14:00
17:00

Thoracic imaging presents significant challenges, with each modality having its limitations: CT-scan provides high spatial resolution but involves ionizing radiation, while lung MRI, radiation-free, suffers from low contrast and artifacts. This thesis aims to synthesize CT images from UTE-MRI using deep learning models. Following an exploratory phase, a 2D GAN based on pix2pixHD was optimized with SPADE normalization and perceptual loss to enhance visual quality. A 2.5D approach and a robust 3D framework, built on nnU-Net with a loss prioritizing anatomical structures, further improved reconstruction. Results were evaluated using task-specific metrics and validated by radiologists, confirming their clinical potential. These advancements promise to expand the clinical adoption of UTE-MRI for non-invasive lung disease monitoring.

Amphi LaBRI