Research on pupil segmentation network based on improved TransUNet

To address the critical challenge of pupil segmentation accuracy in medical image analysis, this paper proposes an innovative pupil segmentation method based on an enhanced TransUNet deep neural network. The proposed architecture introduces three key improvements: a novel multi-kernel self-selective attention mechanism that optimizes local feature extraction through weighted fusion of different convolution modules; a modified cross-attention Transformer that enables efficient parallel processing of spatial information in both horizontal and vertical dimensions; and an advanced feature fusion module that effectively integrates global and local information through multi-scale feature weighting. Extensive experiments on the OpenEDS dataset demonstrate the method's superior performance, achieving state-of-the-art results with average precision of 97.7%, average recall of 97.41%, and average Dice coefficient of 98.15%. These significant improvements in segmentation accuracy not only advance the field of medical image analysis but also provide a robust foundation for applications in ophthalmological diagnosis, human-computer interaction, and cognitive science research. The proposed methodology offers valuable insights into designing efficient attention mechanisms for medical image segmentation tasks, contributing to both theoretical understanding and practical applications in healthcare imaging systems.