Anatomical multiatlas segmentation using local texture statistical properties for matching descriptor with machine learning

The use of anatomical multiatlas methods has proven to be one of the most competitive techniques for brain images segmentation. The majority of these methods are based on visual criteria of similarity between groups of an atlas to select a representative patient image to be segmented. However, this criterion is not necessarily linked to the performance of the segmentation. To overcome this dilemma, we present in this article, a new concept of preselection of an anatomical atlas group, which is based on machine learning and using an adapted descriptor that can give an efficient and more precise segmentation of the patient image. The new descriptor, local texture statistical properties for matching descriptor with only affine registration, is adapted from the local texture of matching (LTEMA) descriptor. The proposed method is tested on real MRI brain images (LONI database provided by USC Neurological Imaging Laboratory), and show the capability and the effectiveness of the proposed local descriptor, it has been compared to three local descriptors: scale-invariant feature transform, speed up robust feature, and LTEMA, as well as the comparison with the registration method. The obtained results show a significant improvement that makes this descriptor recommended for segmentation techniques.