Research On Self-supervised Medical Image Segmentation Algorithm Based On Transformer

Medical image segmentation is an important tool for assisting medical diagnosis and has a wide range of applications in clinical practice.With the rapid development of deep learning,the position of convolutional neural networks in the field of computer vision has been challenged.As a new emerging technology,Transformer has shown outstanding performance in natural scene image segmentation,which is attributed to its ability to capture distant features.To study the effect of Transformer-based segmentation models on medical images,this thesis proposes a new self-supervised medical image segmentation architecture.By designing proxy tasks for pre-training,the difficulty of the lack of large-scale annotated data in medical images is solved.Extensive experiments were conducted on two mainstream datasets to validate the effectiveness of the proposed TF-UNet model in medical image segmentation tasks.Moreover,comparisons with other mainstream models showed that the proposed model performed well in most scenarios.Furthermore,to balance the accuracy and efficiency of segmentation,this thesis further proposes an improved model,TF-KDNet,based on knowledge distillation.It uses the TF-UNet model as the teacher model and trains a smaller student model to transfer the knowledge of the teacher model by minimizing the distance between the outputs of the teacher and student models.Through experiments,it has been demonstrated that this approach can significantly reduce computational costs and memory usage while maintaining accuracy,providing more alternative options for medical image segmentation.