Semi-Supervised Multi-Phase Image Segmentation and Application to Deep-Gray-Matter Segmentation in MRI Brain Images

Unsupervised image segmentations are usually implemented without human interactions, but the segmentation is sometime incorrect for complicated images, especially when the features of different classes are very close. On the other hand, supervised image segmentation, utilizing the features obtained by machine-learning and then applying some classification algorithms to the features, can usually get much more satisfying results. But supervised methods, are usually time-consuming, and only efficient for a specific type of data for each method. By a trade-off, semisupervised segmentation integrates the advantages of both supervised segmentation and unsupervised segmentation. In this paper, we proposed a semi-supervised multi-phase image segmentation framework which is motivated by image matting and central-gray-matter segmentation for magnetic resonance images (MRI). In our framework, an image is divided into two parts at the beginning, i.e., the known parts (labeled data) and the unknown parts (unlabeled data). The image segmentation is then to determine the unknown parts only. The class of a pixel in unknown part will be determined by not only its own features and the features of the known parts, but also its distance from the known parts. Experimental results demonstrate that our method outperforms unsupervised methods. Our method is also more efficient than supervised methods in the sense that there is no data required for training in order to obtain features for classification.