Gamma Gaussian Mixture Modeling for Fibroglandular Tissue Segmentation in MR Images

Abstract

In this manuscript the main contribution is to find out the best way to model the tissue intensity distribution behaviour and the superiority of Gamma Gaussian Mixture model (GaGMM) over the existing approaches which are based on Gaussian mixture model (GMM) for segmenting out the fibroglandular tissues. This has been achieved by defining the probability density function of pixel intensities of fibroglandular tissue from Gaussian to Gamma in breast MR images. In this approach, the pixel intensities of fibroglandular and adipose tissues are modelled by GaGMM. The pre-processing of MR images is done for removal of the ghost artifacts, minimization of the speckle noise, and correction of gain field. Also, the expectation maximization algorithm is employed to estimate and optimize the mixture parameters. The average of the means of Gamma and Gaussian mixtures is chosen as optimal threshold value for the segregation of pixel intensities of the tissues. The results show that the GaGMM is fast and adaptable for fibroglandular tissue segmentation. The results of GaGMM are assessed qualitatively on the dataset with the images and ground truth separately by calculating the performance evaluation measures - Dice Similarity Coefficient (DSC), Jaccard Index (JI) and Dissimilarity Index (DI). The values of DSC, JI and DI of GaGMM are improved by 48%, 60.7%, and 50.5% respectively. The proposed approach improved the performance of segmenting the fibroglandular tissue. The results show the improvement in accuracy of fibroglandular tissue segmentation by proposed distribution approach in comparison to the existing Gaussian distribution approach.