Quantitative MRI parameter mapping with extended phase graphs and recurrent inference machines

11/30/2022

Quantitative Magnetic Resonance Imaging is an imaging technique that allows the quantitative assessment of inherent tissue properties which often suffers from long scan times, preventing its use in routine clinical evaluation. A new class of deep learning frameworks, called model-based deep learning nets, incorporate the magnetic resonance signal model into the learning process to estimate parametric maps of the tissues, alleviating the need for a large number of training datasets and further pushing acceleration rates. However, most of the nets previously used in this context have employed a pure exponential curve to model the magnetic resonance signal, which does not account for stimulated or indirect echoes, or inhomogeneity of the B1 field. In this work, one of such models, the Recurrent Inference Machine (RIM), is adapted to perform T2 mapping with a more accurate signal model based on the Extended Phase Graphs (EPG) concept, while also considering the influence of the effective B1 field. Additionally, Recurrent Inference Machines with forward model and gradient dictionaries (RIMFoGraD) are proposed, which preserve the configuration of the original RIM, but speed up the inference process through dictionaries of pre-calculated echo-modulation curves and their gradients for a large range of parameters. RIMFoGraD was able to estimate T2 320x320 pixel maps 380 times faster than the RIM implemented with the EPG model with no loss in accuracy, and was 80% faster than a pattern-based recognition approach with 1 ms T2 precision, with a median T2 difference of 2.83 ms from this method on the brain parenchyma.

Authors in the community:

• 

  C

  Catarina Neves de Carvalho

  ist189791

Supervisors of this institution:

• 

  R

  Rita Homem de Gouveia Costanzo Nunes

  ist31063

industrial-biotechnology - Industrial Biotechnology

eng - English

Instituto Superior Técnico