Figure 9
The automatic segmentation pipeline. (a) Normalized T1 and T2 maps and Z-spectrum images acquired with various saturation B1 amplitudes and at various frequency offsets Δω (3 ppm shown) for an illustrative DU145 mouse. (b) Non-background voxels were concatenated into an observation matrix and transformed by a trained independent component analysis transform set to generate three independent component (ICs). As part of training, the ICs were sorted in order of increasing mutual information with respect to the input and each IC was weighted. A different set of IC weights was used for each of the two tumour types. (c) The ICs were then input to the Gaussian mixture model set to five clusters.
