Robust Chest CT Image Segmentation of COVID-19 Lung Infection based on limited data

<strong>Background:</strong> The coronavirus disease 2019 (COVID-19) affects billions of lives around the world and has a significant impact on public healthcare. For quantitative assessment and disease monitoring medical imaging like computed tomography offers great potential as alternative to RT-PCR methods. For this reason, automated image segmentation is highly desired as clinical decision support. However, publicly available COVID-19 imaging data is limited which leads to overfitting of traditional approaches. <strong>Methods:</strong> To address this problem, we propose an innovative automated segmentation pipeline for COVID-19 infected regions, which is able to handle small datasets by utilization as variant databases. Our method focuses on on-the-fly generation of unique and random image patches for training by performing several preprocessing methods and exploiting extensive data augmentation. For further reduction of the overfitting risk, we implemented a standard 3D U-Net architecture instead of new or computational complex neural network architectures. <strong>Results:</strong> Through a k-fold cross-validation on 20 CT scans as training and validation of COVID-19, we were able to develop a highly accurate as well as robust segmentation model for lungs and COVID-19 infected regions without overfitting on limited data. We performed an in-detail analysis and discussion on the robustness of our pipeline through a sensitivity analysis based on the cross-validation and impact on model generalizability of applied preprocessing techniques. Our method achieved Dice similarity coefficients for COVID-19 infection between predicted and annotated segmentation from radiologists of 0.804 on validation and 0.661 on a separate testing set consisting of 100 patients. <strong>Conclusions:</strong> We demonstrated that the proposed method outperforms related approaches, advances the state-of-the-art for COVID-19 segmentation and improves robust medical image analysis based on limited data. The code and model are available under the following link:<br> https://github.com/frankkramer-lab/covid19.MIScnn背景：2019冠状病毒病（COVID-19）已波及全球数十亿人口，对公共卫生体系造成重大冲击。若要实现量化评估与疾病监测，计算机断层扫描（computed tomography, CT）等医学影像技术可作为逆转录聚合酶链式反应（reverse transcription-polymerase chain reaction, RT-PCR）检测方法的替代方案，展现出巨大应用潜力。正因如此，自动化图像分割作为临床决策支持工具的临床需求极为迫切。然而，公开可获取的COVID-19影像标注数据较为匮乏，导致传统机器学习模型极易出现过拟合问题。 方法：为解决上述数据匮乏引发的问题，本研究提出一种针对COVID-19感染区域的创新型自动化分割流程，通过引入变异数据库以适配小样本数据集的训练需求。本方法聚焦于模型训练阶段的实时动态图像块生成：通过多种预处理手段与大规模数据增强策略，生成独特且随机的图像块用于模型训练。为进一步降低过拟合风险，本研究采用标准三维U-Net（3D U-Net）架构，而非新型或计算复杂度较高的神经网络架构。 结果：通过对20例COVID-19患者的CT扫描影像开展k折交叉验证（k-fold cross-validation）以完成模型训练与验证，本研究成功构建了针对肺部组织与COVID-19感染区域的高精度、高鲁棒性分割模型，且在有限数据条件下未出现过拟合问题。本研究基于交叉验证结果与各类预处理技术对模型泛化能力的影响，通过敏感性分析，对本分割流程的鲁棒性展开了详细分析与讨论。本方法的戴斯相似系数（Dice similarity coefficients）结果如下：在验证集上，模型预测的COVID-19感染区域与放射科医师标注的分割结果之间的系数为0.804；在包含100例患者的独立测试集上，该系数为0.661。 结论：本研究证实，所提出的分割方法优于同类现有方案，推动了COVID-19影像分割领域的当前最优水平，并提升了小样本条件下医学影像分析的鲁棒性。本研究的代码与模型可通过以下链接获取：https://github.com/frankkramer-lab/covid19.MIScnn