Intervening Nidal Brain Parenchyma and Risk of Radiation-Induced Changes After Radiosurgery for Brain Arteriovenous Malformation: A Study Using an Unsupervised Machine Learning Algorithm

Abstract

OBJECTIVE: To assess the sensitivity and specificity of arteriovenous malformation (AVM) nidal component identification and quantification using an unsupervised machine learning algorithm and to evaluate the association between intervening nidal brain parenchyma and radiationinduced changes (RICs) after stereotactic radiosurgery. METHODS: Fully automated segmentation via unsupervised classification with fuzzy c-means clustering was used to analyze the AVM nidus on T2-weighted magnetic resonance imaging studies. The proportions of vasculature, brain parenchyma, and cerebrospinal fluid were quantified. These were compared with the results from manual segmentation. The association between the brain parenchyma component and RIC development was assessed. RESULTS: The proposed algorithm was applied to 39 unruptured AVMs in 39 patients (17 female and 22 male patients), with a median age of 27 years. The median proportion of the constituents was as follows: vasculature, 31.3%; brain parenchyma, 48.4%; and cerebrospinal fluid, 16.8%. RICs were identified in 17 of the 39 patients (43.6%). Compared with manual segmentation, the automated algorithm was able to achieve a Dice similarity index of 79.5% (sensitivity, 73.5%; specificity, 85.5%). RICs were associated with a greater proportion of intervening nidal brain parenchyma (52.0% vs. 45.3%; P = 0.015). Obliteration was not associated with greater proportions of nidal vasculature (36.0% vs. 31.2%; P = 0.152). CONCLUSIONS: The automated segmentation algorithm was able to achieve classification of the AVM nidus components with relative accuracy. Greater proportions of intervening nidal brain parenchyma were associated with RICs.