The validity of segmentation is an important issue in image processing because it has a direct impact on surgical planning. Binary manual segmentation is not only time-consuming but also lacks the ability of differentiating subtle intensity variations among voxels, particularly for those on the border of a tumor and for different tumor types. Previously we have developed an automated segmentation method that yields voxel-wise continuous probabilistic measures, indicating a level of tumor presence. The goal of this work is to examine three accuracy metrics based on two-sample statistical methods, against the estimated composite latent ground truth derived from several experts' manual segmentation by a maximum likelihood algorithm. We estimated the distribution functions of the tumor and control voxel data parametrically by assuming a mixture of two beta distributions with different shape parameters. We derived the resulting receiver operating characteristic curves, Dice similarity coefficients, and mutual information, over all possible decision thresholds. Based on each validation metric, an optimal threshold was then computed via maximization. We illustrated these methods on MR imaging data from nine brain tumor cases, three with meningiomas, three astrocytomas, and three other low-grade gliomas. The automated segmentation yielded satisfactory accuracy, with varied optimal thresholds.
展开▼
