Research On Lung 4D-CT Image Registration Algorithm Based On Prediction

Lung 4D-CT provides important imaging information for the diagnosis and treatment of lung diseases.It adds time axis to traditional 3D-CT images,monitors the respiratory of human body through external devices,and dynamically collects CT images of patients’ whole respiratory cycle.The lung 4D-CT images reflect the changes of lung tissue with respiratory movement,which effectively reduces the artifact caused by the breathing movement in the image.The registration of lung 4D-CT images is very important for the motion estimation of lung tissue,lung ventilation,localization of lung tumor and the accuracy of radiotherapy.It’s also widely used in lung segmentation,monitor the development of the disease and to judge the effect of the treatment.But the lung organ is a typical motor organ,it is disturbed by the breathing movement and the heart pulsation,which makes the collected image of lungs more deformed.And during the process of breathing,the density of lung image will be changed by the influence of stimulated air.The above two factors become the biggest challenge in the registration of lung 4D-CT images.To solve above questions in lung 4D-CT images registration,two kinds of lung 4D-CT image registration algorithms based on prediction are studied to improve the accuracy of lung image registration.Firstly,we study the registration algorithm based on block similarity prediction.It assumes that if the image block is similar to the gray feature,the corresponding deformation field should be similar when the image is registered to the same reference image.In this algorithm,we take advantage of the multiphase characteristics of the lung 4D-CT.First of all,the images of different phases are registered to the selected reference images,and their corresponding deformation fields are obtained.After that,the floating image is partitioned,for each target image block,the most similar candidate image blocks are found on each phase image according to a certain similarity rule.Then,we define a criteria function to filter candidate image blocks.The deformation field coefficient is solved by the gray relation between the target image block and the selected blocks.Finally,the initial deformation field of the floating image is predicted by using the non-local mean algorithm,and then the middle image can be obtained.The intermediate image obtained from regression prediction is more similar to the reference image,and the local deformation is smaller than the floating image between the reference image.Therefore,the fine registration is less difficult and the registration effect is better.Experimental results show that the accuracy of the algorithm is better than the traditional registration algorithm.Secondly,we study registration algorithm based on regression prediction.The experimental results of block-based similarity prediction show that the idea of predicting the initial deformation field and the intermediate image in order to improve the registration accuracy is feasible for the larger deform between floating image and the reference image.So we further study the learning based prediction algorithm.This method makes full use of the similarity of lung breathing motion while taking advantage of the multiphase characteristics of lung 4D-CT.We first register the different phase images to the reference image,and build the training sample set with the known image appearance and the corresponding deformation field.We use multidimensional support vector regressor to establish the regression model of image appearance and deformation.Based on the image appearance-deformation regression model,the floating image is partitioned into blocks and input into the regression model.The prediction deformation field corresponding to the image block is obtained.After that,the deformation field is spliced into the whole image.Then the predicted initial deformation field and the intermediate image can be obtained accordingly.Finally,the registration between reference image and the intermediate image are refined.The experiments are carried out on the lung 4D-CT dataset.The experimental results show that the algorithm is superior to the traditional image registration algorithm in both visualization and quantization.