Accurately locating and segmenting objects in medical images is crucial for providing precise data support. Traditional Chinese medicine diagnosis relies on pupillary features to reflect the condition of internal organs and pathologies. Combining image segmentation technology with pupil features can offer practical tools for medical treatment. However, obtaining and sharing medical image data is challenging, posing a hurdle for supervised learning approaches. Semi-supervised learning, which can use labeled and unlabeled data, can improve accuracy and generalization and is suitable for medical image processing. This method combines deep learning and traditional Chinese medicine to achieve medical image pupil segmentation with limited data, paving the way for disease discrimination. It effectively addresses the scarcity of data and the inadequacy of supervised learning methods for traditional TCM medical images. This approach provides more accurate and practical TCM image analysis and diagnosis tools.