Intra-observer Variability in Segmentation of Glioblastomas
MetadataShow full item record
Background: Glioblastomas are primary malignant brain tumors with poor prognosis. In study settings where tumor growth and treatment responses are assessed, tumor volumes are measured. To obtain valid volume measures there is a need for accurate and reproducible volume segmentations. In this study, we wanted to investigate the reproducibility of segmentation on magnetic resonance imaging (MRI) images with different resolution, and using different software. Material and methods: MRI images from consecutive patients were retrospectively collected, 10 with low image-resolution and 20 with high resolution. Two software solutions were used for repeated semi-automatic segmentation of glioblastomas, BrainVoyager QX, and the plugin SX software for OsiriX. BrainVoyager was used for segmentation of both the 10 images with thick slices and the 20 images with thin slices, while OsiriX was used for segmentation of images with thin slices only. Reproducibility was assessed with measures of agreement, reliability and the Dice similarity coefficient (DSC), which is a measure of spatial overlap. In addition, the time used for each segmentation was recorded. Results: For the segmentation of the low resolution images in BrainVoyager, the repeatability coefficient was 2.5 ml, the DSC was 0.984, and the mean time used was 6.5 ± 3.3 minutes. For segmentation of images with thin slices in BrainVoyager the repeatability coefficient was 2.4 ml, the DSC was 0.971, and the mean time used was 46.3 ± 21.4 minutes. This segmentation time was significantly longer than the 14.3 ± 5.5 minutes used for segmentation of the same tumors in OsiriX. There was a significant difference of-0.6 ml between repeated segmentations in OsiriX, and a repeatability coefficient could not be calculated. The DSC between repeated segmentations in OsiriX was 0.938. Conclusion: In this study we found that the BrainVoyager software can be used for reproducible segmentation of brain tumor volumes. The significant time saved with segmentations in OsiriX indicates that it is possible to make volumetric segmentation more feasible in a clinical setting. Unfortunately, the significant volume difference found using the SX software plugin in OsiriX means that this plugin cannot be recommended until further studies of the reproducibility have been performed.