The Computer Vision Center, Eurecat and Universitat de Barcelona, together in the fight against COVID-19
Currently, with the PCR and antigen test, clinicians have really good tools for diagnosing the COVID-19 disease when the SARS-CoV-2 virus is active within the patient organism. However, these tests have two drawbacks: first, when patients suffer pulmonary infection, these tests do not provide precise information regarding its extension, and second, once the virus is eliminated by the immune system, the tests will be negative even if the patient is still affected by pneumonia induced by the original COVID-19 infection, a condition that can last for weeks.
Therefore, for getting a more complete perspective of the severity of the disease, clinicians have to use imaging techniques, as computed tomography (CT) scans. But, despite the increase in COVID-19’s detection accuracy through the use of CT images, the reading time necessary to interpret 3D CT volumes and to extract the morphological properties of the lesion can greatly increase the workload of radiologists. However, the use of Computer Vision and Artificial Intelligence tools can help to sensibly reduce the interpretation time.
With this objective in mind, Dr. Petia Radeva, CVC researcher and professor at the Universitat de Barcelona (UB), Giuseppe Pezzano and Drs Oliver Díaz and Vicent Ribas, from the Eurecat technology centre, have developed an automated method for COVID-19 detection using chest CT images, together with the segmentation of the Ground-Glass-Opacities (area of increased attenuation due to air displacement by fluid, airway collapse, fibrosis, or a neoplastic process) and other solidifications/fibrosis present inside the lungs, powered by deep learning strategies to support decision-making process. This study was recently published in the journal Computers in Biology and Medicine.
The procedure is simple: the lungs are firstly segmented from the input CT image to reduce the searching area. Afterward, the detection algorithm is used to analyse the lungs’ area in order to detect the presence of COVID-19. In the case of a positive finding, the CT image is processed to identify the areas affected by the disease.
This algorithm was tested with 79 COVID-19 CT volumes and 110 CT slices for three open-access COVID-19 CT image repositories, achieving an average accuracy for lesion segmentation near 99%. No false positives were observed in the detection network after 10 different runs.
The robustness and the accuracy of this work open up a wide range of other possible applications of this method. For example, the proposed network could be adapted, using fine-tuning, for studying the worst cases of pneumonia, the diffused metastasis or other lung diseases. Also, this methodology could be applied to detect and segment a large variety of organs in other fields of medical imaging analysis.
Reference: CoLe-CNN+: Context learning – Convolutional neural network for COVID-19-Ground-Glass-Opacities detection and segmentation. Giuseppe Pezzano, Oliver Díaz, Vicent Ribas and Petia Radeva, Computers in Biology and Medicine, September 2021, 136 104689. https://doi.org/10.1016/j.compbiomed.2021.104689