Simulation-based amortised inference for analysis of multi-dimensional spectra in combined diffusion-relaxation MRI

Motivation: Multi-dimensional diffusion-relaxometry spectra may provide detailed insights into different microenvironments within tissue. However, traditional methods for producing multi-dimensional spectra are unstable and computationally demanding. Goal(s): To demonstrate the efficacy of simulation-based inference (SBI) to efficiently estimate multi-dimensional spectra in combined diffusion-relaxation MRI. Approach: Using SBI, we combine a generative signal model with a neural posterior estimator to approximate D-T2 distributions. We test our method on simulated data, including a brain phantom derived from microscopy. Results: SBI was able to identify multiple, distinct T2-D compartments effectively, even when peaks had small separation. SBI enables parameter inference in seconds, far faster than classical methods. Impact: This study applied simulation-based inference to multi-dimensional MRI, enabling rapid, precise and robust microstructural tissue characterisation. SBI's computational speed will particularly benefit high-dimensional data, as the number of parameters grows exponentially with dimensions in classical inference, making conventional methods slow.