A Maximum Likelihood Classification method for image segmentation considering subject variability

In this paper, we present a new statistical model for Maximum Likelihood Classification (MLC) algorithm to improve the image segmentation/classification performance. MLC has been widely used in many classification applications. For supervised MLC, the parameters of the statistical model are obtained from the training dataset at the learning step. However, in the previous studies, the feature values of different classes are assumed to have similar distributions for different subjects. This is not true in many real world situations. The considerable differences across subjects have not obtained much attention before. To conquer this difficulty, we model the mean of feature values of each subject and the feature values as two groups of dependent random variables. This is made possible by using a bivariate Gaussian mixture model to fit the image data of different subjects. In this way, class membership depends on both the feature values and another random variable that captures subject-specific information. We apply our method to simulated image data and our experimental results show that the proposed model could improve the classical supervised MLC segmentation results when there are considerable differences across subjects.