AN APPROACH TO RIGHT VENTRICLE SEGMENTATION IN MR IMAGES
Introduction: Magnetic Resonance Imaging (MRI) is considered a reference modality for non-invasive assessment of cardiac morphology and function. Important information on the heart’s contractility can be obtained from cine MRI images using the short-axis view. MRI imageprocessingimplies segmentation of cardiac ventricles as a prerequisite for evaluation ofventricular function. Variability among subjects makes subjective ventricle segmentation a difficult and time consuming task. We show how a computer vision approach can be used to automatically segment theright ventricle, particularly demanding for its characteristic morphology.
Methods: MRI images used were acquired at the Diagnostic Imaging Centre, Oncology Institute of Vojvodina,SremskaKamenica, Serbia.Automatic segmentation initialises by locatinga region of interest (ROI) containing the heart, using morphological activity detection between consecutive cardiac cycle frames. Due to extremely thin wallof the right ventricle intensity based segmentation is difficult. To overcome this problem, we userobust adaptive contrast adjustmentthat maximizes separation of bright blood volumes and dark cardiac wall and papillary muscles. Finalsegmentation threshold is determinedbyanalysis of intensity distributions in the located ROI.
Results and Conclusions: Proposed automatic ROI definition algorithm successfully locates the heart region in practically all images, whilesubsequent automatic right ventricle segmentation algorithm was able to obtain the necessary ventricle information in all imaging planes except those near the apex and base of the heart. These regions exhibit significant differences in intensity statistics and appearance of the ventricle while contributing little to true ventricle volume. The proposed automatic segmentation approach is therefore an efficient alternative for extracting the right ventriclein MR images(completing in less than a minute) compared to manually segmented images, requiring significantly longer effort. Focus of future work will be on constructing more robust algorithms for near-apex and near-base slices able to model their specific local properties.