Performance Analysis of Clustering Algorithms in Brain Tumor Detection from PET Images

Abstract

Brain metastases remain fatal and challenging, and their early detection is imperative. With the advancement in non-invasive imaging techniques, positron emission tomography, as a functional imaging, has been widely employed in oncological studies, including pathophysiological mechanisms of the tumors. While manual analysis and integration of dynamic 4D PET images are challenging and inefficient. Therefore, automated segmentation is adopted to improve the efficiency and accuracy. In recent years, clustering-based image segmentation has been gaining popularity in detecting tumors. This thesis applies three clustering-based algorithms to automatically identify and segment metastatic brain tumors from dynamic 4D PET images of mice. The clustering algorithms used include K-means and Gaussian mixture model clustering in combination with pre-processing principal component analysis, independent component analysis and post-processing connected component analysis. The performances of three clustering algorithms in execution time and accuracy were evaluated by the Jaccard index and validated by time activity curve. The results indicate that K-means clustering is the best-performing among the three clustering methods when combined with independent component analysis, and the post-processing method connected component analysis has significantly improved the performance of K-means clustering.