P01.088 Brain resection cavity delineation for radiation target volume definition in glioblastoma patients using deep learning. (19th September 2018)

• Record Type: Journal Article
• Authors: Herrmann, E; Ermis, E; Jungo, A; Blatti-Moreno, M; Knecht, U; Aebersold, D M; Manser, P; Reyes, M
• Is Part Of: Neuro-oncology. Volume 20(2018)Supplement 3
• Page Start: iii250
• Page End: iii251
• Publication Date: 2018-09-19
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noy139.130
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: Background: We present results of an automated segmentation method for resection cavity in Glioblastoma (GBM) patients using deep learning technologies. The approach fully segments the resection cavity and yields estimates of uncertainty. We compare the approach to a previously reported machine learning method and show the superiority in performance of the deep learning approach while yielding estimates of segmentation uncertainty, which can be used by an operator to monitor computer's results. Material and Methods: We included post-OP, pre- and post-contrast T1, T2 and FLAIR MRI datasets of 30 patients with newly diagnosed, surgically resected and histologically confirmed GBM. Radiation Cavity (RC) were defined as surgical cavity plus blood products. Three radiation oncologists manually delineated the RC using all four MRI sequences. For fully automatic segmentation, we developed a deep learning cavity segmentation method, which was trained on a subset of 25 cases, and utilizes all four MRI sequences to learn RC delineations. 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 6-fold cross-validation using the imaging data. Results: 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 (0.83, 0.14), (0.81, 0.12), and (0.81, 0.13). 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=1.46, p=0.69). 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). Uncertainty maps generated by the deep learning approach were visually compared against the segmentation errors, and a good spatial overlap was found between uncertainty and RC delineation errors. This allow us to conclude on the reliability and good calibration properties of the deep learning approach. Conclusion: The proposed deep learning approach 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 but outperform previously presented approaches, while providing estimates of segmentation uncertainty, which can leverage the interactions between these systems and human experts.