Rapid annotation of 3D medical imaging datasets using registration-based interpolation and adaptive slice selection

Accurate ground truth generation for 3D datasets is essential in many anatomy recognition and disease understanding applications. Interpolation techniques can substantially reduce the cost of manual segmentation for 3D images by only requiring a subset of 2D slices to be manually segmented, from which 3D segmentation is reconstructed through inter-slice label propagation. In this paper we present two enhancements, applying registration-based interpolation for 3D segmentation reconstruction and adaptive slice selection, to further speed up interpolation-based annotation. Comparing to shape-based interpolation methods, our technique can deliver accurate 3D anatomical segmentation with almost 50% reduction in the number of manually labeled slices.