Eijgelaar Roelant S, Visser Martin, Müller Domenique M J, Barkhof Frederik, Vrenken Hugo, van Herk Marcel, Bello Lorenzo, Conti Nibali Marco, Rossi Marco, Sciortino Tommaso, Berger Mitchel S, Hervey-Jumper Shawn, Kiesel Barbara, Widhalm Georg, Furtner Julia, Robe Pierre A J T, Mandonnet Emmanuel, De Witt Hamer Philip C, de Munck Jan C, Witte Marnix G
Department of Radiation Oncology, The Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX Amsterdam, the Netherlands (R.S.E., M.v.H., M.G.W.); Department of Radiology and Nuclear Medicine, Amsterdam UMC, Location Vrije Universiteit Amsterdam, Amsterdam, the Netherlands (M.V., F.B., H.V., J.C.d.M.); Neurosurgical Center Amsterdam, Amsterdam UMC, Location Vrije Universiteit Amsterdam, Amsterdam, the Netherlands (D.M.J.M., P.C.D.W.H.); Institutes of Neurology & Healthcare Engineering, University College London, London, England (F.B.); Faculty of Biology, Medicine & Health, Division of Cancer Sciences, University of Manchester and Christie NHS Trust, Manchester, England (M.v.H.); Neurosurgical Oncology Unit, Department of Oncology and Hemato-Oncology, Università degli Studi di Milano, Humanitas Research Hospital, IRCCS, Milan, Italy (L.B., M.C.N., M.R., T.S.); Department of Neurologic Surgery, University of California-San Francisco, San Francisco, Calif (M.S.B., S.H.J.); Department of Neurosurgery, Medical University Vienna, Vienna, Austria (B.K., G.W.); Department of Biomedical Imaging and Image-guided Therapy, Medical University Vienna, Vienna, Austria (J.F.); Department of Neurology & Neurosurgery, University Medical Center Utrecht, Utrecht, the Netherlands (P.A.J.T.R.); and Department of Neurologic Surgery, Hôpital Lariboisière, Paris, France (E.M.).
Radiol Artif Intell. 2020 Sep 30;2(5):e190103. doi: 10.1148/ryai.2020190103. eCollection 2020 Sep.
To improve the robustness of deep learning-based glioblastoma segmentation in a clinical setting with sparsified datasets.
In this retrospective study, preoperative T1-weighted, T2-weighted, T2-weighted fluid-attenuated inversion recovery, 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 missing. 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 sparsified training and center-specific training were tested using Wilcoxon signed rank tests for paired measurements.
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 < .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.
Accurate and automatic segmentation of glioblastoma on clinical scans is feasible using a model based on large, heterogeneous, and partially incomplete datasets. Sparsified training may boost the performance of a smaller model based on public and site-specific data.Published under a CC BY 4.0 license.
在临床环境中利用稀疏数据集提高基于深度学习的胶质母细胞瘤分割的稳健性。
在这项回顾性研究中，使用了多模态脑肿瘤图像分割（BraTS）数据集中的117例患者（中位年龄64岁；四分位间距[IQR]，55 - 73岁；76名男性）的术前T1加权、T2加权、T2加权液体衰减反转恢复序列以及增强后T1加权磁共振成像（MRI），再加上来自六家医院的634例胶质母细胞瘤患者（中位年龄59岁；IQR，49 - 69岁；382名男性）的具有相似成像模态的临床数据集（2012 - 2013年）。有增强后图像上的专家肿瘤轮廓标注，但对于各种临床数据集，有一个或多个序列缺失。卷积神经网络DeepMedic在有和没有特定部位数据的完整和不完整数据组合上进行训练。引入了稀疏训练，即在训练期间随机模拟缺失序列。使用威尔科克森符号秩检验对配对测量结果来测试稀疏训练和特定中心训练的效果。
仅在BraTS数据上训练的模型在BraTS测试数据上进行分割时，中位骰子系数得分为0.81，但在临床数据上仅为0.49。即使排除有缺失序列的测试数据，稀疏训练也提高了性能（校正P <.05），中位骰子系数得分达到0.67。在稀疏训练期间纳入特定部位数据导致更高的模型性能，骰子系数得分大于0.8，与基于所有完整和不完整数据的模型相当。对于使用BraTS和临床训练数据的模型，纳入特定部位数据或稀疏训练没有影响。
使用基于大型、异质性和部分不完整数据集的模型，在临床扫描上对胶质母细胞瘤进行准确和自动分割是可行的。稀疏训练可能会提高基于公共和特定部位数据的较小模型的性能。根据知识共享署名4.0许可发布。
