Research On Few-shot Human Brain Segmentation Algorithm Based On Metric Learning

The use of Deep Neural Networks(DNNs)for medical image segmentation has made significant advancements in recent years.However,deep neural networks usually require a large amount of data and annotations for training.Image acquisition equipment and patient privacy make data and annotations expensive and costly to obtain for medical images.To solve this problem,this paper proposes a unified framework for generalized low-shot(one-shot and few-shot)medical image segmentation based on Distance Metric Learning(DML).Most of the existing methods solve the problem of lack of annotations while assuming abundance of data,but the framework proposed in this paper works in the case of extremely scarce data and annotations,which is an ideal method for dealing with rare diseases.The framework learns the multimodal mixture representation for each category through DML,and performs dense prediction based on the cosine distance between the pixels’ embedded vector and the category representations.The multimodal mixture representations effectively utilize the similarity between subjects and the variations within the category to overcome the overfitting caused by extremely limited data.In addition,this paper proposes adaptive mixing coefficients for multimodal mixing distribution to adaptively select the modes that are more suitable for the current input.These representations are implicitly embedded as weights of the fully connected layer,so that the cosine distance can be effectively calculated through forward propagation.In addition to the adaptive mixing coefficient,the framework also adopts three improved strategies that are suitable for medical images to make the framework obtain better performance.Experiments are performed on brain magnetic resonance images,compared with standard DNN-based(3D U-Net)and classical registration-based(ANTs)methods,the framework proposed in this paper has obtained superior results in few-shot segmentation.When a single sample with annotation is used for segmentation training,the Dice coefficient on the brain MRI image reaches 81%,compared with 52%and 72%for U-Net and ANTs,respectively.In order to further verify the ability of the framework to segment other image modalities and other body tissues or organs images,we use the proposed framework to conduct experiments on abdominal CT and breast ultrasound images,and good results are also obtained.In addition,although the usage scenarios for this research is different from some exsiting methods,this paper compares the framework with some state-of-the-art methods(including low-shot learning and non-low-shot learning)through reasonable modification of experimental conditionsthe to further illustrate the adaptability and effectiveness of the proposed framework.