DeepChest: Intelligent diagnosis and triage of COVID-19
© Image: atriumhealth.org
A new interdisciplinary clinical research group, Intelligent Global Health (iGH), hosted at the Machine Learning and Optimization Laboratory (MLO), has developed a novel deep learning model to automate diagnosis and triage in COVID-19 from over a thousand lung ultrasound (LUS) images, collected in the CHUV emergency room.
The model, DeepChest, matches and slightly outperforms expert evaluation by clinicians. The model was developed by Hugo Schmutz as an MSc thesis in iGH, and co-supervised by Dr Mary-Anne Hartley (clinician researcher and head of iGH) and Jean-Baptiste Cordonnier (MLO-PhD student). It makes use of a state-of-the-art transformer architecture to better aggregate images at a patient level and be robust to variability in the number of images acquired per patient.
LUS is a simple, non-invasive imaging technique to view the lung surface at the patient’s bedside. It is a cheap hand-held device, emitting no radiation and is pluggable into a mobile phone. This is particularly useful in COVID-19, where its portability enables decentralised respiratory evaluations (at home or in triage centres rather than in the hospital) and simple inter-patient disinfection.
The models are now being improved for greater interpretability and implementation by MSc students Mariko Makhmutova,Lilia Ellouz, and Deeksha Rao and further patients will be recruited by Dr Noemie Boillat and Dr Thomas Brahier at CHUV.