Segmentation of Basal Nuclei and Anatomical Brain Structures Using Support Vector Machines

Abstract

Segmentation of structures inside of the brain are essential for planning computer assisted surgery. Structures such as basal nuclei are difficult to detect in MR images because they have fuzzy edges, and exhibit few changes on the gray level intensities relative to the anatomical structures surroundings. The traditional techniques of image processing thus cannot be used to segment the basal nuclei. We propose a new processing pipeline conformed by five modules: 1) Image acquisition on DICOM format using MRI inside clinics or hospitals. 2) An image pre-processing for improvement on its contrast in a specific window. 3) An anisotropic filter applied to these images. 4) An automatic selection of input vectors for the Support Vector Machine applied. This selection expands radially, where the operator give only a center point of the structures to be detected. This software searches in a radial direction for the supporting vectors that belong to the basal nucleus, and for those corresponding to the background, based on the histogram. 5) Once the Support Vector Machine is trained in the previous module, the generated model serves to classify the image. Finally, the structures are properly detected and segmented, and are then validated by a neuro-anatomist by comparing the segmentation made by our software with a manual segmentation. In conclusion, we propose a new processing pipeline that includes a classifier using Support Vector Machines (SVM) for segmentation that can be used for image guided surgery.