Week 6 - Alexandra Roberts
Last week I shared the network performance on learning the mapping between 7T and downsampled 7T (3T analog) scans. This week I prepared a training dataset of real 3T and 7T scans acquired on a General Electric (GE) scanner. The pre-processing routine includes reconstructing the susceptibility weighted images (SWI) from gradient echo data, registering the 3T dataset to the 7T dataset, and normalizing the intensity of both datasets. In the fused image below, the green pixels indicate similar intensity, and the magenta pixels indicate differing intensity between the registered 3T slice and the 7T slice. Notice most of the difference is around the vessels and what appears to be phase artifact from the 7T.
After these measures, each of the 9 cases was split into 3D patches of size 32x32x32. The new residual is shown below. The residual is now a volume rather than a 2D slice.
A sample 3D patch is shown below.
Comments
Post a Comment