WRCANet: wavelet residual cross-attention network for meibomian gland segmentation

To address challenges in automatic meibomian gland segmentation for meibomian gland dysfunction (MGD) diagnosis, such as size variation and blurred edges, this study proposes a wavelet residual cross-attention network (WRCANet). The network utilizes wavelet residual blocks to preserve boundary details and a cross-attention-based global feature aggregation module to handle varying gland sizes. Evaluated on both in-house and public datasets, WRCANet achieves Dice scores of 90.20% and 84.60% respectively, outperforming other state-of-the-art methods. By effectively enhancing gland boundary discrimination, the model shows significant potential for precise morphological analysis and auxiliary clinical diagnosis of MGD.