Automatic statistical diagnosis of COVID-19 based on multi-modal CT feature extraction

Background and purpose: Computed tomography (CT) is highly sensitive to lung-related abnormalities as a non-invasive method and has become an essential tool for screening and diagnosing Coronavirus disease 2019 (COVID-19). To reduce the stress of work for physicians and speed up diagnosis, we propose a novel automatic diagnosis pipeline for COVID-19 based on high-dimensional radiomic features extracted from multimodal CT scans (multi-geometric and multiscale). Materials and methods: There are 746 CT scans involved in this study, where 349 CT scans are COVID-19 positive and 397 CT scans are COVID-19 negative. All of them are from the public dataset. We first construct a transfer learning-based auto-segmentation model with a morphological post-processing block to improve the lung region segmentation. Then the radiomics feature extraction is guided by the proposed multi-modal CT scans strategy. In addition, our automatic diagnosis pipeline is driven by a well-designed loss function. We also explain the diagnosis capability from the related theory of linear subspace spanned by multi-modal radiomics features. Results: Under the 10-fold cross-validation strategy, our approach can achieve an improvement in diagnostic performance of 5. 77%, 7. 78%, 7. 74%, 7. 78%, 7. 45% compared to the radiomic features extracted from the original CT scans, and diagnosis performance is promoted to 91.53%, 86.46%, 86.47%, 86.46%, 86.95% in terms of AUC, Acc, F1, Recall and Precision in public datasets. Conclusions: We demonstrate a statistically significant improvement of the proposed statistical learning method compared to the state-of-the-art machine learning-based diagnosis approaches. Thanks to theoretical support and excellent diagnostic performance, our method can be deployed in clinical auxiliary diagnosis, releasing the overstretched medical resources.