Anais do X SIBGRAPI (1997) #ART64


Segmentação do Miocárdio do Ventrículo Esquerdo em Ressonância Magnética Usando Conectividade Fuzzy e Expansão de Camadas

Sérgio Shiguemi Furuie, SP&D, Divisão de Informática, InCor/HCFMUSP
Griselda E. Jara, INPE
Marina F. S. Rebelo, SP&D, Divisão de Informática, InCor/HCFMUSP
Luiz Cláudio Parzianello, SP&D, Divisão de Informática, InCor/HCFMUSP
Marco Antônio Gutierrez, SP&D, Divisão de Informática, InCor/HCFMUSP

Abstract. A truly 3D technique for segmentation of left ventricle (LV) myocardium in MRI is presented, based on fuzzy connectedness and incorporation of voxels with minimum cost. Direct segmentation of LV muscle based only on intensity or texture is difficult because of inhomegeneity and background with similar characteristics. We propose a two step technique: first, segmentation of LV cavity and then an orderly search of LV muscle keeping roughly the endocardium morphology. The LV cavity is extracted in all volumes applying competitive fuzzy connectedness between cavity and myocardium. From the endocardium surface, connected voxels with minimum insertion cost and below a given limit are incorporated in an orderly fashion (layers), emulating a balloon expansion model. The cost function for incorporating a voxel has three components: a) spatial gradient to avoid other objects with distinct boundaries; b) textures to allow some local inhomogeneities; c) shape similarity to the endocardium surface in order to restrict leaking to other organs with similar texture. Studies with phantoms (contrast=48 %) were carried out and presented good results (less than 6 % of misclassified voxels) for SNR up to 12 dB (Gaussian noise). The proposed technique was also applied to gated-MRI of heart, closely matching manually segmented LV myocardium in images with good quality.

[ full version | contact authors | project home page ]