DentSeg3: A Three-Branch Specialized Fusion Network for Dental Lesion Segmentation in Color Images

A three-branch fused semantic segmentation framework for color dental images is proposed for the joint segmentation of caries, cavities, cracks, tooth, and background. Accurate lesion segmentation is challenging due to subtle visual cues, inter-patient variability, and class imbalance. To address this, a neural network integrating three complementary branches is designed: (i) DeepLabv3 + with a ResNet-18 backbone to model large anatomical structures; (ii) a hierarchical Vision Transformer (ViT)-based branch with patch embedding only in the first block and self-attention over progressively downsampled feature maps, enabling contextual validation of spatially dispersed lesions; and (iii) a dual-attention branch combining CBAM and a Texture-Aware Attention (TAA) module to enhance fine-grained textures. These designs enable functional specialization and effective multi-scale fusion. Experimental results against classical clustering, encoder-decoder, and recent attention- and transformer-based methods show that the fused model achieves the best overall performance among the evaluated methods, achieving 96.2% accuracy, 93.6% Dice, 92.9% IoU, 93.8% precision, and 93.7% recall with data augmentation. Ablation studies confirm that gains mainly arise from branch fusion. Grad-CAM analysis reveals complementary activation patterns, improved boundary delineation, and robustness to background variability. Overall, the proposed framework provides a robust semantic segmentation solution with consistent cross-device generalization across different imaging conditions.