Efficient Thin Structure Segmentation via Edge, Directional, and Topology-Preserving Priors

Thin, elongated structures---pavement and structural cracks, pulmonary fissures, and retinal vessels---share segmentation challenges: extreme aspect ratios, irregular trajectories, low contrast, and sensitivity to fragmentation. Accurate pixel-level segmentation with structural continuity remains difficult, particularly for real-world deployment. This dissertation tests the hypothesis that shared morphological priors can be exploited through principled architectural design.

Three architectures address these challenges. dCrack++ introduces Edge-Guided Attention (EGA) for thin irregular structures, alongside the dCrack61k dataset (61,943 images, 36 sources). It achieves a Dice score of 0.65 and an mIoU of 0.780 on dCrack61k, best F1 and IoU on CHASE_DB1 retinal vessels, and a Dice score of 0.882 on pulmonary fissures, outperforming UNet++, DeepLabV3, and nnU-Net. Its 186M parameters and isotropic 3x3 convolutions limit deployment and directional modeling.

STRIPE addresses this limitation with the Anisotropic Feature Module (AFM), replacing isotropic 3x3 convolutions with asymmetric 1x7 and 7x1 convolutions for encoder-level directional encoding. STRIPE reduces parameters 36 times (5.1M) and FLOPs 20 times, recovering fissure accuracy within 1 percentage point of dCrack++ (Dice 0.873). Testing shows encoder-level directional modeling contributes more to connectivity than decoder attention, though global context is limited.

THREAD addresses this limitation by replacing local-only refinement with the Strip Attention Bridge (SAB), which captures global directional context through fully parallelizable strip pooling with linear complexity O(n) and the Topology-Preserving Gate (TPG) with Sobel-inspired initialization. SAB reduces fragmentation by 46% on Crack500 (Component Ratio: 1.044 vs. 1.941; values closer to 1 indicate better topology preservation). THREAD achieves state-of-the-art results on four crack benchmarks while using 48.6% fewer FLOPs than previous methods (9.33G vs. 18.16G), enabling 92 FPS on a GPU and 29.0 FPS on a Raspberry Pi 5.

Cross-domain analysis confirms the central hypothesis: dCrack++ establishes strong baseline performance on retinal vessels and fissures, STRIPE achieves competitive crack segmentation without domain-specific training, and THREAD attains a Dice score of 0.866 on fissures with the best topology preservation on CrackSeg9k. These results show that efficient, domain-agnostic architectures can match specialized accuracy while improving structural continuity.