Research On Key Techniques Relevant To Deformable Image Registration Algorithms In Adaptive Radiotherapy

Radiotherapy is one of the three effective means of treating tumor,and its therapeutic outcome depends two aspects:ensuring adequate target dose and sparing normal tissue.Radiotherapy techniques ranging from three dimensional conformal radiotherapy(3D-CRT)to intensity-modulated radiation therapy(IMRT),and then to image guided radiotherapy(IGRT)and adaptive radiotherapy(ART),develop towards both goals.Specifically,3D-CRT can conform beams to targets with complex shapes.On the basis of 3D-CRT,IMRT achieves a conformal distribution of dose.With in-room images,IGRT corrects setup errors which may occur in each treatment fraction.However,due to effects induced by treatment and normal physiological processes,the locations,shapes and volumes of some regions of interest(ROI)may vary during the treatment course,which may lead to inadequate irradiation of tumor or an overshoot of organs at risk(OAR).ART is an effective solution to these problems described above,which adapts to current anatomies by replanning or modifying the original plan.Deformable image registration(DIR)algorithm,as one of the key techniques in ART,correlates the planning images to in-room images.The accuracy of the correlation determines the effectiveness and reliability.Therefore,it is necessary to verify the accuracy of DIR algorithms.Taking this as the starting point,we carried out a series of related researches.First of all,four sorts of representative intensity-based DIR algorithms,which are the optical flow-based(HS,HSLK and FFD),demons-based(OD,MD,SFD,DFD and DISC),level-set-based(LS)and spline-based(BSpline),respectively,are reviewed and implemented in this study.All DIR algorithms but DISC coded in GPU are achieved in MATLAB.For the problem concerning the lack of researches on the validation of the accuracy of DIR with multimodality images,this study retrospectively collects CT and CBCT images from twenty-one patients who were diagnosed with nasopharyngeal carcinoma(NPC)and ask an experienced physician to delinear four ROIs including the bilateral parotid glands,the bilateral submandibular glands,cervial vertebras and vertebral foramen.With DIR algorithms,we propagate contours in the planning CT and compare them with contours outlined by the physician on CBCT images for evaluating DIR accuracy in terms of contour propagation and comprehensively comparing differences among them.The experimental results indicate that all DIRs do not necessarily perform better than rigid registration.Generally,the DIR performe better for bony structures than for soft-tissue organs,and the DIR performance tends to vary for different ROIs with different degrees of deformation as the treatment proceeds.For the problem concerning the lack of complexity of phantom in DIR accuracy validation with physical phantom,this study designs and fabricates a highly anthropomorphic phantom on the basis of the abdomen CT of a real patient.There are various simulational structures including liver,kidney,spleen,stomach,vertebras and so on in this phantom.These simulational structures are similar to those in the real patient in terms of volume,shape and CT value and spatial position.In addition,195 fiducial markers(metal balls with the diameter of 1-2 mm)are evenly implanted in the interior and on the surface of artificial organs,and over areas among simulational organs.Variations of the position of fiducial markers before and after the deformation are regarded as ground truths for DIR accuracy validation.The experimental results show that the vast majority of DIR algorithm can significantly improve the accuracy of the rigid registration.The performance of DIR algorithm is related to image contrast and ROI properties(volume,shape,etc.,).For all solid organs,DIR has a higher accuracy on the surface than in the interior,while for individual organs,differences in DIR accuracy between surface and interior are not found.Some preliminary results of researches concerning DIR accuracy validation have been made in this paper,however,further studies are necessary for some details such as the parameter optimization in the implementation of DIR algorithms,the quantification of manually-contouring error in clinical data validation and the distribution of fiducial markers in physical phantom verification.Moreover,improving the existing DIR algorithms based on these validation results will be our research orientation for the future.For example,according to the conclusion shown from this study,which is the accuracy of DIR algorithm depends on the ROI contrast,we could artificially increase the contrast of some ROIs and add the registration result of local-contrast enhanced images into the registration of original images as a prior information or constraint for improving the registration accuracy of original images.