Research On Medical Image Segmentation Algorithm Based On Deep Learning

Medical image segmentation can obtain abundant anatomical structure and pathological information,which is the key to auxiliary diagnosis,treatment and postoperative evaluation.Because of the complexity of medical image structure,such as fuzzy boundary and irregular shape of tumor,accurate 3D brain tumor contour is challenging.Therefore,image segmentation method based on deep learning is considered to be an indispensable tool in the process of medical image segmentation,which solves many problems.However,medical radiation images in the field of biomedical image analysis are mainly obtained by different radiology instruments,such as heart images,which can be obtained by X-ray,CT and MRI.Due to different imaging physical principles and different modes of data distribution mismatch,the established network model may not perform well.Therefore,in view of the above-mentioned problems in medical image segmentation,this paper mainly carries out the following research:(1)Aiming at the problem that the structure of actual medical image is complex and it is difficult to achieve accurate segmentation,this paper proposes a medical image segmentation method based on MAC-IOUC-3DSFCNN,and verifies it on brain tumor data set.Firstly,3D residuals are added to the classic 3DU-Net model to further deepen the network structure and obtain more abstract features.In addition,adaptive blocks are introduced to adapt feature mapping,so as to promote the bidirectional flow of feature information.Then,Haar-like features and multimodal information are introduced to construct a context framework to obtain a more accurate 3D structure of brain tumor lesions.Finally,an improved IOU constraint loss function is proposed to solve the problem of unbalanced data,so as to further improve the 3D segmentation performance of network model.The experimental results show that MAC-IOUC-3DSFCNN can achieve accurate segmentation of medical brain tumor images under the conditions of large noise and unclear boundary.(2)In order to solve the problem of inconsistent data distribution due to different imaging physical principles and different modes in actual medical diagnosis,a cross modal medical image segmentation method based on UMDA-SNA-SFCNN is proposed and verified on cross modal heart data set.Firstly,the spatial attention mechanism is added to the traditional domain adaptive network structure to highlight important regions and suppress irrelevant regions,so as to reduce the negative migration in the alignment process of different domains and improve the segmentation accuracy.Then,a multi-layer domain discriminator is added to the adaptive learning of confrontation domain to connect multi-layer features and segmentation mask,so as to achieve fine-grained alignment of feature domains.Finally,a large number of experiments are carried out on public data sets.The experimental results show that this method is superior to some traditional domain adaptive methods and realizes unsupervised segmentation of different data distributions.