Iterative Probabilistic Voxel Labeling: Automated Segmentation for Analysis of The Cancer Imaging Archive Glioblastoma Images

BACKGROUND AND PURPOSE: Robust, automated segmentation algorithms are required for quantitative analysis of large imaging datasets. We developed an automated method that identifies and labels brain tumor-associated pathology by using an iterative probabilistic voxel labeling using k-nearest neighbor and Gaussian mixture model classification. Our purpose was to develop a segmentation method which could be applied to a variety of imaging from The Cancer Imaging Archive. MATERIALS AND METHODS: Images from 2 sets of 15 randomly selected subjects with glioblastoma from The Cancer Imaging Archive were processed by using the automated algorithm. The algorithm-defined tumor volumes were compared with those segmented by trained operators by using the Dice similarity coefficient. RESULTS: Compared with operator volumes, algorithm-generated segmentations yielded mean Dice similarities of 0.92 0.03 for contrast-enhancing volumes and 0.84 +/- 0.09 for FLAIR hyperintensity volumes. These values compared favorably with the means of Dice similarity coefficients between the operator-defined segmentations: 0.92 +/- 0.03 for contrast-enhancing volumes and 0.92 +/- 0.05 for FLAIR hyperintensity volumes. Robust segmentations can be achieved when only postcontrast T1WI and FLAIR images are available. CONCLUSIONS: Iterative probabilistic voxel labeling defined tumor volumes that were highly consistent with operator-defined volumes. Application of this algorithm could facilitate quantitative assessment of neuroimaging from patients with glioblastoma for both research and clinical indications.