ScanCovid IA: Artificial Intelligence for Predicting Infection Severity

Changed on 23/06/2020
What if the combined analysis of our health data (clinical, biological, scanners... ) allowed health workers to predict and refine the severity of the virus attack? A scientific feat made possible by artificial intelligence but also by close collaboration between teams from the Gustave Roussy Institute, the Kremlin-Bicêtre Hospital - APHP, the Owkin startup and Inria.
Photo ScanCovid IA
L'équipe partenaire à l'Institut Gustave Roussy

What is the genesis of your project?

For several years now, the OPIS project-team has been working on a sustained basis with the doctors of the Imaging Department of the Institut Gustave Roussy (IGR) on issues related to learning and artificial intelligence and their contribution to medical imaging for the diagnosis, prognosis and follow-up of cancer patients.

The IGR team contacted our research team at the beginning of March 2020 to launch a collaborative project: to work on the prediction of the severity of Covid-19 damage from AI analyses of 3D thoracic scans of patients.

Project holder : Emilie Chouzenoux (EPC OPIS)

Partners : Institut Gustave Roussy, Hôpital Bicêtre AP-HP, Owkin CentraleSupélec

#IA #data #prediction

The project then expanded. Our initial consortium was joined by a team of radiologists from the Kremlin-Bicêtre Hospital - APHP and by a Parisian startup, Owkin, a specialist in AI for medicine. The team, thus strengthened, tackled an even more ambitious goal: to combine in AI analysis a wide variety of heterogeneous data (scanner, biological, clinical and patient medical history data) in order to predict even more accurately the severity of the infection.

How is it developing today and what are its objectives?

The project received the "Mission Covid" label from Inria at the beginning of May 2020. A first research article was submitted in mid-May in a prestigious medical journal: we show the importance of integrating the information present in chest scans to reliably predict the severity of the disease.


We plan to continue our study by diversifying the hospital centers from which the patient cohorts under consideration come, and by increasing the volume of annotated images.

In this project, the long-term goal for our project team is to develop open-source software to help radiologists in the task of automatically segmenting and quantifying different types of lung lesions.

How do you work with your partners?

Multi-weekly video-conference meetings are held with representatives of the four partners, in order to best coordinate the progress of the project.

The OPIS project-team and the start-up Owkin are attentive to the needs of the doctors, and take into account all their feedback on both the methodology and the results obtained (both quantitative and visual) by the AI approaches developed. The objective of our collaboration is to converge towards the most reliable and relevant analysis of medical data possible.