, the study demonstrates that physics-informed neural networks can offer a potentially more data-efficient alternative to traditional fitting methods for qMRI, with potential to improve the robustness in parameter estimation.