The Research Of Brain Extraction Algorithm Based On Fully Convolotional Network

Brain extraction from magnetic resonance imaging(MRI)data is an essential step in many neuroimaging applications.The accuracy and robustness of brain extraction,therefore,is crucial for the accuracy of the entire brain analysis process and it will affect following image processing directly.There are two kinds of methods for brain extraction: manual brain extraction and automatic brain extraction.The accuracy of manual way is high,however,it costs a lot of time and labor.Furthermore,this method depends on professional skills of the operator and is affected by the operator's subjective feeling.The automatic way is relatively easy and can be precise and stabilized.Hence,the study of extracting brain automatically has become mainstream.CNN(convolutional neural network)and FCN(fully convolutional neural network)is introduced in this paper.Afterwards,a method of brain extraction which is based on FCN is proposed.We also put forward a further classification for pixels around the border of brain due to the fact that most of errors of the extraction results are around the boundary between brain and non-brain.In addition,because of the excellent character of SegNet network,our proposed method is high-efficiency.The main work in this thesis is as follows.(1)Based on the analysis of current brain extraction algorithm and the defect of extraction brain with CNN method,wo proposed the method of FCN.(2)We proposed a SegNet based brain extraction method,and because of wrong segmentation of the border of brain,we used two different methods to do further segmentation.Finally,the results is better than before.(3)Tesing and analyzing proposed algorithm on real dataset,results manifest that our method is useful.We used three public dataset(OASIS,IBSR and LPBA40)to validate the effectiveness of our method.For the OASIS,our approach outperforms the other popular brain extraction packages and two well-established deep learning based methods with a mean Dice coefficient of 0.983.For the other two datasets,though the Dice coefficient of our approach is inferior to other deep learning based methods,the Sensitivity and Specificity of our approach is superior to theirs and for the IBSR,the mIoU(mean Intersection over Union)of our method is the best.The computational cost of our method is superior to other methods because of the network we chose.The final results of experiments manifest that the segmentation results of this method is accurate and better.The proposed method may prove useful for large-scale studies and clinical trials.