Abstract
<title>Abstract</title> <p>Peri-implant bone loss is a primary radiographic indicator of implant compli- cation, yet its assessment from panoramic radiographs remains manual and subject to inter-observer variability. This study proposes BoneNet, a single- stage detection architecture built on a YOLOv8m backbone in which standard C2f blocks are replaced by Dilated Residual Attention Blocks (DRAB) – dual-rate dilated convolutions fused through a Convolutional Block Attention Module (CBAM) – trained with a Clinically-Graded Bone (CGB) loss assigning severity-proportional penalty weights across five bone-loss grades. BoneNet was selected after benchmarking nine candidate architectures on a balanced radio- graph corpus, achieving mAP@50–95 of 0.9660, a 3.83-point improvement over the strongest baseline (YOLOv8m, 0.9277), with fewer parameters (23.07M vs. 25.8M) and lower cost (70.4 vs. 78.7 GFLOPs). A cross-attention fusion network combining 512-dimensional image embeddings with a 27-feature tabular branch raised severity-grade classification accuracy from 50.79% (image-only) to 85.96%. Bone-loss depth was estimated using a gradient-boosted regressor (XGBoost), achieving a mean absolute error of 0.638 mm, within the clinically cited ±1 mm tolerance. Stratifying depth error by grade revealed a statistically significant, monotonically increasing bias (Kruskal-Wallis H = 36.76, p < 0.0001), with only 42.9% of Grade 4 predictions within tolerance versus 100% for No-Bone-Loss cases. Dual-modality explainability used EigenCAM for the image branch and SHAP for the tabular branch, both converging on clinically plausible attributions. This work contributes a validated architectural refinement over a benchmarked baseline, evidence that tabular features complement image-only deep learning at limited dataset scale, and a transparent account of where the system’s predictions remain unreliable.</p>