Research On Medical Image Segmentation Algorithm Based On Meta-learning Theory

With the improvement of people’s living standards,more and more people realize the importance of physical health.For patients,early screening and diagnosis of diseases can effectively reduce mortality.The segmentation of medical images is a very important step for disease screening and diagnosis.In recent years,Significant advances has been made in medical image segmentation tasks based on deep learning.However,most high-performance deep learning segmentation models need massive golden standard data as support,which is undoubtedly a huge challenge for medical images,because the acquisition of medical image segmentation annotations is time-consuming,labor-intensive and is prone to errors due to the different experience of doctors.Therefore,researches on small sample image segmentation is of great value for clinical applications.In this paper,a retrospective study of dynamic contrastenhanced magnetic resonance imaging(DCE-MRI)and ultrasound image data of breast cancer was conducted.Based on the attention mechanism and meta-learning theory,noninvasive and accurate medical image segmentation algorithms was proposed.The main research of the paper is as follows:(1)Aiming at the problem of context information aggregation in breast tumor segmentation,a network for aggregation and segmentation of context information inside and outside the image is proposed.The context information aggregation module inside and outside the image simultaneously considers the context information inside the breast cancer image and the molecular subtype context information of the breast cancer image to further enhance the pixel-level feature representations of the encoded features,and finally improve the segmentation performance of the network.(2)To solve the problem of similar intensity distribution of breast cancer ultrasound images,large differences in tumor shape and blurred boundaries,which make it difficult to be accurately segmented,a hybrid convolutional attention few-shot learning segmentation model is proposed.The model guides the network to extract more robust feature representations in channel and spatial dimensions by embedding channel convolutional attention blocks and spatial convolutional attention blocks in the network.In addition,few-shot learning can improve the generalization ability of the model with only a small number of samples,so that it is easier to deal with the complex breast tumor segmentation task.(3)A bidirectional convolutional recurrent graph attention few-shot learning model is proposed to address the heterogeneity of breast tumors and the need for a large number of labeled samples to train high-performance deep learning segmentation models.To learn the pharmacokinetic prior knowledge rich in DCE-MRI images,a bidirectional gated recurrent unit is used to extract the temporal information in the DCE-MRI image sequence,and a graph attention unit is used to extract the spatial structure information in the tumor corresponding to each sequence.In addition,the few-shot learning scheme driven by immunohistochemistry utilizes the unique immunohistochemical phenotype and molecular subtype characteristics of breast tumors.By learning the MRI imaging phenotype features of specific molecular subtype during training,the network parameters are optimized,improves the generalization ability of the model to different molecular subtype tumors,and achieves accurate segmentation of tumor images in the case of a small number of samples.