SliceProp: A Slice-Wise Bidirectional Propagation Model for Interactive 3D Medical Image Segmentation

Interactive medical image segmentation methods have become increasingly popular in recent years. These methods combine manual labeling and automatic segmentation, reducing the workload of annotation while maintaining high accuracy. However, most current interactive segmentation frameworks are limited to 2D image data, and are not suitable for 3D image data due to the large size and high complexity of 3D data, as well as the challenges posed by information asymmetry and sparse annotation. In this paper, we propose SliceProp, an interactive segmentation framework that implements slice-wise Label Bidirectional Propagation (LBP) for 3D medical image segmentation. SliceProp extends the interactive 2D image segmentation algorithm to 3D image segmentation, and can handle 3D data with large size and high complexity. Moreover, equipped with a Backtracking Feedback Check (BFC) module, SliceProp effectively addresses the issues of information asymmetry and spatial sparse annotation in 3D medical image segmentation. Additionally, we adopt an uncertainty-based criterion to pri-oritize the slices to be refined interactively, which enhances the efficiency of the interaction process by enabling the model to focus on the regions with the most unreliable predictions. SliceProp is evaluated on two datasets and achieves promising results compared to state-of-the-art methods.