Structure-preserving dental plaque segmentation via dynamically complementary information interaction

Jian Shi, Rui Xu, Baoli Sun, Tiantian Yan, Zhi-Hui Wang, Haojie Li

Abstract: In clinical practice, accurate dental plaque segmentation plays an important role in the diagnosis of oral problems such as dental caries and periodontitis. Existing methods consistently exhibit undesired structural distortions owing to the intensive variations in shape and the ambiguous boundaries of the plaques. In this paper, we introduce a Structure-Preserving Dynamic Complementary Interaction Network (SPDINet) that facilitates complementary information interaction between main plaque segmentation sub-network and auxiliary boundary sub-network to address this problem. Mutual-attention module (MA) and gradient guided refinement module (GGR) achieve this dynamic interaction at the feature level and result level, so as to preserve perceptual-pleasant details and further avoid structural distortion for plaque segmentation. In MA module, A bi-directional cross-task Mutual-attention mechanism reduces misleading attentions and distributes attention responses to emphasize parts that one task overlook while another task highlights. In GGR module, gradient boundaries derived from mask prediction category-wisely refine the boundary probability map to generate a more accurate boundary around the object, which in turns significantly boosts the performance of segmentation and implicitly implements the information interaction. Experiments on two recent dental plaque segmentation datasets, including SDPSeg-S and SDPSeg-C, show SPDINet establishes new state-of-the-art results.