Topographic classification of nuclear medicine images for tumor detection

Abstract

Topographic classification is a nonlinear technique used to enhance nuclear medicine images for tumor detection. Second-order directional derivatives are computed at each pixel location after performing a least-squares fit of the underlying surface using a bivariate cubic polynomial. The eigenvalues and their corresponding eigenvectors computed from the Hessian matrix determine which topographic feature is assigned to the image pixel. Parameter selection for the mask size, curvature threshold, and angle thresholds are chosen to yield the "best" classified image. The classifier is applied to clinical images of cancer patients provided by the Department of Nuclear Medicine at the University of Arizona. Background noise associated with the photon-starved data is suppressed using a Difference-of-Gaussians (DOG) filter prior to pixel classification. Results indicate the feasibility of using this technique to isolate possible tumor sites which will assist the clinician during patient examination.