Validation of appearance-model based segmentation with patch-based refinement on medial temporal lobe structures

Abstract

This paper presents a new automatic segmentation scheme that combines active appearance (AAM) modeling and patch-based label fusion into a segmentation framework. AAM, which uses eigen-decomposition to analyze the statistical variation of image intensity and shape infor- mation over the population, is used to capture the global shape char- acteristics of the structure of interest with a generative model, while patch-based label fusion, which uses a non-local means method to com- pare the image local intensity properties, is applied to locally refine the segmentation results along the structure boundary area to improve the segmentation accuracy. The proposed segmentation scheme is used to segment human medial temporal lobe structures, which have low inten- sity contrast in MRI and complexity in shape. The experiments demon- strate that this new segmentation scheme is computationally efficient and robust. In a leave-one out validation with fifty-four normal young subjects, the method yields a mean Dice κ of 0.87 for hippocampus, 0.81 for amygdala, 0.73 for para-hippocampal complex cortex, and 0.73 for perirhinal cortex between manual and automatic labels.