Application Of Hidden Markov Model In Segmentation Of Brain MR Images

Magnetic Resonance Imaging(MRI)has high spatial resolution,high tissue contrast and non-invasive characteristics,which is widely used in surgical guidance,radiotherapy and other aspects.MR image processing is the basis of diagnosing brain diseases.And image segmentation is the first step of image processing,which can assist in the identifying the morphology of white matter,gray matter and cerebrospinal fluid.The pixel values of brain tissue show a heavy-tailed distribution,which is not suitable to be described by Gaussian random variable.In contrast,α-stable distribution is a more suitable choice.Based on the assumption that the logarithm values of pixel data of all kinds of brain tissue in the images of brain MR follow the α-stable distribution,the hidden Markov random field model is applied to segment brain tissue.Compared with the segmentation methods in the existing literature,our segmentation method has higher accuracy for the same MRI data.The data processed in this article include the brain tissue MR images of 18 group s of normal subjects and corresponding ones after being segmented by experts,which are downloaded from a publicly available database IBSR(https://www.nitrc.org/projec ts/ibsr/).Firstly,energy minimization MICO(Multiplicative intrinsic component optim ization)algorithm is employed to estimate and correct the bias field of MR images.An d we obtain 18 groups of MR images after removing the bias field.Then the pixel values of three kinds of brain tissues(white matter,gray matter and cerebrospinal fluid)are isolated from the corrected MR images.And hypothesis tests are performed on the distribution of pixel values.We use the separation methods of threshold and image superposition to obtain the pixel values of the three types of brain tissue,and use the Chi-square test method to conduct the distribution fitting test of the pixel values from the brain tissue.According to the result of tests,the pixel values of the three tissues neither obey the Gaussian distribution nor the a-stable distribution.Fortunately,the logarithm values of the pixel data of three brain tissues obey the αstable distribution.Finally,the logarithmic pixel values of brain tissue of 18 groups of MR images after bias field correction are substituted into the Hidden Markov model which follows the α-stable distribution.At this time,the observable set is the image pixel value set after bias field correction,and the hidden state set is the segmentation result set of image segmentation.And the segmentation maps of 18 groups of images are obtained.Using Dice similarity coefficient and Accuracy as evaluation indicators,the average Dice similarity coefficient of image segmentation is 0.9298,and the average Accuracy value is 0.9343.Compared with the result in existing literature,for the same 18 groups of data,under the assumption that pixel values follow the α-stable distribution.the average Dice similarity coefficient of segmentation is 0.8941,the average Accuracy value is 0.9063,and the average accuracy Precision under the Gaussian distribution assumption is 0.8636(See Diego Castillo-Barnes et al,2017,FRONT NEUROINFORM),which shows that the segmentation method proposed in this paper significantly improves the accuracy of segmentation.