Optimal Transport Feature Selection Method For Medical Image Segmentation Transfer Learning And Its Application

Nowadays,image segmentation has become one of the hotspots of in-depth learning.Medical image segmentation technology has developed successfully,but in the clinical environment,due to the differences of hospitals,medical equipment,patient groups and other factors,the generalization ability of these models is often difficult to achieve the expectation.Transfer learning can associate two domains and realize the accurate segmentation of target domain samples with the help of labeled source domain samples,so as to improve the performance of the model.Most domain adaptive transfer learning models will use all the features of two domains for training.and often incorrectly use domain independent features for model learning,which may have a negative impact on the performance of the model.At present,most of the transfer learning feature selection methods are aimed at the field of image classification,and most of the algorithms are difficult to solve.In the field of image segmentation,there is no reliable feature selection method.This paper proposes an optimal transport feature selection method for medical image segmentation and transfer learning,that is,the general feature selection module of transfer learning.The module is divided into two stages,sample selection stage and feature selection stage.In the sample selection stage,this paper proposes the segmentation accuracy weight,and uses this weight to re weight the cost matrix of the original optimal transport problem,and constructs a weighted optimal transport method to select samples,so that the selected samples can more specifically describe the segmented image features of the source domain and the target domain,increase the transport accuracy,and make the number of samples in the two domains equal,so as to carry out the feature transport in the next stage.In the feature selection stage of the feature selection module in this paper,the entropy regularization optimal transport is used to match the features of the sub samples of the source domain and the target domain with the same number.Through the moving transport features between the domains,according to the obtained optimal coupling matrix,a list f sorted according to the similarity of the same features in the representation space of the shared source domain and the target domain is obtained,and the domain invariant features for image segmentation are selected..This method makes full use of the original feature information of the image to avoid some feature selection methods can not effectively use the original feature information of the image,and some use the eigenvalue decomposition method to increase the amount of calculation and solve the problem of difficulty.Finally,according to list F,the selected features are used in the domain adaptive transfer learning segmentation model.In this paper,the COVID-19 segmentation data set is used for experiments,and the module is applied to unsupervised domain adaptive transfer learning tasks and unsupervised domain adaptive transfer learning tasks.For sample selection,the random selection,ot1 and OT2 methods are compared with the segmentation accuracy weighted optimal transport in this paper,and then the ascending feature selection and random feature selection methods are compared with the feature selection methods in this paper to highlight the advantages of this method.Experiments show that the dice indexes of seg-jdot,e-uda and self-assembly models after using the transfer learning general feature selection module reach 78.1%,78.3% and 77.9% respectively,which are higher than 75.1%,76.3% and 76.7% of the dice indexes before use.It can be seen that this method can improve the effective adaptation of adaptive algorithm and model performance.