Abstract
Conventional machine learning-based segmentation (e.g., ensemble learning) suffers from the problem of unnatural shapes of the extracted figures because decision-making by the constructed classifier is carried out voxel by voxel or local region by local region independently. In this paper, we propose an ensemble learning algorithm that constructs a segmentation process based on the statistical shape feature of an organ. We describe a novel loss function for evaluating the shape of an extracted figure using a statistical shape model of the organ and an algorithm to minimize the loss function which combines conventional error loss with proposed loss. The results of experiments using an artificial image are presented to confirm the basic performance of the algorithm. In addition, the results of experiments involving spleen segmentation using 80 clinical CT volumes are presented to validate the clinical usefulness of the algorithm. Based on these results, it is concluded that the proposed algorithm reduces unnatural shapes of the extracted organs and provides significantly superior segmentation performance as compared to conventional ensemble learning-based segmentation.
