Boundary-Aware ResUNet for Small and Ambiguous Lesion Segmentation in Dental Panoramic Radiographs under Class Imbalance

This study investigates a residual U-Net (ResUNet)based boundary-aware training strategy for the segmentation of small and low-contrast pathological classes in dental panoramic radiographs, where severe class imbalance makes findings such as caries and periapical lesions particularly hard to delineate. On a seven-class panoramic radiograph dataset we benchmark several encoder–decoder backbones and loss functions and analyze the results per class and with boundary-based metrics. The experiments show that the proposed boundary-aware ResUNet yields a limited but consistent improvement in the overall Dice score (from 0.596 to 0.607) and, more importantly, produces more balanced results in terms of boundary metrics and small-class performance, improving the per-image Dice of every foreground class over a plain region loss. Combining the boundary-trained and a complementary Tversky-trained ResUNet with test-time augmentation, the proposed system reaches the best test Dice (0.6168), IoU (0.446) and macro precision (0.667) together with the lowest boundary errors (HD95 280.9, ASSD 266.7), and improves the two hardest classes, caries and periapical lesions, by about 21 and 35 percent relative to a single boundaryaware model, all without any additional training. Beyond the empirical study we derive an exact decomposition of the pooled Dice score into a class-mass weighted average of per-class Dice scores, together with a sensitivity bound that quantifies how much aggregate improvement a given class can produce at all. Instantiated on our test set the bound shows that caries and periapical lesion jointly cap the attainable aggregate gain at about 0.045 Dice, so a modest aggregate margin is a structural property of the metric under extreme imbalance rather than evidence of a small clinical effect. This motivates per-class and boundary-aware evaluation on this task.