Purpose: COVID-19 has become a global pandemic and is still posing severe health risk for publics. Accurate and efficient segmentation of pneumonia lesions in CT scans is vital for the disease diagnosis and treatment decision-making. To automatically segment the pneumonia lesion, we proposed a novel unsupervised approach using cycle-consistent generative adversarial network(cycle-GAN) which could liberate people from heavy burden of manual labeling and accelerate the diagnosis process.
Methods: To restrain the CT volume of interest used for lesion segmentation, a previously proposed method was used to delineate the whole lung first. The cycle GAN was developed to generate synthetic healthy CT lung volume from infected patients. Then, the pneumonia lesions were extracted by subtracting the synthesis healthy CT from the infected CT one. Noise and isolated points were eliminated using maximum connected component, threshold segmentation and gaussian smoothing.
Results: Two public datasets (Coronacases and Radiopedia) were used to evaluated our approach. The dice reached 0.709 on average in the Coronacases, and 0.653 in the Radiopedia. Both are comparable with existing supervised approaches and outperforms previous unsupervised methods. Meanwhile, our approach reaches a specificity of 0.682 and sensitivity of 0.811 in Coronacases, and 0.655 and 0.706 respectively, in Radiopedia. Comparing with existing label-free method, the proposed method outperforms in all index except specificity for Coronacases which varies due to lesion shape, size and position.
Conclusion: The proposed unsupervised method could automatically delineate the pneumonia lesion in a highly efficient and accurate way. It can also be used as a baseline for further manual modification and a quality assurance tool in clinical diagnosis.