KISEG: A Three-Stage Segmentation Framework for Multi-level Acceleration of Chest CT Scans from COVID-19 Patients

During the ongoing COVID-19 outbreak, it is critical to perform an accurate diagnosis of COVID-19 pneumonia by computed tomography (CT). Although chest lesion segmentation plays a pivotal role in computer-aided diagnosis (CAD), accuracy is hindered by the lack of a publicly available CT dataset with manual annotation. In addition, for clinical deployment, how to balance the accuracy versus efficiency for the semantic segmentation model remains challenging. To address these issues, we construct the first CT dataset of COVID-19 pneumonia with pixel-wise lesion annotations. We propose a three-stage framework, called KISEG (Key and Intermediate frame of Segmentation), to enhance performance on serial CT image segmentation with multi-level acceleration. We first take a policy to divide frames of serial CT into two groups, key frames and intermediate frames. Then KISEG employs a main model (accurate but cumbersome) for key frame segmentation. And third, an auxiliary model was employed for intermediate frame segmentation with incorporating the information of key frames during the fusion module. Moreover, we propose a Gaussian Kernel Dropout for data augmentation. Experiments on our dataset demonstrate that our proposed KISEG achieves comparable accuracy with state-of-the-art methods and fewer GFLOPs, speeding up from 2.88\(\times \) to 9.16\(\times \). This dataset has been made public for further research of COVID-19 for AI community, released on http://ncov-ai.big.ac.cn/download.