With the rapid development of national fitness and professional sports, the problem of sports trauma has become more and more prominent, which has caused serious impact on the training and competition of athletes. It is urgent to improve the level of diagnosis and treatment. The analysis technique of identifying and diagnosing cells and tissues for athletes in medical images with the help of deep learning algorithms has gradually become a popular research direction in the field of medical image diagnosis. Convolutional neural network (CNN), as an efficient deep learning algorithm, is widely used in the field of medical image diagnosis. However, since CNN models need to initialise the parameters before training, various problems may arise when the initial parameters are not properly selected. Firstly, for the initial weights of the CNN model, the traditional method is to use random initialisation, which leads to problems such as slow training speed and low diagnostic accuracy of the model. Secondly, for the selection of the hyperparameter of the model, the traditional method is to use manual adjustment or grid search, which not only consumes a lot of time and computational resources, but also usually fails to select the most suitable hyperparameter, which leads to the problems of lower diagnostic accuracy of the model. In order to solve the above problems, this paper firstly proposes a new self-supervised medical image segmentation architecture. By designing an agent task for pre-training, the model is better able to extract and process the visual information of medical images, and then fine-tuned on the segmentation task, as a way to solve the difficulty of the lack of large-scale labelled data for medical images. The effectiveness of the proposed algorithm for athletes in the medical image segmentation task is verified through a large number of experiments conducted on two mainstream datasets. And the comparison with other mainstream models shows that the model performs well in most scenarios.