Effect of Number of Coupled Structures on the Segmentation of Brain Structures

This paper reports the effect of the coupling information on the performance of model-based segmentation of the brain structures from magnetic resonance images (MRI). We have developed a three-dimensional, nonparametric, entropy-based, and multi-shape method that benefits from coupling of the shapes. The proposed method uses principal component analysis (PCA) to develop shape models that capture structural variability and integrates geometrical relationship among different structures into the algorithm by coupling them (limiting their independent deformations). At the same time, to allow variations of the coupled structures, it registers each structure individually when building the shape models. It defines an entropy-based energy function which is minimized using quasi-Newton algorithm. Probability density functions (pdf) are estimated iteratively using nonparametric Parzen window method. In the optimization algorithm, analytical derivatives are used for maximum speed and accuracy. Sample results are given for the segmentation of caudate, thalamus, putamen, pallidum, hippocampus, and amygdala illustrating superior performance of the proposed method compared to the most similar method in the literature. The similarity of the results obtained by the proposed method with the expert segmentation is 4% to 12% higher than that of the most similar method. Experimental studies show that the proposed coupling method, which regulates shape variability during segmentation, improves accuracy of the results of the proposed method by 6% and those of the other method by 1%. In addition, the more the structures are used in the coupling process, the more accurate the results are obtained.