Preliminary assessment of semiautomated segmentation tool for GBM neurosurgery: a BraTS-Toolkit approach

Introduction The current standard of care for glioblastoma (GBM) is surgical resection, followed by adjuvant radiotherapy. Complete resection remains challenging because of unclearly defined tumor margins. Presurgery tumor segmentation is generally done by a neurosurgeon on the contrast-enhanced T1 (T1c) scan. A semi-automated tumor segmentation can ease this process, exploiting also information contained in other scans besides T1c. The aim of this study is to propose a settlement to qualitatively assess segmentations on a multicentric GBM cohort. The Brats-Toolkit [1] tool was considered. Methods Preoperative MRI studies of 72 GBM patients (41 males, 31 females, mean age 63 yr.) surgically treated at Fondazione IRCCS San Gerardo were considered. Images were acquired with heterogeneous protocols, all containing T1, T1c, T2 and T2 FLAIR scans. T1c images were manually segmented by a neurosurgeon, as to clinical practice (unique Region of Interest containing core + necrosis + core enhanced). BraTSPreprocessor, BraTS-Segmentor (8 Docker images from the repository), and BraTS-Fusionator (SIMPLE method) were applied, providing as output core + necrosis, core-enhanced and edema contours [2], [3]. Manual and BraTS segmentations were visually evaluated by a senior neuroradiologist. Results/Discussion MRI studies were classified in 6 groups based on scans characteristics (Group1: 3T - isotropic T1, T1c, FLAIR - anisotropic T2; Group2: 1.5T - isotropic T1, T1c, FLAIR - anisotropic T2; Group3: 3T - isotropic T1c, FLAIR - anisotropic T1, T2; Group4 1.5T - isotropic T1c, FLAIR - anisotropic T1, T2; Group5: 1.5T - isotropic T1, T1c - anisotropic FLAIR, T2; Group6: 1.5T - isotropic T1c - anisotropic T1, FLAIR, T2). A 5-point scale was defined: 1 (correctly overlapped), 2 (not-perfectly overlapped but including all pathologic tissue), 3 (miss of pathologic tissue inclusion, not compromising overall tumor core segmentation), 4 (miss of pathologic tissue inclusion resulting in substantial of tumor core miss), 5 (misclassification of normal tissue as tumor core or of edema as tumor core). The 5-point scale is independently applied to: core + necrosis, coreenhanced, core + necrosis + core enhanced, edema. Preliminary application of the proposed assessment to 15 patients gave results in Fig. 1. Conclusion A protocol to qualitatively assess GBM segmentations on multicentric MRIs was proposed and applied to 15 patients, showing a Brats-Toolkit GBM contour quality generally comparable with neurosurgeons (Fig. 1, Fig 2). After evaluations on the remaining patients are finished, we will be able to drive more robust conclusions. A further evaluation refinement is envisioned, in which patients will be further classified based on tumor characteristics. Novelty A protocol for GBM segmentation assessment, taking into account protocol heterogeneity and tumor subregions, is proposed and preliminarily assessed. Impact The results of a robust evaluation performed according to this protocol will allow to tune semi-automatic strategies usage in clinical practice. Disclosure I or one of my co-authors have no financial interest or relationship to disclose regarding the subject matter of this presentation. References [1] F. Kofler et al., “BraTS Toolkit: Translating BraTS Brain Tumor Segmentation Algorithms Into Clinical and Scientific Practice,” Front. Neurosci., vol. 14, p. 125, Apr. 2020, doi: 10.3389/FNINS.2020.00125/BIBTEX [2] B. H. Menze et al., “The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS),” IEEE Trans. Med. Imaging, vol. 34, no. 10, p. 1993, Oct. 2015, doi: 10.1109/TMI.2014.2377694 [3] Bakas S. et al, "Identifying the Best Machine Learning Algorithms for Brain Tumor Segmentation, Progression Assessment, and Overall Survival Prediction in the BRATS Challenge, 2019, https://doi.org/10.48550/arXiv.1811.02629