Pattern Analysis and Applications Computer Vision and Pattern Recognition
RoDNet: spatial-frequency collaborative representation learning with bounded activation calibration for dental anomaly detection
Abstract
Automated anomaly detection in dental panoramic radiographs is clinically important, yet substantial lesion-scale variation and interference from high-intensity restorative or implant-related structures continue to hinder reliable feature representation. To address these challenges, we propose RoDNet, a dental anomaly detection framework that combines a Spatial-Frequency Collaborative Pyramid Network (SF-CPN) with a Bounded Activation Calibration Module (BACM). Within SF-CPN, a Wavelet-based Receptive Expansion Module (WREM) and a Frequency-Gated Aggregation Module (FGAM) serve complementary roles in enhancing structural context and lesion-sensitive details. WREM expands the effective receptive field while preserving local spatial cues, which supports lesion representation across different scales. FGAM applies frequency-aware feature modulation within the attention pathway to improve the representation of weak boundaries and subtle textures. Complementing these representation modules, BACM places statistics-free, learnable bounded mappings before the attention and feed-forward transformations to calibrate high-magnitude feature responses associated with artifact-source structures. Across publicly available dental panoramic radiograph datasets and additional benchmarks, RoDNet consistently improves multi-scale detection performance. Compared with YOLOv12n, it achieves absolute AP gains of 1.2, 0.8, and 1.5 percentage points on ODet3, DXR-Pano, and PDR-10, respectively. Code and dataset acquisition instructions will be available at https://github.com/JiachunWu/RoD.