Combining “Vertical” and “Horizontal” Features from Medical Images

Abstract

Landmark points, when they can be located reliably over a sample of medical images, serve both to generate a feature space of biometric “shape variables” and also to specify an unwarping by which variation of pixel values associated with displacements of the landmarks can be corrected with respect to all landmarks simultaneously. There thus arise two separate spaces of descriptors, the “horizontal” and the “vertical.” Combinations of features from this pair of spaces for prediction or discrimination can be visualized in one or the other space alone by injecting precisely calibrated crosstalk between these channels. This essay analyzes a set of brain images in this way to highlight the effect of schizophrenia upon anatomy in the midsagittal plane.