Smooth volumetric copy-paste for ARDS CT database augmentation to improve lung segmentation

Our goal is to segment the lungs in computed tomography (CT) images of patients suffering from acute respiratory distress syndrome (ARDS). Indeed, comparing lung CT images at different respiratory conditions (e.g., end-inspiratory and end-expiratory), allows quantifying such phenomena as alveolar recruitment or cyclic hyperinflation, that help intensive care clinicians to adjust the artificial ventilation settings. The aim of the project is to improve the robustness of our deep model trained to segment the lungs in patients suffering of ARDS.To overcome the limitation of available annotated data from ARDS patients, and increase the diversity of lung morphologies, the MASTER student will use CT scans from patients with normal lungs and incorporate heterogeneous aeration patterns, as well as lesions of varying extent, density, and location, learned from ARDS patients. In a nutshell, the main contribution will be a tool capable of generating an ARDS CT-lung database with an unprecedented diversity of lesions and aeration patterns, with ground truth for both training and evaluation purposes.