Research And Implementation Of Medical Image Segmentation Generalization Algorithm

With the development of deep learning in the field of medical image application,the method based on deep learning has replaced the algorithm of manual feature extraction,and has exceeded the accuracy of manual annotation.However,it is difficult to annotate medical images,resulting in scarce data volume,which cannot meet the demand of large data volume required by deep learning training.A simple solution is to join in training of data from multiple sources,but various types of medical image acquisition equipment and parameter Settings is complex,resulting in huge differences between hospital to collect image style,namely,the phenomenon of the domain drift,when training to complete in the source model in the forecast of the new target domain,usually performance will decline a lot,This problem of poor generalization has become one of the major obstacles in applying deep learning models to clinical practice.In order to overcome the domain the impact of the drift phenomenon,this thesis proposed a new feature regularization LSRML yuan learning framework,the framework can be known to the target domain of the nuclear magnetic resonance(MR)images to produce more accurate segmentation results,the segmentation based on meta learning generalization framework by fitting virtual source domain and target domain to simulate the real differences between the domain of drift.In addition,a domain discriminator module is designed to generate category prediction of potential features,which can confuse the sensitivity of the trunk network to data distribution and force the trunk model to learn taskrelated semantic features to complete the segmentation task,and then,an image reconstruction module is designed,which reconstructs the foreground and background of the target domain image respectively,so that the encoder can be encouraged to learn the more discriminant features of the unmarked image,and the decoder can reconstruct the target mask more easily by using this feature.In this thesis,the proposed algorithm is evaluated by using the prostate mask annotation provided by the radiology department of a third class A hospital in Beijing and the data from six public data fields.The results show that the proposed algorithm has good segmentation and generalization performance.