Precision pterygium segmentation using hybrid encoder and multiscale deep learning

Pterygium is a common ocular surface disease characterized by conjunctival fibrovascular proliferation, it poses diagnostic challenges in resource-limited regions. Manual segmentation methods are time-consuming and subjective, necessitating automated solutions to enhance clinical screening efficiency. This study improves the Transunet framework for pterygium semantic segmentation by integrating a hybrid encoder (combining CNN and Vision Transformer) with a novel decoder structure. The decoder incorporates a large kernel grouped attention gate (LGAG) and a multi-scale depthwise convolution block (MSDCB), effectively addressing soft tissue boundary segmentation. LGAG optimizes feature fusion through attention-guided spatial modulation, while MSDCB employs parallel depthwise convolutions to capture multi-scale contextual features. Evaluated on a dataset of 520 anterior segment images, the enhanced model achieved an 83.41% Intersection over Union (IoU), outperforming the baseline Transunet (81.26% IoU). It demonstrates superior performance in soft tissue boundary delineation and morphological adaptability. This work provides an efficient tool for ophthalmic intelligent diagnosis, potentially reducing clinical workloads and improving screening accuracy.