Gradient-oriented profiles for unsupervised boundary classification

We present a method for unsupervised boundary classification by producing and analyzing intensity profiles. Each profile is created by sampling an ellipsoidal neighborhood of voxels oriented along the image gradient. The profile is analyzed via nonlinear optimization to find the best fitting cumulative Gaussian. The parameters of the cumulative Gaussian parameterize the boundary directly yielding: (1) extrapolated intensity values for voxels located far inside and outside of the boundary; and (2) estimates the boundary location and boundary width. For these parameters, intrinsic measures of confidence are established to eliminate low-confidence parameter estimates. Neighborhoods overlap considerably, yielding sufficient high-confidence estimates for a thorough survey of the boundary. Gradient oriented profiles are demonstrated on artificially generated 3D test data and proved to accurately parameterize and classify the boundary