Ziebart Andreas, Stadniczuk Denis, Roos Veronika, Ratliff Miriam, von Deimling Andreas, Hänggi Daniel, Enders Frederik
Department of Neurosurgery, University Hospital Mannheim, Medical Faculty Mannheim, University of Heidelberg, Mannheim, Germany.
Department of Software Engineering, Clevertech Inc., New York, NY, United States.
Front Oncol. 2021 May 11;11:668273. doi: 10.3389/fonc.2021.668273. eCollection 2021.
Reliable on site classification of resected tumor specimens remains a challenge. Implementation of high-resolution confocal laser endoscopic techniques (CLEs) during fluorescence-guided brain tumor surgery is a new tool for intraoperative tumor tissue visualization. To overcome observer dependent errors, we aimed to predict tumor type by applying a deep learning model to image data obtained by CLE.
Human brain tumor specimens from 25 patients with brain metastasis, glioblastoma, and meningioma were evaluated within this study. In addition to routine histopathological analysis, tissue samples were stained with fluorescein and analyzed with CLE. We trained two convolutional neural networks and built a predictive level for the outputs.
Multiple CLE images were obtained from each specimen with a total number of 13,972 fluorescein based images. Test accuracy of 90.9% was achieved after applying a two-class prediction for glioblastomas and brain metastases with an area under the curve (AUC) value of 0.92. For three class predictions, our model achieved a ratio of correct predicted label of 85.8% in the test set, which was confirmed with five-fold cross validation, without definition of confidence. Applying a confidence rate of 0.999 increased the prediction accuracy to 98.6% when images with substantial artifacts were excluded before the analysis. 36.3% of total images met the output criteria.
We trained a residual network model that allows automated, on site analysis of resected tumor specimens based on CLE image datasets. Further studies are required to assess the clinical benefit CLE can have.
对切除的肿瘤标本进行可靠的现场分类仍然是一项挑战。在荧光引导的脑肿瘤手术中实施高分辨率共聚焦激光内镜技术（CLE）是术中肿瘤组织可视化的一种新工具。为了克服观察者依赖的误差，我们旨在通过将深度学习模型应用于CLE获得的图像数据来预测肿瘤类型。
本研究评估了25例脑转移瘤、胶质母细胞瘤和脑膜瘤患者的人脑肿瘤标本。除了常规组织病理学分析外，组织样本用荧光素染色并用CLE进行分析。我们训练了两个卷积神经网络并为输出建立了预测水平。
从每个标本中获得了多个CLE图像，总共13972张基于荧光素的图像。对胶质母细胞瘤和脑转移瘤进行两类预测后，测试准确率达到90.9%，曲线下面积（AUC）值为0.92。对于三类预测，我们的模型在测试集中实现了85.8%的正确预测标签率，这通过五折交叉验证得到证实，无需定义置信度。在分析前排除有大量伪影的图像时，应用0.999的置信率可将预测准确率提高到98.6%。总图像的36.3%符合输出标准。
我们训练了一个残差网络模型，该模型允许基于CLE图像数据集对切除的肿瘤标本进行自动现场分析。需要进一步的研究来评估CLE可能具有的临床益处。
