Domain Transformation for PMMA-Inserted Vertebral Body Segmentation

Abstract

Learning-based medical image segmentation has been advanced with the collection of datasets and various training methodologies. In this work, we target bone cement (polymethylmethacrylate [PMMA]) inserted vertebral body segmentation, where the target dataset was relatively scarce, compared to a large-scale dataset for the regular vertebra segmentation task. We presented a novel domain transformation framework, where a large-scale training set for our target task was generated from the existing dataset of a different domain. We proposed two main components: label translation and image translation. Label translation was proposed to filter out unnecessary regions in a segmentation map for our target task. In addition to label translation, image translation was proposed to virtually generate PMMA-inserted images from the original data. The synthesized training set by our method successfully simulated the data distribution of the target task; therefore a clear performance improvement was observed by the dataset. By extensive experiments, we showed that our method outperformed baseline methods in terms of segmentation performance. In addition, a more accurate shape and volume of a bone were measured by our method, which satisfied the medical purpose of segmentation.