Bilateral Edge‐Augmented Recurrent Horizontal Feature‐Shift Aggregation Network for Tooth Instance Segmentation With FDI ‐Based Numbering

Tooth instance segmentation in panoramic radiographs is crucial for computer‐aided dental diagnosis, yet remains challenging due to low contrast, structural similarity and the coexistence of permanent and deciduous teeth. In this paper, we propose BEHA‐Net, a Bilateral Edge‐Augmented Recurrent Horizontal Feature‐Shift Aggregation Network for joint tooth instance segmentation and Fédération Dentaire Internationale (FDI)‐compliant tooth numbering. BEHA‐Net integrates two key components: (1) a recurrent horizontal feature‐shift aggregation (HFA) module that captures cross‐tooth contextual correlations via leveraging the horizontal arrangement topology of dental arches and aggregating multi‐scale dependencies across adjacent and non‐adjacent teeth; and (2) a bilateral edge‐augmented (BEA) module that refines boundary representations by edge detail enhancement and noise suppression. Extensive experiments on the Tufts and O 2 PR benchmarks demonstrate the superiority of our method over existing approaches.