Fully Automated Segmentation of the Brain Resection Cavity for Radiation Target Volume Definition in Glioblastoma Patients

Automated brain tumor segmentation methods are computational algorithms that yield a 3-dimensional delineation of tumor compartments from multimodal magnetic resonance imaging (MRI). Our standalone software, using machine-learning technology, performs fully automatic Glioblastoma (GBM) segmentation (necrosis, edema, contrast-enhanced, and non-contrast enhanced tumor) in 5 minutes. The software was evaluated clinically for pre- and post- operative (OP) tumor definition (machine vs. human rater) in GBM, without determining the resection cavity (RC). The aim of the present study is to evaluate if fully automated segmentation of the RC in GBM patients is non-inferior to human rater segmentation for radiation target volume definition. Post-OP, pre-contrast, and post-contrast T1, T2, and FLAIR MRI datasets of 30 patients with newly diagnosed, surgically resected and histologically confirmed GBM were included. RCs were defined as surgical cavity plus blood products. Three radiation oncologists manually delineated the RC using all four MRI sets. For fully automatic segmentation, we developed a cavity segmentation method. The MRI data was skull-stripped and co-registered for automatic segmentation. Manual delineations were utilized for training the machine learning system. Additionally, a manual correction step was performed with naive brushing for the most obvious errors. We evaluated the segmentation method in terms of dice-coefficient (DC) (volumetric overlap) and estimated volume measurements using the Kruskal-Wallis test followed by a 3-fold cross-validation using the imaging data. Median DC and interquartile range (IQR) of the different pairings of expert raters (reported as tuples) were (0.85, 0.08), (0.84, 0.07), and (0.86, 0.07). The results of the automatic segmentation compared to the three different raters were (0.76, 0.23), (0.74, 0.22), and (0.73, 0.19). After the manual correction step, the results improved to (0.83, 0.1), (0.8, 0.1), and (0.81, 0.09). We did not detect a statistically significant difference regarding the distribution of the measured volumes for the different raters and methods (Kruskal-Wallis test: chi-square=1.79, p=0.77). In two out of the 30 cases, the RC was completely missed by the automatic method. The main sources of error consisted in over- or under-segmentation of the RC due to signal inhomogeneity (especially in T2 and FLAIR sequences) and over-segmentation due to similar intensity patterns (edema, subarachnoid space, or ventricles). The current prototype yields good results for the segmentation of the RC. However, blood products and air can affect its performance. Compared to human experts, the computer-generated results are still subpar. Nonetheless, with this proof of concept study, we generate first promising results, and plan next directions to improve results.