RFFA-Net: Recursive Filtering and Feature Aggregation Driven Network for Tooth Segmentation in CBCT image

Automatic tooth segmentation has significantly improved the accuracy and efficiency of clinical diagnosis, while also providing a technical foundation for the development of patient-specific treatment strategies. Although significant progress has been made in automatic tooth segmentation research, existing methods still struggle to accurately segment tooth structures, when confronted with the presence of metal artefacts and blurred boundaries in dental CBCT images. To address these challenges, this paper proposes an encoder-decoder network (RFFA-Net) driven by recursive filtering and feature aggregation strategies for three-dimensional segmentation of tooth structures. First, a Dense Mamba (DM) module is introduced into the encoder of the proposed network, which combines recursive filtering and dense connections to enhance feature extraction capabilities and reduce the interference caused by metal artifacts in CBCT images. Second, a dynamic multi-scale feature fusion (DMFF) module is designed at the network bottleneck. This module adaptively aggregates multi-scale features extracted from the encoder to better capture the irregular shapes of teeth. Finally, to address the issue of blurred tooth boundaries in CBCT images, an Inter-Channel-Spatial Attention (ICSA) module is introduced. This module employs a gating mechanism to dynamically adjust the weights of both channel and spatial attention, thereby effectively fusing global semantic information and local details. We conducted experiments on two commonly used dental CBCT datasets. The results indicate that our method outperforms existing state-of-the-art tooth segmentation methods, demonstrating its effectiveness, robustness, and superiority. In particular, in the presence of metal artifacts and blurred boundaries commonly found in CBCT images, the proposed method still maintains stable and accurate segmentation performance.