Study On Structure Mismatching Problem In Multimodal MRI Image Registration Based On Mutual Information

In clinical science,multimodal magnetic resonance imaging(MRI)is a modern imag-ing technique for medical examination and diagnosis,it provides radiologists with com-prehensive information about tissues and lesions.However,as MRI process consumes plenty of time,patient's movement during imaging couldn't be avoided comprehensively,the same tissue's positions between multimodal images are usually different.It seriously affects radiologists'diagnosis of diseases.Multimodal image registration methods could unify the coordinates in MRI images and match the tissues.It improves the accuracy of medical diagnosis,surgical planning and radiotherapy.In the field of multimodal registra-tion,mutual information(MI)is a widely used similarity metric,whereas MI is sensitive to'structure mismatch phenomenon',the phenomenon affect the accuracy of MI-based registration method.To resolve the mismatch problem in MI-based registration method,in multimodal MRI images,the cause of formation,influence mechanism and solutions for structure mismatch problem were deeply studied in this dissertation.The innovations were summarized as follows:(1)For multimodal MRI images,to analyze structure mismatch's influences on MI distribution,structure mismatch in MI-based registration was deeply studied.The mech-anism of structure mismatch was comprehensively analyzed in perspective of multimodal MRI imaging.Structure mismatch modeling was used to analyze the mechanism.Analy-sis results confirm that,structure mismatch causes the deviation of MI's global maximum,it makes global maximum deviated from the registration target.This conclusion forms a solid foundation for designing multimodality MRI image registration method.(2)The anatomical structure of brain is special—the soft tissues of brain is surrounded by skull.In the multimodal MRI image of brain,caused by tissues around skull,the struc-ture mismatch phenomenon is significant.To deal with the mismatch structures around skull,atlas was introduced to compensate mismatch structure's information,an atlas-based registration method was proposed.Atlas was used as an information bridge between struc-ture mismatch images.Through spatial transformations between atlas and multimodal im-ages,the deformation field between registration images was generated,which indirectly compensated the mismatch structure's information.Then based on the deformation field,the multimodal MRI images were registered.Model images and brain MRI images were employed to test the registration method.Results indicate that,for multimodal MRI im-ages,the proposed method effectively resolves structure mismatch's influences on MI-based registration and improves the registration accuracy.(3)As many organs locates in chest,the anatomical structures of chests are individ-ually different.And the fat tissue around thoracic is an important factor caused structure mismatch phenomenon.For multimodal registration of chest MRI images,to deal with the mismatch fat tissue around thoracic,morphological processing was introduced to com-pensate mismatch structure's information,a morphological structure compensation based registration method was proposed.Active contour model segmentation and morphological processing were combined to compensate the missing fat structure in chest MRI images,which directly compensated the mismatch structure's information.Then we calculated the deformation field between compensation image and reference image.Finally,according to the deformation field,the multimodal MRI images with mismatch structures were reg-istered.Model images and chest MRI images were used in the registration experiments.Results indicate that,for multimodal chest MRI images,the proposed method effectively resolves structure mismatch's influences on MI-based registration and improves the reg-istration accuracy.(4)To further improve the compensation accuracy of mismatch structures,Gener-ative Adversarial Networks(GAN)was used to enhance the structure compensation,a GAN-based multimodal chest MRI image registration method was proposed.By learn-ing characteristics and distribution of tissue structures,GAN generated a compensation image whose contours are more consistent with human tissues.Then we calculated the deformation field between compensation image and reference image.Finally,based on the deformation field between compensation image and reference image,floating image was registered to reference image.Model images and chest MRI images were used in reg-istration experiments.Results indicate that,for chest MRI images,the proposed method obtains higher compensation accuracy and improves the multimodal registration accuracy.In this dissertation,for multimodal MRI images,structure mismatch problem in MI-based registration was deeply studied.Based on the springboard of compensating the missing structures,a series of multimodal MRI image registration methods were proposed.Experiment results confirm the correctness of theoretical analysis and the effectiveness of the proposed methods.