Laukamp Kai Roman, Pennig Lenhard, Thiele Frank, Reimer Robert, Görtz Lukas, Shakirin Georgy, Zopfs David, Timmer Marco, Perkuhn Michael, Borggrefe Jan
Institute for Diagnostic and Interventional Radiology, Faculty of Medicine and University Hospital Cologne, University of Cologne, Kerpener Straße 62, 50937, Cologne, Germany.
Department of Radiology, University Hospitals Cleveland Medical Center, Cleveland, OH, USA.
Clin Neuroradiol. 2021 Jun;31(2):357-366. doi: 10.1007/s00062-020-00884-4. Epub 2020 Feb 14.
Volumetric assessment of meningiomas represents a valuable tool for treatment planning and evaluation of tumor growth as it enables a more precise assessment of tumor size than conventional diameter methods. This study established a dedicated meningioma deep learning model based on routine magnetic resonance imaging (MRI) data and evaluated its performance for automated tumor segmentation.
The MRI datasets included T1-weighted/T2-weighted, T1-weighted contrast-enhanced (T1CE) and FLAIR of 126 patients with intracranial meningiomas (grade I: 97, grade II: 29). For automated segmentation, an established deep learning model architecture (3D deep convolutional neural network, DeepMedic, BioMedIA) operating on all four MR sequences was used. Segmentation included the following two components: (i) contrast-enhancing tumor volume in T1CE and (ii) total lesion volume (union of lesion volume in T1CE and FLAIR, including solid tumor parts and surrounding edema). Preprocessing of imaging data included registration, skull stripping, resampling, and normalization. After training of the deep learning model using manual segmentations by 2 independent readers from 70 patients (training group), the algorithm was evaluated on 56 patients (validation group) by comparing automated to ground truth manual segmentations, which were performed by 2 experienced readers in consensus.
Of the 56 meningiomas in the validation group 55 were detected by the deep learning model. In these patients the comparison of the deep learning model and manual segmentations revealed average dice coefficients of 0.91 ± 0.08 for contrast-enhancing tumor volume and 0.82 ± 0.12 for total lesion volume. In the training group, interreader variabilities of the 2 manual readers were 0.92 ± 0.07 for contrast-enhancing tumor and 0.88 ± 0.05 for total lesion volume.
Deep learning-based automated segmentation yielded high segmentation accuracy, comparable to manual interreader variability.
脑膜瘤的体积评估是治疗规划和肿瘤生长评估的重要工具，因为与传统直径测量方法相比，它能更精确地评估肿瘤大小。本研究基于常规磁共振成像（MRI）数据建立了专用的脑膜瘤深度学习模型，并评估其自动肿瘤分割性能。
MRI数据集包括126例颅内脑膜瘤患者（I级：97例，II级：29例）的T1加权/T2加权、T1加权增强（T1CE）和液体衰减反转恢复序列（FLAIR）。对于自动分割，使用了基于所有四个MR序列运行的既定深度学习模型架构（3D深度卷积神经网络，DeepMedic，BioMedIA）。分割包括以下两个部分：（i）T1CE序列中增强对比的肿瘤体积，以及（ii）总病变体积（T1CE序列和FLAIR序列中病变体积的总和，包括实体肿瘤部分和周围水肿）。成像数据的预处理包括配准、颅骨剥离、重采样和归一化。在使用70例患者（训练组）的2名独立阅片者的手动分割对深度学习模型进行训练后，通过将自动分割结果与由2名经验丰富的阅片者一致完成的真实手动分割结果进行比较，对56例患者（验证组）的算法进行评估。
在验证组的56例脑膜瘤中，深度学习模型检测出55例。在这些患者中，深度学习模型与手动分割结果的比较显示，增强对比的肿瘤体积的平均骰子系数为0.91±0.08，总病变体积的平均骰子系数为0.82±0.12。在训练组中，2名手动阅片者的阅片者间变异性对于增强对比的肿瘤为0.92±0.07，对于总病变体积为0.88±0.05。
基于深度学习的自动分割产生了较高的分割精度，与手动阅片者间变异性相当。
