Principal Component Analysis (PCA)

This module yields so-called eigenimages from a stack of 2D images or 3D structures, thereby facilitating subsequent image classification to resolve heterogeneity.

The PCA1) is a prerequisite for data complexity reduction in order to perform a subsequent Classification; alternatively, it can also be directly performed within the classification logic. The PCA yields eigenimages (volumes in the 3D case) which represent the heterogeneity or covariance within the data. Pre-computing the PCA before classification allows to inspect the resulting eigenimages and to decide how many eigenimages should be used for image classification. In general, the first eigenimages will correspond to actual variations in structure and/or angular orientation, whereas later eigenimages will rather represent variations in the noise. Using an external mask (e.g. a circle) as additional input allows to suppress background noise. Using external eigenimages from a previous PCA as input allows to combine eigenimages from different groups of data. The latter may be useful when working on data with orientation bias.

Parameters Description
Dimension Number of calculated dimensions (Eigen images)
Eigenimages Use external eigen images
Using mask Use an external mask
Input Description
Input Stack of input images
Eigen Images External Eigen images
Mask External mask
Output Description
Eigenimages Stack of generated Eigen images
New/Changed Header Values Description
eigenValue Eigenvalue of the corresponding eigenimage (or volume in 3D case)

Hall, P. et al. (2000). Merging and splitting eigenspace models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(9), 1042-49.