Segmenting lesion regions of Coronavirus Disease 2019 (COVID-19) from computed tomography (CT) images is a challenge owing to COVID-19 lesions characterized by high variation, low contrast between infection lesions and around normal tissues, and blurred boundaries of infections. Moreover, a shortage of available CT dataset hinders deep learning techniques applying to tackling COVID-19. To address these issues, we propose a deep learning-based approach known as PPM-Unet to segmenting COVID-19 lesions from CT images. Our method improves an Unet by adopting pyramid pooling modules instead of the conventional skip connection and then enhances the representation of the neural network by aiding the global attention mechanism. We first pre-train PPM-Unet on COVID-19 dataset of pseudo labels containing1600 samples producing a coarse model. Then we fine-tune the coarse PPM-Unet on the standard COVID-19 dataset consisting of 100 pairs of samples to achieve a fine PPM-Unet. Qualitative and quantitative results demonstrate that our method can accurately segment COVID-19 infection regions from CT images, and achieve higher performance than other state-of-the-art segmentation models in this study. It offers a promising tool to lay a foundation for quantitatively detecting COVID-19 lesions.

从计算机断层扫描 (CT) 图像中分割 2019 年冠状病毒病 (COVID-19) 的病变区域是一项挑战，因为 COVID-19 病变的特点是变异大、感染病变与正常组织周围的对比度低以及感染边界模糊。此外，可用 CT 数据集的短缺阻碍了应用于解决 COVID-19 的深度学习技术。为了解决这些问题，我们提出了一种基于深度学习的方法，称为 PPM-Unet，用于从 CT 图像中分割 COVID-19 病变。我们的方法通过采用金字塔池化模块而不是传统的跳过连接来改进 Unet，然后通过辅助全局注意力机制增强神经网络的表示。我们首先在包含 1600 个样本的伪标签的 COVID-19 数据集上预训练 PPM-Unet，产生一个粗略的模型。然后我们在由 100 对样本组成的标准 COVID-19 数据集上微调粗 PPM-Unet，以实现精细 PPM-Unet。定性和定量结果表明，我们的方法可以从 CT 图像中准确地分割出 COVID-19 感染区域，并且在本研究中比其他最先进的分割模型具有更高的性能。它提供了一个很有前途的工具，为定量检测 COVID-19 病变奠定了基础。