Improving the robustness of deep-learning models in lung segmentation and registration: generating CT image pairs at different inflation levels with dense lung lesions for data augmentation

The heterogeneous aeration of diseased lung parenchyma can be assessed using computed tomography (CT). This assessment is crucial for identifying the patient's phenotype in cases of severe lung function impairment, such as acute respiratory distress syndrome (ARDS), to personalize artificial ventilation settings. To do so, lung tissues must be delineated and aligned in scans acquired at different inflation levels. Deep-learning models implemented by team achieve these tasks in a matter of minutes with uncertainties close to inter-expert variability. Our approach consists in exploiting the complementarity and redundancy of information carried by scans representing the same lungs at different inflation levels. To increase the success rate and accuracy, it is necessary to expand and diversify the training database beyond the scans that can be collected in clinics. 

The main objective of this MSc project is to improve the accuracy and robustness of our models. To this end, the candidate will have to develop a method capable of generating highly realistic pairs of synthetic CT scans with different morphologies and lesions, as well as with heterogeneous density changes between simulated inflation levels. Improvements to model architecture and training strategy may also be considered, depending on the candidate's skills.