ROI-based binary caries segmentation in intraoral periapical radiographs using a ResNet-conditioned U-Net with FiLM-modulated decoder

Accurate segmentation of dental caries in intraoral periapical radiographs remains challenging because lesions are often small, low-contrast, and visually confounded by surrounding anatomical structures. To address this problem, an ROI-based binary caries segmentation framework is proposed using a ResNet conditioned U-Net with a FiLM-modulated decoder. The method combines a standard 4-level U-Net segmentation stream with a lightweight TinyResNetContext branch that extracts a global representation from the same ROI input. This context vector is used to generate feature-wise modulation parameters that adapt decoder feature refinement at multiple stages, enabling case-specific lesion delineation. The study was conducted on a dataset of 1,887 intraoral periapical radiographs collected from SIBAR Institute of Dental Sciences, for which binary lesion masks were generated through CVAT-based annotation. Following preprocessing, dataset partitioning, and training-time augmentation, the proposed model was evaluated against U-Net, U-Net++, and ResNet34 baselines under both ROI-based and non-ROI settings using an identical training and evaluation protocol. Experimental results showed that ROI-based training consistently outperformed full-image segmentation across all baseline families, confirming the importance of lesion-focused inputs. Among all evaluated models, the proposed method achieved the strongest overall test-set performance, with Precision = 0.8302, Recall = 0.9049, F1-score = 0.8659, Dice = 0.8659, IoU = 0.7635, and AP50 = 0.9792. Checkpoint-level statistical analysis further demonstrated significant improvements over U-Net (ROI) and all non-ROI baselines, while also indicating favorable generalization relative to competing ROI-based models. In addition, a multi-seed robustness analysis on ROI-based models confirmed that the observed performance gains were reproducible across different random initializations. These findings suggest that integrating global ROI context into decoder refinement provides a competitive and computationally lightweight strategy for binary caries lesion segmentation under predefined ROI-based radiographic analysis settings. The present framework should be interpreted as an ROI-based segmentation approach rather than a fully end-to-end automated clinical caries detection and segmentation workflow operating directly on uncropped radiographs.