Jiaxing Tian, Fang Wang, Hua Yu, Chunzhuang Yang
By supplementing high-level features with spatial information and progressively strengthening boundary-distance constraints during training, the proposed method improved both pterygium-region segmentation and localization of indistinct boundaries. The resulting pixel-wise segmentation may provide a basis for subsequent lesion measurement and morphological analysis. However, its generalizability across institutions and imaging devices requires further validation using independent external datasets.
PURPOSE: Pixel-wise segmentation of pterygium lesions provides a basis for quantitative assessment of lesion area, the extent of corneal involvement, and morphological changes. However, the relatively small lesion area, limited local texture contrast, and indistinct interface between the pterygium head and the cornea can lead to incomplete segmentation and contour displacement. This study aimed to develop a fully automatic segmentation method that improves both lesion-region identification and boundary localization.
METHODS: We developed an improved DeepLabV3+ framework with Xception as the backbone. An implicit neural representation (INR)-inspired coordinate-conditioned feature enhancement (CCFE) module was inserted between the backbone and the atrous spatial pyramid pooling module to jointly model spatial coordinates and high-level semantic features. Training employed a progressive mixed loss comprising Focal Loss, Dice Loss, and distance-field Boundary Loss. The model was evaluated on 303 anterior-segment images and compared with representative segmentation models from multiple methodological categories. All experiments were independently repeated using five different random seeds.
RESULTS: The proposed method achieved an mIoU of 93.26 ± 0.22%, a foreground IoU of 89.34 ± 0.39%, an HD95 of 11.51 ± 0.58 pixels, and a Boundary F1 score of 0.4923 ± 0.0046, yielding the best mean performance across all four metrics among the compared models. Compared with Swin-UMamba, the strongest competing model overall, the proposed method improved mIoU and foreground IoU by 1.84-3.78 percentage points, respectively, reduced HD95 by 6.71 pixels, and increased Boundary F1 by 0.0525. Ablation experiments showed that coordinate-conditioned feature enhancement and progressive boundary supervision each independently improved segmentation performance, while their combination further improved foreground coverage and reduced large local contour deviations.
CONCLUSION: By supplementing high-level features with spatial information and progressively strengthening boundary-distance constraints during training, the proposed method improved both pterygium-region segmentation and localization of indistinct boundaries. The resulting pixel-wise segmentation may provide a basis for subsequent lesion measurement and morphological analysis. However, its generalizability across institutions and imaging devices requires further validation using independent external datasets.