On information fusion to improve segmentation of MRI sequences

Abstract

Abstract The paper deals with the problem of segmentation of MRI sequences of vertebrae, in the form of images of their multiple slices, using the Dempster–Shafer theory. This leads to the study of 3-D deformations of the scoliosis. The motivation comes from the inadequacy of the existing techniques based on X-ray image analysis. Such analysis cannot deal with, on the one hand, the complex anatomical structures (“scoliotic rachis”), and the spongy tissue peri-rachidian, and, on the other hand, the choice of slices and the problem of the residue irradiation present in each examination. The main contributions of the paper are: • New architecture for the fusion of MRI data sets. • A novel method to exploit the information contained in MRI sequence. • Model for knowledge representation adapted to specificity of information available (Dempster–Shafer theory). • Choice of the discriminating parameters for the statistical expertise. • Construction of the belief functions. • Choice of the decision criterion. Starting from segmentation by active contour (snake) [Deformable contour: modelling, extraction, detection and classification, Ph.D. Thesis, Wisconsin–Madison University, 1994; Proceedings of the 15th International Conference on Pattern Recognition, vol. 4, Barcelona, 2000, p. 17; Int. J. Comput. Vision 1 (3) (1987) 211], we upgrade it in an attempt to present the doctor with a degree of belief concerning their membership of the contour of the vertebra. We illustrate the proposed fusion architecture by application to actual MRI sequences of the vertebrae, and include perhaps the first example of 3-D reconstruction of the lumbar rachis starting from the results obtained during fusion.