Enhancing pupil segmentation with an improved Swin-Unet network

Pupil changes can signal a range of diseases or exhibit different behavioral performance, traditional methods for accurate pupil measurement often face limitations due to non-standard conditions, resulting in reduced robustness in complex scenarios. To address this, this paper introduces an improved Swin-Unet network for pupil segmentation, which integrates Coordinate Attention (CA) and Efficient Multi-scale Attention (EMA) to enhance feature learning. Focal Loss is employed to handle class imbalance, improving the model's focus on challenging samples. The improved Swin-Unet network demonstrates superior performance on the Han <i>et al</i>.’s dataset and CASIA dataset, on the Han <i>et al</i>.’s dataset, the model attained a mDice coefficient of 0.966, and a mHD95 value of 1.641. On the CASIA dataset, the model achieved a mDice coefficient of 0.717, and a mHD95 value of 71.802. The model offers new methodologies for pupil segmentation.