• daras@iti.gr
  • zarpalas@iti.gr
  • theosem@iti.gr

VCL's rapid response to COVID-19 crisis

VCL is conducting cutting edge AI research for detecting COVID-19 infections in CT-scans and X-Rays. Our aim is to develop a data-driven tool-set based on a large amount of annotated data from patients with COVID-19 infection. In collaboration with the Humanitas Research hospital we have access to anonymized data from COVID-19 patients, annotated by HUM radiologists, as well as openly available datasets. Our ultimate goal is to provide a solution that can support health organisations by providing guidance in:

  • Prioritising infected organs during diagnosis
  • Prioritising patients who need ICU admission, based on their health history, underlying diseases and severity of infection symptoms
In order to reach our goal, we develop diverse methods for medical image segmentation and classification segmentation (along with data augmentation) in order to set up an ecosystem able to discern the patients in the following classes:

  • Level 1: Patients who can be discharged and monitored from their home
  • Level 2: Patients requiring hospitalization but not intensive care
  • Level 3: Patients who need ICU admission

Classifying types of pneumonia infections

In order to classify pneumonia infections and further explore pathogenic lesions in CT-scans, one of the most essential steps is the lung boundaries identification from the surrounding thoracic tissues [2].

CT-scan lung segmentation

To this end, one of the main steps of our approach is to perform lung segmentation on CT-scans. We are employing machine learning solutions and neural networks for acquiring accurate lung segments [3].

Pulmonary infection identification

A large amount of COVID-19 cases is related with various types of pulmonary infections such as Ground-Glass Opacities (GGO), consolidation, vascular enlargment in the lesion, or traction bronchiectasis [1]. Therefore, one of the main steps of our pipeline is identifying pneumonia related lung infections.

As a response to the recent coronavirus outbreak. X-Ray and CT-scan pneumonia related datasets have been complemented with samples from COVID-19 patients, however, up to this time COVID-19 samples remain limited. Therefore, we are conducting experiments for classifying pneumonia infections from X-Rays and CT-scans including COVID-19 samples, while our methods are constantly updated depending on the availability of new data.

Data augmentation

Despite the tremendous amount of COVID-19 related pneumonia cases since December 2019, there still is a lack of related data and CT-scans. In addition, annotating medical data requires extensive human effort, significant expertise and time. Therefore, apart from standard approaches for augmenting medical images (such as affine and elastic transformation) we plan to explore latest advances in Generative Adversarial Networks (GAN) [4] to increase the number of samples of the utilized datasets

Data Fusion

Despite the fact that valuable information can be obtained by segmenting infected organs, a more solid solution would jointly examine the visual findings with other clinical metadata of individual patients, such as underlying diseases or health history. However, due to confidentiality issues, clinical metadata are not publicly available. The dataset provided by the Humanitas hospital includes anonymized Electronic Medical Records, such as demographic variable, clinical and laboratory data and Sepsis Organ Failure Assessment (SOFA) score. In the presence of such data our final toolset can be considerably more effective, by employing data fusion solutions, for fusing the infection-related visual findings with other clinical metadata.

References:

[1] Zhao, W., Zhong, Z., Xie, X., Yu, Q. and Liu, J., 2020. Relation between chest CT findings and clinical conditions of coronavirus disease (COVID-19) pneumonia: a multicenter study. American Journal of Roentgenology, pp.1-6.

[2] Mansoor, A., Bagci, U., Foster, B., Xu, Z., Papadakis, G.Z., Folio, L.R., Udupa, J.K. and Mollura, D.J., 2015. Segmentation and image analysis of abnormal lungs at CT: current approaches, challenges, and future trends. RadioGraphics, 35(4), pp.1056-1076.

[3] Azad, R., Asadi-Aghbolaghi, M., Fathy, M., & Escalera, S. (2019). Bi-Directional ConvLSTM U-Net with Densley Connected Convolutions. In Proceedings of the IEEE International Conference on Computer Vision Workshops (pp. 0-0).

[4] Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J. and Greenspan, H., 2018. GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification. Neurocomputing, 321, pp.321-331.