AB038. Establishing a reliable semi-automated segmentation method for assessing chemoradiotherapy effects on the non-tumoral brain

Background: Current voxel-based morphometry (VBM) studies of chemoradiotherapy effects on healthy tissues of the glioblastoma multiforme (GBM) brain face a challenge with neuroanatomical distortions (tumor, tumor edema, and resection cavities) and limited comparisons can be drawn across studies due to lack of a universally accepted software package. Our aim is to compare current semi-automated segmentation methods and optimize them for reliability in investigating the effects of chemoradiotherapy on GBM patients.

Methods: A publicly available dataset was used based on predefined inclusion and exclusion criteria. VBM pipelines CAT12 and FSL were tested and optimized to reduce the impact of neuroanatomical distortions. T1-weighted images were screened, and post-processed with FSL and CAT12. Gray matter (GM) and white matter (WM), and cerebrospinal fluid (CSF) volumes of whole brain, tumour-containing and non-tumor containing hemispheres, pre- and post-chemoradiotherapy were calculated and analyzed with Wilcoxon signed-rank tests. Agreement and consistency between FSL and CAT12 were assessed using Bland-Altman plots and intraclass correlation coefficients (ICCs).

Results: Post-chemoradiotherapy GM volumes were significantly reduced in whole brain with a compensatory significant increase in CSF volumes, while WM volumes had no significant changes. Similar trends were noted in tumor-containing and non-tumor-containing hemispheres. Bland-Altman plots showed good agreement between FSL and CAT12 processed GM and WM volumes of whole brain, tumor-containing, and non-tumor-containing hemispheres. ICC ≥0.70 was observed in GM [0.70 (0.53–0.82)] and WM [0.75 (0.60–0.85)] volumes of non-tumor-containing hemisphere, and WM [0.71 (0.55–0.83)] volumes of whole brain. GM volumes of tumor-containing hemisphere had good agreement but surprisingly, poor consistency [0.50 (0.25–0.68)]. CSF volumes in non-tumor-containing hemisphere had better agreement and consistency [0.55 (0.32–0.71)] than whole brain [0.49 (0.25–0.67)] and tumor-containing hemisphere CSF [0.36 (0.10–0.58)] volumes. Visual inspection revealed both CAT12 and FSL mis-segmented in the presence of neuroanatomical distortion although CAT12 was more susceptible in the presence of a hematoma.

Conclusions: VBM studies of chemoradiotherapy effects on the brain post-tumor resection remain challenging due to neuroanatomical distortions. A reliable alternative is to use non-tumor-containing hemispheres with no anatomical distortion. Should tumor-containing brains be used, FSL is a more suitable choice, especially in the presence of hematoma distortion.

Keywords: Voxel-based morphometry (VBM); glioblastoma; chemoradiotherapy