Brain Tumor Semi-automatic Segmentation on MRI T1-weighted Images using Active Contour Models

Abstract

Brain tumor is a collection of abnormal growth in brain tissue. One of the methods to diagnose brain tumor is using magnetic resonance imaging (MRI) to produce images of brain tissue, on which the radiologist will perform manual segmentation of the tumor boundary. Manual segmentation poses a challenge in a large number of images. A Computer Aided Diagnosis (CAD) system can be designed to perform an automated segmentation of tumor boundary, thus providing more efficient and objective results. In this work, we compared and analyze the performance of snake active contour (SAC), morphological active contour without edge (MACWE), and morphological geodesic active contour (MGAC) segmentation algorithms on 3049 brain MRI T1-weighted images containing glioma, meningioma, or pituitary tumor. The performance of these algorithms quantified using the Jaccard Similarity Index (JSI) and the Hausdorff Distance (HD). The best segmentation results were obtained by the MGAC with the average JSI and HD of 71.18% and 4.04 pixels, respectively. The JSI of MGAC segmentation was highest for meningioma (77.94%) and lowest for glioma (66.31%) while a random shift in contour initialization affected the glioma and pituitary tumors more than the meningiomas, as shown by the respective post-shift JSI of 76.42%, 76.84%, and 85.98% accuracy for glioma, pituitary, and meningioma.