Research Of Improved Multi-atlas Segmentation Algorithm Based On Infant Brain MR Images

Infant period is the critical period of human brain development.During this period,infants are likely to get brain diseases.It is very important to evaluate the infant brain development status accurately and diagnose brain development abnormal in time.As an important modern medical imaging technology,Magnetic Resonance Imaging has many advantages.First,it can image brain tissue in multi-modality,multi-parameter and multi-direction.Second,it is non-invasive and has no radiation damage to human.Third,MR images have high spatial resolution and high contrast between different type of tissues.Magnetic Resonance Imaging has been widely used in quantitative infant brain development studies and pathology evaluation analyses,where image segmentation plays an important role.However,segmentation of infant brain MR images is challenging due to complex structure of human brain,low quality of images.This paper proposed a multi-atlas based segmentation framework that adopts non-local methods to fuse image patches and further incorporate texture information with image patches and anatomic text-book.In summary,the major work and contribution of this paper include:(1)The theory and the framework of multi-atlas based segmentation method has been studied.Then the non-local patch-based multi-atlas segmentation algorithm has been studied and implemented.The whole algorithm includes image pre-processing,image linear registration,label propagation,histogram-based image intensity normalization,patch-based label fusion.(2)A rotation-invariant image patch has been proposed,used for patch-based label fusion,to overcome the impact of heavy noise and weak discriminative infant MR image patches in segmentation.Specifically,the image-patch was incorporated with the texture orientation information,which extract from rotation-invariant local binary pattern.Furthermore,the author combined the multi-scale feature image patch with the proposed rotation-invariant image patch,to produce accurate and robust segmentation.(3)The author proposed a texture-based searching volume selection algorithm to build anatomic text-books.This method was proposed to overcome the significant difference among different infant brain MR images.Specifically,the proposed method first extracted LBP feature of target image patch's neighborhood and several atlas image searching volume.Then,LBP features of different searching volumes were compared with target image neighborhood.Finally,the most similar search volume was chosen to build anatomic text-book.(4)The proposed method was evaluated on seven group of infant brain MR images by using a leave-one-out cross-validation.The results show that the proposed segmentation methods achieved a high accuracy for the Dice ratio.The influence on segmentation accuracy of different parameters was also studied.