Integrating Imperfect Cross-Domain Segmentation Maps into Dental Radiograph Classification: Robustness of Soft Probability Weighting versus Hard Masking

Authors

• Raed Bulbul, Department of Computer Engineering, Islamic University of Gaza, Gaza, PalestineFollow
• Mohammed Alhanjouri, Department of Computer Engineering, Islamic University of Gaza, Gaza, PalestineFollow

Abstract

Tooth segmentation maps transferred across dental radiograph datasets are imperfect, and the way they are integrated into a diagnostic classifier determines whether their errors reach the dental diagnosis. This study quantifies such errors on external panoramic radiographs and compares hard masking with soft probability weighting for tooth-level classification of dental pathology. A U-Net and a conditional diffusion model were trained on STS-2D-Tooth and frozen; the diffusion model refined the U-Net probability map in six image-conditioned denoising steps. On two independent external datasets from Bangladesh (319 radiographs) and Tunisia (59 radiographs), U-Net maps reached Dice 0.787 and 0.871, and 73% and 87% of missed tooth pixels lay within 4 pixels of the tooth boundary. Diffusion refinement improved Dice by +0.0100 (95% confidence interval (CI) +0.0085 to +0.0116) and +0.0089 (+0.0048 to +0.0139), mainly by recovering missed tooth pixels. On DENTEX 2023, raw, hard-masked and soft-weighted tooth crops with identical geometry were classified as caries versus other pathology by five-seed ResNet-50 ensembles and evaluated once on 504 test annotations after a validation-only pre-test lock. Under five controlled map perturbations, hard masking lost up to 0.022 area under the receiver operating characteristic curve (AUC) per severity level, whereas soft probability weighting remained within 0.003, with 95% CIs of the slope difference excluding zero in 6 of 10 comparisons and consistently under erosion and boundary deletion, which mimic the under-segmentation at tooth boundaries observed externally. With unperturbed maps, all conditions reached AUC 0.961–0.974; soft-weighted U-Net and diffusion-refined inputs achieved AUCs of 0.9739 and 0.9729, respectively (ΔAUC -0.0010; 95% CI -0.0070 to +0.0042). Soft probability weighting provided greater robustness to the tested map errors, with consistent benefits across both segmentation models under erosion and boundary deletion.

Recommended Citation

Bulbul, Raed and Alhanjouri, Mohammed (2026) "Integrating Imperfect Cross-Domain Segmentation Maps into Dental Radiograph Classification: Robustness of Soft Probability Weighting versus Hard Masking," Mesopotamian Journal of Artificial Intelligence in Healthcare: Vol. 4: Iss. 1, Article 8.