Robust Deep Learning–based Segmentation of Glioblastoma on Routine Clinical MRI Scans Using Sparsified Training

• DOI: 10.1148/ryai.2020190103
• Publication Date: 2020-09-30T13:56:37Z
• Authors: Roelant S. Eijgelaar; Martin Visser; Domenique M. J. M...; Frederik Barkhof; Hugo Vrenken; Marcel van Herk; Lorenzo Bello; Marco Conti Nibali; Marco Rossi; Tommaso Sciortino; Mitchel S. Berger; Shawn Hervey-Jumper; Barbara Kiesel; Georg Widhalm; Julia Furtner; Pierre A. J. T. Robe; Emmanuel Mandonnet; Philip C. De Witt...; Jan C. de Munck; Marnix G. Witte
• Keywords: 03 medical and health sciences; 0302 clinical medicine; Radiological and Ultrasound Technology; Radiology Nuclear Medicine and imaging; Artificial Intelligence; Original Research; 3. Good health

Purpose To improve the robustness of deep learning–based glioblastoma segmentation in a clinical setting with sparsified datasets. Materials and Methods In this retrospective study, preoperative T1-weighted, T2-weighted, T2-weighted fluid-attenuated inversion recovery, postcontrast T1-weighted MRI from 117 patients (median age, 64 years; interquartile range [IQR], 55–73 76 men) included within Multimodal Brain Tumor Image Segmentation (BraTS) dataset plus (2012–2013) similar imaging modalities 634 59 IQR, 49–69 382 six hospitals were used. Expert tumor delineations on images available, but for various datasets, one or more sequences missing. The convolutional neural network, DeepMedic, was trained combinations complete incomplete data without site-specific data. Sparsified training introduced, which randomly simulated missing during training. effects center-specific tested using Wilcoxon signed rank tests paired measurements. Results A model exclusively BraTS reached median Dice score 0.81 test only 0.49 improved performance (adjusted P < .05), even when excluding sequences, to 0.67. Inclusion led higher scores greater than 0.8, par based all For data, inclusion no consequence. Conclusion Accurate automatic scans is feasible large, heterogeneous, partially may boost smaller public Supplemental material available article. Keywords: Adults, Brain/Brain Stem, CNS, MR-Imaging, Oncology, Segmentation, Technology Assessment Published under CC BY 4.0 license.