Segmentation-Guided Deep Learning Framework for Automated Abnormality Detection in Panoramic Dental Radiographs

Artificial intelligence (AI) has achieved significant advances in biomedical image analysis; however, many state-of-the-art deep learning models rely on large, balanced datasets and substantial computational resources, which are often unavailable in medical imaging. Panoramic dental radiographs present additional challenges due to non-uniform resolution, geometric distortion, overlapping anatomical structures, and the small spatial extent of diagnostically relevant abnormalities. In this study, we propose a segmentation-guided deep learning framework for abnormality detection in panoramic dental radiographs under limited and imbalanced data conditions. Instead of directly predicting image-level labels, segmentation is introduced as an intermediate supervisory signal to guide the model toward clinically relevant anatomical and pathological regions. Robust segmentation models are first trained to localize teeth and abnormalities, and the pretrained encoders are subsequently reused as frozen feature extractors for downstream abnormality classification. The proposed framework is evaluated on the Tufts Dental Database, which provides expert-annotated anatomical and abnormality segmentation masks but lacks explicit image-level labels. By deriving class labels from segmentation annotations, the proposed approach effectively bridges pixel-level supervision and image-level diagnosis. Experimental results demonstrate reliable abnormality detection performance at a reasonable computational cost, while systematic evaluation across multiple segmentation architectures shows that segmentation-guided representation learning improves robustness and interpretability in small and imbalanced dental imaging datasets.