Table 2 Number of trainable weights in different topologies.

From: Minimal Linear Networks for Magnetic Resonance Image Reconstruction

 

#Tunable weights N = 128 matrix size, Nch = 13 channels, NTS = 7 Time segments, NN = 12 Neighbors, Acc = 4

AUTOMAP4

1,409,286,144

Nch⋅N4/Acc + 2⋅N4 + C

Full general-purpose linear transform

1,744,830,464

2Nch⋅N4/Acc

k + I MLN, TS independent

36,012,032

2Nch⋅NSegments⋅NN⋅N2 + 2N2⋅NTS

MLN, TS shared

5,398,528

2⋅N2⋅Nch/Acc + 2N2⋅NSegments⋅NN + 2N2⋅NTS

MLN, no B0 correction

5,111,808

2⋅N2⋅Nch⋅NN

  1. The full general-purpose linear transform transforms is similar to the main layer of AUTOMAP - transforming directly from signal data to reconstructed image.
Back to article page