Computational Lung
Our focus is the acute respiratory distress syndrome (ARDS). Our main goal is to help the clinicians choose the best settings of mechanical ventilation for each patient. The reference imaging technique is computed tomography (CT), which gives access to three-dimensional information of air-to-tissue ratio within the parenchyma. Changes between CT scans acquired at different respiratory conditions (e.g., end-inhale and end-exhale) are used to assess such phenomena as alveolar recruitment, cyclic hyperinflation, etc.
  • Lung segmentation: Assessment of pathological changes within the lungs requires the delineation of the latter, which is hampered by dramatic changes in contrast occurring in ARDS and some other diseases. The challenge depends on the type and spread of the desease but a general trend is the risk of under-estimation of the lung volume or even missing some important regions. To overcome these  issues, we are investigating both analytical [1] and data-driven approaches. Our current developments focus on using deep learning techniques to train segmentation models. In addition to using deep neural networks for the lung segmentation task, we also apply this methodology within the COVID-19 transversal project to segment the motion mask used for Ventilation Imaging (CTVI).
  • Lung registration: Accurate alignment of pulmonary structures between different CT scans is a key to subsequent computation of 3D maps representing recruitment or inflation. Here also both conventional and learning-based approaches are investigated. The applications mainly include ARDS, but also the chronic obstructive pulmonary disease (COPD) in collaboration with STROBE laboratory (Grenoble).

References :

[1] A. Morales Pinzón, M. Orkisz, J.-C. Richard, and M. Hernández Hoyos. Lung Segmentation by Cascade Registration. Innovation and Research in BioMedical engineering, 38(5):266 – 280, October2017. doi: 10.1016/j.irbm.2017.07.003. URL