Characterizing morphology differences from image data using a modified fisher criterion

Image-based morphometry of cells, tissues, and organs is an important topic in biomedical image analysis. We propose a novel method to characterize the morphological information that discriminates between two populations of morphological exemplars (cells, organs). We first demonstrate that the application of standard techniques such as Fisher linear discriminant analysis (FDA) can lead to undesirable errors in characterizing such information. We then describe an adaptation of the FDA technique that utilizes a least squares projection error to regularize the final solution. We show results comparing the FDA, modified FDA, and principal component analysis (PCA) techniques utilizing a contour-based characterization of both simulated and real images of cell nuclei.