Research On Atlas-Based Segmentation Algorithm For The Infant Brain MR Images

With the development of modern medical imaging,magnetic resonance imaging(MRI)has become an important measure to assist doctors diagnose brain diseases,and the work load of doctors can be greatly reduced due to the computer aided diagnosis system,the diagnosis can be made more timely and more accurate.Compared with the adult brain MR images,the pre-processing and segmentation of infant brain MR images are more difficult.Due to the small brain capacity,lower imaging resolution,the inversion of gray matter and white matter,larger noise,and the partial volume effect,the segmentation algorithms have been presented for adult brain images cannot be applied to infant brain images directly.At present,there is still no perfect algorithm to make a high precision segmentation of tissue or lesion for infant brain MR images.Therefore,in view of the characteristics of infant brain images,the automatic segmentation methods of three-dimensional brain tissue have been studied in the face of healthy infants and infants with white matter lesions respectively.The main research work and results of this thesis are described as follows:(1)After studied in the non-local patch based method,the sparse representation classification has been applied in the label fusion stage instead of non-local patch method.Moreover,the text feature that computed from the gray level co-occurrence matrix has been added in the gray feature,bring more reference information for sparse representation.The improved method is no longer dependent on the similarity between patches,less affected by the gray change,and solves the problem of redundancy information caused by the original method.(2)Multi-atlas fusion algorithm based on dictionary learning has been studied and implemented,and a Fisher criterion based dictionary-learning algorithm for infant brain MR images segmentation is proposed.Firstly,the basic model of dictionary learning has been introduced,and the two major improvement directions have been described subsequently.Face to the classification problems,the thesis has constructed a discriminative learning dictionary,which is calculated the within class scatter and between class scatter of the sparse representation coefficients based on the Fisher criterion,increasing the discrimination of the cofficients.Then,aimed at the characteristics of brain MR image,a separate dictionary should be learned for each pixel in the whole brain image segmentation.(3)An automatic framework for segmenting brain tissue and white matter lesion simultaneously has been studied and implemented.This framework make full use of different modality information.On the basis of previous researches,the lesion segmentation stage and tissue segmentation stage has been mixed together;the probability prior information is added into the Robust FCM model to segment 5 class of brain tissue;and the other modality information is also applied to provide more reference for the detection of white matter lesions,finally we have achieved a result of extracting the lesion precisely without affecting the brain tissue segmentation.(4)The cross validation experiments of these researched algorithms have been carried out,and the comparisons with other algorithms have been made as well.The experimental results show that the algorithms proposed in this thesis can achieve good stability and segmentation accuracy.