A statistical approach for estimating brain tumor induced deformation

Abstract

A general statistical approach for predicting anatomical deformations is presented. Emphasis in this paper is on estimating deformations induced in the brain anatomy due to tumor growth. The presented approach utilizes the principal modes of co-variation between deformed (after tumor growth) and undeformed (before tumor growth) anatomy to estimate one given the other. In particular, with a statistical model constructed from a number of training samples, a patient's brain anatomy prior to tumor growth is estimated based on the patient's tumor-bearing images. This approach is suitable for use in registering a patient's tumor-bearing images to an anatomical atlas for purposes of surgical, or radio-surgical planning. The proposed approach is tested on a data set of 40 axial 2D brain images of normal human subjects. A biomechanical model was used to simulate tumor growth in each image of the data set. Pairs of deformed and undeformed anatomy were generated by tracking locations of 94 landmark points. The quality of the estimates of the undeformed anatomy are evaluated using the leave-one-out method. Results indicate good estimation accuracy considering the relatively small sample size.