Applying training hidden features to joint curve evolution for brain MRI segmentation

According to the level of information provided in images, segmentation techniques can be categorized into two groups. One is region-labeling, which obeys the intensity-based classification methods. Although modeling the tissue intensity is straightforward by applying local statistical methods and spatial dependencies, the results might suffer from noise and incomplete data. The second group of techniques applies active contour models, in which the objective is to find the optimal partition of the image domain using a closed or open curve by using prior constraints on the shape variation. However, estimating optimal curve is intractable due to the incomplete observation data. This paper extends a previously reported joint active contour model for medical image segmentation in a new Expectation-Maximization (EM) framework, wherein the evolution curve is constrained not only by a shape-based statistical model but also by applying a hidden variable model from the image observation. In this approach, the hidden variable model is defined by the local voxel labeling computed from its likelihood function, depended on the image functions and the prior anatomical knowledge. Comparative results on segmenting putamen and caudate shapes in MR brain images confirmed both robustness and accuracy of the proposed curve evolution algorithm.