Comparing Image Segmentation Techniques for Determining 3D Orbital Cavernous Hemangioma Size on MRI.

PURPOSE: To measure orbital cavernous hemangioma size using 3 segmentation methods requiring different degrees of subjective judgment, and to evaluate interobserver agreement using these methods. METHODS: Fourteen patients with orbital cavernous hemangiomas were included in the study. Pretreatment T2-weighted MRIs were analyzed by 2 observers using 3 methods, including 1 user-dependent image segmentation method that required high degrees of subjective judgment (ellipsoid) and 2 parameter-dependent methods that required low degree of subjective judgment (GrowCut and k-means clustering segmentation). Interobserver agreement was assessed using Lin's concordance correlation coefficients. RESULTS: Using the ellipsoid method, the average tumor sizes calculated by the 2 observers were 1.68 ml (standard deviation [SD] 1.45 ml) and 1.48 ml (SD 1.19 ml). Using the GrowCut method, the average tumor sizes calculated by the 2 observers were 3.00 ml (SD 2.46 ml) and 6.34 ml (SD 3.78 ml). Using k-means clustering segmentation, the average tumor sizes calculated by the 2 observers were 2.31 ml (SD 1.83 ml) and 2.12 ml (SD 1.87 ml). The concordance correlation coefficient for the ellipsoid, GrowCut, and k-means clustering methods were 0.92 (95% CI, 0.83-0.99), 0.12 (95% CI, -0.21 to 0.44), and 0.95 (95% CI, 0.90-0.99), respectively. CONCLUSIONS: k-means clustering, a parameter-dependent method with low degree of subjective judgment, showed better interobserver agreement compared with the ellipsoid and GrowCut methods. k-means clustering clearly delineated tumor boundaries and outlined components of the tumor with different signal intensities.