TY - JOUR
T1 - Robust deep learning–based segmentation of glioblastoma on routine clinical MRI scans using sparsified training
AU - Eijgelaar, Roelant S.
AU - Visser, Martin
AU - Müller, Domenique M.J.
AU - Barkhof, Frederik
AU - Vrenken, Hugo
AU - van Herk, Marcel
AU - Bello, Lorenzo
AU - Nibali, Marco Conti
AU - Rossi, Marco
AU - Sciortino, Tommaso
AU - Berger, Mitchel S.
AU - Hervey-Jumper, Shawn
AU - Kiesel, Barbara
AU - Widhalm, Georg
AU - Furtner, Julia
AU - Robe, Pierre A.J.T.
AU - Mandonnet, Emmanuel
AU - Hamer, Philip C.De Witt
AU - de Munck, Jan C.
AU - Witte, Marnix G.
N1 - Publisher Copyright:
© 2020, Radiological Society of North America Inc.. All rights reserved.
PY - 2020/9
Y1 - 2020/9
N2 - 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 re-covery, and postcontrast T1-weighted MRI from 117 patients (median age, 64 years; interquartile range [IQR], 55–73 years; 76 men) included within the Multimodal Brain Tumor Image Segmentation (BraTS) dataset plus a clinical dataset (2012–2013) with similar imaging modalities of 634 patients (median age, 59 years; IQR, 49–69 years; 382 men) with glioblastoma from six hospitals were used. Expert tumor delineations on the postcontrast images were available, but for various clinical datasets, one or more sequences were miss-ing. The convolutional neural network, DeepMedic, was trained on combinations of complete and incomplete data with and without site-specific data. Sparsified training was introduced, which randomly simulated missing sequences during training. The effects of spar-sified training and center-specific training were tested using Wilcoxon signed rank tests for paired measurements. Results: A model trained exclusively on BraTS data reached a median Dice score of 0.81 for segmentation on BraTS test data but only 0.49 on the clinical data. Sparsified training improved performance (adjusted P, .05), even when excluding test data with missing sequences, to median Dice score of 0.67. Inclusion of site-specific data during sparsified training led to higher model performance Dice scores greater than 0.8, on par with a model based on all complete and incomplete data. For the model using BraTS and clinical training data, inclusion of site-specific data or sparsified training was of no consequence. Conclusion: Accurate and automatic segmentation of glioblastoma on clinical scans is feasible using a model based on large, heteroge-neous, and partially incomplete datasets. Sparsified training may boost the performance of a smaller model based on public and site-specific data.
AB - 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 re-covery, and postcontrast T1-weighted MRI from 117 patients (median age, 64 years; interquartile range [IQR], 55–73 years; 76 men) included within the Multimodal Brain Tumor Image Segmentation (BraTS) dataset plus a clinical dataset (2012–2013) with similar imaging modalities of 634 patients (median age, 59 years; IQR, 49–69 years; 382 men) with glioblastoma from six hospitals were used. Expert tumor delineations on the postcontrast images were available, but for various clinical datasets, one or more sequences were miss-ing. The convolutional neural network, DeepMedic, was trained on combinations of complete and incomplete data with and without site-specific data. Sparsified training was introduced, which randomly simulated missing sequences during training. The effects of spar-sified training and center-specific training were tested using Wilcoxon signed rank tests for paired measurements. Results: A model trained exclusively on BraTS data reached a median Dice score of 0.81 for segmentation on BraTS test data but only 0.49 on the clinical data. Sparsified training improved performance (adjusted P, .05), even when excluding test data with missing sequences, to median Dice score of 0.67. Inclusion of site-specific data during sparsified training led to higher model performance Dice scores greater than 0.8, on par with a model based on all complete and incomplete data. For the model using BraTS and clinical training data, inclusion of site-specific data or sparsified training was of no consequence. Conclusion: Accurate and automatic segmentation of glioblastoma on clinical scans is feasible using a model based on large, heteroge-neous, and partially incomplete datasets. Sparsified training may boost the performance of a smaller model based on public and site-specific data.
UR - http://www.scopus.com/inward/record.url?scp=85101450837&partnerID=8YFLogxK
U2 - 10.1148/ryai.2020190103
DO - 10.1148/ryai.2020190103
M3 - Article
C2 - 33937837
AN - SCOPUS:85101450837
VL - 2
SP - 1
EP - 9
JO - Radiology: Artificial Intelligence
JF - Radiology: Artificial Intelligence
IS - 5
M1 - e190103
ER -