Segmentation d’images de résonance magnétique de diffusion pour l’aide au diagnostic des pathologies cérébrales fœtales.

Abstract

Fetal MRI is a complementary modality to ultrasound examination. The segmentation of fetal MRIs is a recent method, quickly becoming an essential step in many clinical applications for antenatal monitoring of maturation or brain malformation. However, the artifacts inherent in this type of image and the low resolution of these images are at the origin of the difficulties encountered in the segmentation of these images. To overcome these, we propose in this memory, two methods of segmentation: (i) the first one is based on the geodesic active contours applied to adult MRIs for the automatic detection of the brain lesions, (ii) the second one is based on the modification of the fuzzy segmentation to achieve the classification of fetal brain MRIs. The first method is a combination of the geodesic active contours function and the Gradient Vector Convolution (GVC) in order to improve the detection of the boundaries of the objects to be segmented. The model has been tested on adult MRIs that contain brain tumors or multiple sclerosis lesions. This model has been satisfactory in adults but not in the fetal case. This led us to use an unsupervised classification especially with fuzzy segmentation models. We have therefore, integrated the local and non-local distance in the term of attachment to the data of the RFCM (Robust Fuzzy C-Means) energy function, and integrate non-local means in the regularization term. An algorithm based on layer-bylayer segmentation of fetal brain regions, has been developed. Quantitative and qualitative results on real cerebral fetal images showed the efficacy and robustness of the proposed method compared to the methods described in the literature.