3D Tomographic Pattern Synthesis for Enhancing the Quantification of COVID-19
The Coronavirus Disease (COVID-19) has affected 1.8 million people and resulted in more than 110,000 deaths as of April 12, 2020. Several studies have shown that tomographic patterns seen on chest Computed Tomography (CT), such as ground-glass opacities, consolidations, and crazy paving pattern, are correlated with the disease severity and progression. CT imaging can thus emerge as an important modality for the management of COVID-19 patients. AI-based solutions can be used to support CT based quantitative reporting and make reading efficient and reproducible if quantitative biomarkers, such as the Percentage of Opacity (PO), can be automatically computed. However, COVID-19 has posed unique challenges to the development of AI, specifically concerning the availability of appropriate image data and annotations at scale. In this paper, we propose to use synthetic datasets to augment an existing COVID-19 database to tackle these challenges. We train a Generative Adversarial Network (GAN) to inpaint COVID-19 related tomographic patterns on chest CTs from patients without infectious diseases. Additionally, we leverage location priors derived from manually labeled COVID-19 chest CTs patients to generate appropriate abnormality distributions. Synthetic data are used to improve both lung segmentation and segmentation of COVID-19 patterns by adding 20% of synthetic data to the real COVID-19 training data. We collected 2143 chest CTs, containing 327 COVID-19 positive cases, acquired from 12 sites across 7 countries. By testing on 100 COVID-19 positive and 100 control cases, we show that synthetic data can help improve both lung segmentation (+6.02% lesion inclusion rate) and abnormality segmentation (+2.78% dice coefficient), leading to an overall more accurate PO computation (+2.82% Pearson coefficient).