Research And Application Of Intelligent Registration Method For Multimodal Medical Images

Image registration is a commonly used technology in medical image processing,which is widely used in various fields of medicine,including lesion detection,disease diagnosis,surgical planning,surgical navigation,and efficacy evaluation.With the vigorous development of a variety of medical imaging technologies,from morphological imaging reflecting anatomical structures to functional imaging reflecting organs and tissues,medical imaging of different modalities carries a wealth of medical diagnostic information from different angles.Rich functional complementary information for clinical treatment can be provided from the fusion processing of images of multiple modalities.The basis of multi-modal image fusion is to register these images.Therefore,the registration technology of multi-modal medical mirrors has gradually become the focus of research.However,there exits many limitations in the traditional multi-modal medical image registration method,which limit its use in actual clinical scenarios,including the problems of poor adaptability and long registration time.In recent years,the powerful functions of deep learning in the field of image processing have made it a mainstream method for studying image problems.In medicine,deep learning has excellent performance in organ segmentation,tumor detection and other tasks.Considering the above situation,this paper proposes a network model based on deep learning to solve the registration of multi-modal medical images.At the same time,considering the difficulty of directly registering multi-modal images,this paper proposes a method of generating medium modalities to ensure the validity of the registration between multi-modal images.The main contributions of this work are as follows:(1)A new method SCG-TPL to generate medium modality is proposed.The m Dixon MR sequence for the challenging lower abdomen generates a medium modality that can simulate CT images,which is synthetic CT images.SCG-TPL is based on the framework of patch learning,which uses the knowledge lever transfer fuzzy c-means clustering(KLTFCM)as a global model of the framework to obtain preliminary classification results,and KL-TFCM can reasonably overcome the individual differences between different samples;The semi-supervised model Lap-SVM is used as a local model of the framework for fine classification.The results show that SCG-TPL only requires a small amount of labor cost and can bring low time overhead,and can produce good-quality synthetic CT images for abdomen with only m Dixon MR sequences.(2)For multimodal medical image registration,a deep network M-i VM based on medium modality is proposed.The network can register MR image to CT image.M-i VM is an improved diffeomorphism registration network model,which is divided into registration network module and spatial transformation module.The structure of the registration network module is a U-net network,whose purpose is to generate a registration field.The spatial transformation module is a deformed spatial transformation network structure(Spatial Transform Network),whose purpose is to transform and interpolate the registration field to obtain the final registration image.The results show that M-i VM can not only accurately realize the registration between multi-modal images,but also reduce the registration time and improve the efficiency of registration.