Efficient Soft-Constrained Clustering for Group-Based Labeling

Ryoma Bise, Kentaro Abe, Hideaki Hayashi, Kiyohito Tanaka, Seiichi Uchida

Abstract: We propose a soft-constrained clustering method for group-based labeling of medical images. Since the idea of group-based labeling is to attach the label to a group of samples at once, we need to have groups (i.e., clusters) with high purity. The proposed method is formulated to achieve high purity even for difficult clustering tasks such as medical image clustering, where image samples of the same class are often very distant in their feature space. In fact, those images degrade the performance of conventional constrained clustering methods. Experiments with an endoscopy image dataset demonstrated that our method outperformed various state-of-the-art methods.