Research On Few-shot Medical Image Segmentation Based On Metric Learning

The segmentation of medical images is a key step for doctors to diagnose and treat,but manual labeling is time-consuming and laborintensive,and there are differences in the standards of segmentation results.With the development of deep learning,deep learning algorithms based on fully supervised learning can achieve good segmentation results under the training of a large amount of labeled data,but the practical application is difficult due to the difficulty of obtaining professional labels and the low generality of the model.The emergence of few-shot learning solves some of the problems,but the essence of few-shot learning is to enable the model to acquire learning ability after training a large amount of labeled data,and then only a small number of new samples are needed as a support set in the inference process without retraining the model.The problem of difficulty in obtaining annotations remains unsolved.In this paper,the segmentation method based on graph theory is used to obtain superpixel pseudo-labels of the original image to replace the real labels,the VGG-16 network is used to complete the feature embedding,and the attention mechanism is introduced to improve the adaptive local prototype module to obtain more local information to reduce the class imbalance.Combined with metric learning to measure the query set and prototype based on cosine similarity to complete image segmentation,use the hybrid loss function of cross entropy and Dice to replace the cross entropy loss function,and introduce prototype alignment regularization in the training process to reuse support set data information.Chapter 4 improves the feature embedding network,redesigns the network structure based on DenseNet,adds the Inception module and channel attention mechanism to obtain more discriminative features,improves the accuracy of subsequent prototype metrics,combines the regional loss function and the boundary loss function,Cope with class imbalance problems that are intractable with single-region loss functions.The experiments in this paper are based on the 2017 Automatic Cardiac Diagnosis Challenge(ACDC)data set,and the effectiveness of the algorithm in this paper is verified through the comparison of various algorithms and their own improved modules.