Context-Aware Caries Segmentation in Periapical Radiographs Using a Hybrid Multi-Task Learning Framework with Partial Annotations

Dental caries remains a prevalent diagnostic challenge, particularly in periapical radiographs, where geometric distortions and anatomical overlap complicate interpretation. Although deep learning has advanced dental image analysis, most segmentation models depend on fully annotated datasets and rarely exploit anatomical context. This study proposes a hybrid Multi-Task Learning (MTL) framework that jointly performs anatomical segmentation and caries detection in scenarios with partial and asymmetric annotations. The method integrates a U-Net++ dual-head architecture with a shared EfficientNet-B4 encoder, supplemented by pseudo-label generation and selective loss masking to handle incomplete ground truth. We hypothesize that learning healthy dental structures provides a contextual scaffold that enhances the identification of pathology. Experiments on an independent test set validate this hypothesis: the proposed MTL model achieved an DSC of 0.6706 and an IoU of 0.5044, outperforming a specialized single-task baseline. Most notably, sensitivity improved by 7.47%, reducing false-negative pixels by 19.9%. These findings demonstrate that context-aware supervision substantially improves detection robustness on complex periapical radiographs, even when full-pixel-level annotations are unavailable. The proposed framework offers a scalable pathway for developing clinically oriented diagnostic tools in real-world settings where annotation completeness is limited.