Using Gaussian Mixture Model Clustering for the Automated Segmentation of Cerebral Arteriovenous Malformations in Magnetic Resonance Images

Abstract

Purpose: The complex vascular structure of cerebral arteriovenous malformations (AVMs) is a serious impediment to radiosurgical treatment planning. Precise delineation of AVMs is crucial to the effectiveness of Gamma Knife radiosurgery (GKRS) and efforts to minimize the risk of adverse radiation effects; however, manual segmentation methods are labor-intensive and prone to variability. Methods: This retrospective study analyzed T2-weighted MRI from 25 AVM patients who underwent GKRS. A panel of three neurosurgeons manually labeled the AVM components to establish a ground truth dataset. Fuzzy c-means, K-means, and Gaussian mixture model (GMM) algorithms were used to automate the clustering of the AVM nidus, brain tissue, and cerebrospinal fluid. Segmentation accuracy was assessed using Dice Similarity Coefficient (DSC). Results: When applied to T2-weighted MRI, all three algorithms demonstrated good segmentation capabilities (average DSC > 0.7) in differentiating the AVM nidus and brain tissue. The GMM distinguished itself with the highest DSC (0.826) in brain tissue segmentation. The GMM also exhibited notable proficiency in CSF segmentation, establishing itself as the most powerful and balanced tool for AVM component analysis. Conclusion: Unsupervised machine learning techniques provide an efficient and highly accurate approach to analyzing AVM components within T2-weighted MRI. The automated segmentation of AVM components using the GMM could enhance the precision of radiosurgical treatment planning while providing a basis for future investigations into predicting complications.