Figure 4.

Neural network architectures for parameter inference. The neural networks receive the complex signal in SVD subspace and infer the underlying tissue parameter vector θ based on different architectures. (a) The multi-pathway implementation (NN multipath) is designed such that each path specializes on the estimation of one parameter, parameter ${\mathrm{\theta }}_{\mathrm{i}}$, i.e. T1, T2 or a PD related scaling factor. (b) The single pathway (NN fwd) architecture has the total number of nodes as the NN multipath but directly infers the parameter vector from the same hidden layers. (c) Parameters can also be derived from an autoencoder network (NN fwd-bck), where T1 and T2 are directly inferred from the latent space and relative PD is computed from the decoded output signal.