Optimal MAP Parameters Estimation in STAPLE - Learning from Performance Parameters versus Image Similarity Information

Abstract

In many medical imaging applications, merging segmentations obtained from multiple reference images (i.e., templates) has become a standard practice for improving the accuracy as well as reliability. Simultaneous Truth And Performance Level Estimation (STAPLE) is a widely used fusion algorithm that simultaneously estimates both performance parameters for each template, and the output segmentation; a more accurate estimation of performance parameters consequently results in more accurate output segmentations. In this paper, we propose a new approach for learning prior knowledge about the performance parameters of each template, and for incorporating it into the Maximum-a-Posteriori (MAP) formulation of the STAPLE, so that more accurate output segmentations can be obtained. More specifically, we propose a new approach to learn, for each structure to be segmented, the relationships between the performance parameters (viz. sensitivity and specificity) and the intensity similarities; we also propose a methodology for transferring this prior knowledge about the performance parameters into the STAPLE algorithm through optimal setting of the MAP parameters. The proposed approach is evaluated for the segmentation of structures in the brain MR images. These experiments have clearly demonstrated the advantages of incorporating such prior knowledge.