A One-shot/Few-shot Interactive Segmentation Method for CT Image Segmentation.

Medical image ground truth annotation is important in clinical workflows for diagnosis, disease staging, prognostication of outcome, treatment planning, and treatment response assessment. However, annotating medical images requires significant human effort, especially for three-dimensional (3D) medical images. Deep learning techniques have significantly contributed to the methods of medical image annotation. Nonetheless, highly accurate automatic segmentation methods still require large-scale labeled data for training, which hinders the application in practice, especially for objects that are newly defined and/or are not well defined in the image and/or are spatially sparse with poor boundaries. Deep-learning(DL)-based interactive segmentation methods offer a trade-off between training cost and prediction performance. By leveraging extra regional prompts, these methods gain an unparalleled ability to locate the target object compared to automatic methods. This region recognition ability is essential for improving DL segmentation performance when training data are limited. In this paper, we present an interactive DL segmentation method for 3D computed tomography (CT) images, especially for the situation of one-shot/few-shot learning. Specifically, we propose a novel click prompt simulation method for training 3D interactive segmentation models inspired by the annotating procedures and focuses of human experts. The proposed prompt simulation consists of the object boundary prompt to indicate the global and local characteristics of the object and the mis-predicted region prompt to refine existing segmentations. We conduct experiments involving multiple anatomical structure segmentation tasks on our publicly available abdomen dataset under the condition of one-shot/few-shot learning. The experimental results demonstrate that the interactive segmentation method is more efficacious than its fully automated counterpart, which highlights the significant potential of applying interactive segmentation methods on the condition of lacking training data.