Toward a fully automatic left ventricle segmentation using cine-MR images

Left ventricle (LV) function is assessed by manually segmenting short axis cardiac cine magnetic resonance (cine-MR) images. It is a labor, time-consuming, operator biased task. A series of difficulties arise from these images, that make automatic segmentation of the LV a challenging task: (i) misalignment of the LV along the stack, (ii) signal intensity variation over the stack and over the slice and (iii) the presence of papillary muscles. In this thesis, the first steps toward a full automatic LV segmentation algorithm based on a single view of the LV are presented: 1) Automatic crop: selects a sub-volume containing the LV in all images and in all temporal frames from the acquired data. It is based on three assumptions: (i) the LV is close of the center of the image, (ii) the LV is circular shaped and (iii) there is a high temporal variability of the image intensity in the myocardium boundaries due the heart beat. 2) Alignment-by-reconstruction: novel technique to solve the misalignment due to respiratory motion, inspired on the work from Sanches et al. [1] in ultrasound; 3) Segmentation: the LV is segmented using active contours in an energy minimization formulation with gradient vector flow (GVF) as external field. The automatic initialization algorithm here implemented is original, and it is based on the property of intersecting chords. Preliminary tests with synthetic and real data from 17 patients were performed with successful results.