Bayesian fusion: Application in medical imaging

In this paper, we present a method of fusion of medical MRI images based on a probabilistic approach to identify a brain lesion. The aim is to obtain a segmented image from two MRI images T1 and T2 representing the same scene captured at the same time by a patient. To develop such an approach, in a first step, pre-processing is necessary (registration and filtering) of two images to prepare them for fusion. Then, considering that the MRI image is made up of three classes “White Matter”, “Gray Matter” and “Cerebrospinal Fluid”, assumes that the resulting image consists of the fusion of these classes. In a second step, the initial parameters of each class (mean and standard deviation) of each image are deduced from a Gaussian modeling of their histograms. These parameters are used to calculate the probabilities a priori of belonging of the pixels to the different classes mentioned above. Finally, a first fusion image is obtained by applying the maximum likelihood method [1]. From the first image, the parameters of each class are updated to be used the second time, and then the same operations are repeated until no pixel changes classes from one iteration to another.