Deformabic Medical Image Registration Based On Sample Equalization Mechanism

Medical image registration is of great significance for medical image analysis and treatment planning.Medical image registration model based on deep learning greatly improves the efficiency of registration by acquiring common feature parameters,which is a research hotspot.However,most of the medical image registration models have the problem that the local registration effect is not ideal,which directly affects the results of medical image analysis.To solve this problem,the following research work is carried out in this paper:Based on the classical supervised learning and unsupervised learning medical image registration models,a large number of experiments are carried out,and the registration results of the two models are comprehensively analyzed from the perspective of voxel sample distribution.The experimental results show that in the process of training,the number distribution of voxel samples with different difficulty degrees is quite different,and there is obvious sample imbalance phenomenon,which makes the model unable to learn well for a small number of difficult samples,which is the main reason affecting the registration effect.In order to solve the problem of sample imbalance,this paper improves the registration model based on supervised learning by referring to the research ideas of sample equilibrium mechanism in the field of target detection.Taking the error distance between the predicted deformation field and Ground Truth as the evaluation index of the difficulty degree of the sample,the contribution degree of the difficult sample is improved by focusing coefficient.On this basis,the sample adjustment factor matrix is established and used to improve the loss function.Taking the U-net network combined with the sample equalization mechanism as an example,the experimental results on multiple data sets show that the registration effect of the sample equalization mechanism for local details is improved,and the overall registration accuracy is improved.In the process of applying the sample balance mechanism to unsupervised learning registration model,due to the lack of the guidance of Ground Truth,it is unable to directly establish the sample difficulty evaluation index.Starting from the spatial neighborhood information,this paper takes local variance(LV)and local cross-correlation(LCC)as the evaluation indexes of sample difficulty,and then constructs the sample adjustment factor matrix to improve the loss function.Then we integrated the sample equilibrium mechanism into classical registration models such as VoxelMorph and SYM-Net.The experimental results show that the registration effect of local complex regions is significantly improved without affecting the overall registration accuracy.Furthermore,the registration accuracy in Dice shows that 33 annotated regions(54 in total)in LPBA40 dataset and 52 annotated regions(62 in total)in Mindboggle-101 dataset are improved.