An automated pixel classification method using surface expansion Application to MRI image sequence

This paper describes an automated pixel classification method using surface expansion. The originality of this work resides in the definition and use of small pictures (called "imagelet") of increasing size centered on the pixel of interest. This allows for the extraction of a set of local and global parameters associated to the pixel investigated. This set of parameters then permits body tissue separation. Classification was obtained using a multilayer artificial neural network. The new approach proposed was applied to main lower limb tissue classification in MRI image sequences. Four kinds of body tissue were taken into consideration in this study (muscle, adipose tissue, cortical bone and spongy bone). A database consisting of 1400 prototypes was created in order to evaluate this method's performances. The classification success rate was found to be 87%. This method therefore proved to be reliable and robust to analyze MRI image sequences in 20 lower limbs, representing about 2000 pictures