Organs at Risk Delineation for Brain Tumor Radiation Planning in Patients with Glioblastoma Using Deep Learning

Abstract

Our standalone software, using deep learning technology, performs fully automatic organs at risk (OAR) segmentation within the brain in 2 minutes for brain radiation planning in patients with glioblastoma (GBM). The aim of the present study is to evaluate, if fully automated segmentation of the OAR within the brain in GBM patients is non-inferior to human rater segmentation for radiation planning. We included post-OP T1w, T1c, T2w and FLAIR MRI datasets of 23 patients with newly diagnosed, surgically resected and histologically confirmed GBM. Manual segmentations of the brain OAR (brainstem, cochleae, eyes, hippocampus, lacrimal glands, lenses, optic chiasm, optic nerves, pituitary gland and retinae) according to Scoccianti et al. (Scoccianti 2015) were performed by 3 experienced radiation oncologists on the T1c MRI sequence. For fully automatic segmentation, we developed an OAR segmentation method. The MRI data was co-registered for automatic segmentation. Manual delineations were utilized for training the deep learning system based on the U-Net architecture. We evaluated the segmentation method on the results of a 6-fold cross-validation considering the imaging data of all patients in terms of dice-coefficient (DC) (volumetric overlap) and volume similarity (VS) using the Kruskal-Wallis test. Median DC and interquartile range (IQR) of the different pairings of expert raters (reported as tuples) were (0.74, 0.19), (0.67, 0.28) and (0.72, 0.20) for all OAR structures combined. The results of the automatic segmentation compared to the three different raters were (0.63, 0.39), (0.65, 0.36) and (0.63, 0.41). Large distribution of DC and VS were observed for small OAR like the left cochlea between the raters [(0.78, 0.12), (0.43, 0.13), (0.49, 0.12); Kruskal-Wallis test: chi-square=43.80, p=0.001)] and the automatic method: [(0.00, 0.23), (0.00, 0.26), (0.00, 0.18; Kruskal-Wallis test: chi-square=81.29, p=0.001)]. The cochlea was missed completely by the automatic method in 9 patients. For larger OAR, such as the brainstem, no statistical difference was detected between measured volumes among raters (Kruskal-Wallis test: chi-square=4.6, p=0.098) [(0.91, 0.02), (0.91, 0.02), (0.89, 0.02)], and compared to the automatic segmentation: [(0.91, 0.02), (0.91, 0.01), (0.90, 0.02)]. The performance of the proposed system is highly dependent on the raters agreement. The higher the agreement among raters, the better predicts the prototype the OAR within the brain. With this proof of concept study, we generate first promising results, and plan next steps to improve results.