Semi-automatic segmentation of knee cartilage in longitudinal MR images by seed transfer

In this paper, we propose an efficient and effective semi-automatic segmentation framework for knee cartilage in longitudinal MR images. The proposed method provides a seamless and efficient combination of initial automatic segmentation and interactive segmentation. The main idea is to generate and transfer the seeds, rather than the labels, of the manual segmentation results of the initial image, to segment sequential images. Here, segmentation seeds are defined as sample voxels that are highly likely to be either the organ of interest or background. Similar to user annotations by scribbles, seeds can be used to determine labels for remaining voxels. Compared to methods that transfer segmentation labels, the proposed method is able to reduce the amount of required user annotations, since it does not require the user to identify and correct erroneous regions in the transferred segmentation labels, while still providing a robust initial segmentation based on transferred initial seeds. We evaluate the effectiveness of the proposed framework by measuring the interaction time of the user to segment the cartilage within the images. We compare the results with previous methods and show that the proposed method reduces the required interaction time.