Slice Segmentation Propagator: Propagating the single slice annotation to 3D volume

Abstract

In this paper, we consider the problem of semi-automatic medical image segmentation, with the goal of segmenting the target structure in a whole 3-D volume image with only a single slice annotation to relieve the user’s annotation burden. Under such a paradigm, the segmentation of the volume is achieved by establishing the correspondence between slices and propagating the reference segmentation. We propose a more medical-suited framework denoted Slice Segmentation Propagator (SSP) that can establish reliable correspondences between slices with local attention, and maintain a running memory bank that effectively mitigates the problem of error accumulation during mask propagation. Additionally, we propose two test-time training strategies to further improve the propagation performance and generalization ability of the framework, namely, a cycle consistency mechanism to suppress error propagation, and an online adaption procedure via artificial augmentation, assisting the model to better generalize towards new structures at test time. We have conducted thorough experiments on three datasets on four anatomy structures, demonstrating promising results on both in-structure and cross-structure (test on different structures from trainset) scenarios.