3D Semantic Segmentation of Airway Abnormalities on UTE-MRI with Reinforcement Learning on Deep Supervision

Cystic fibrosis (CF) monitoring traditionally relies on CT scans, which involve radiation exposure concerns. Ultra-short echo time (UTE) MRI has emerged as a promising radiation-free alternative for lung imaging. However, automated segmentation of CF lesions on UTE MRI has not been reported yet, mainly due to lower signal-to-noise ratio and contrast compared to CT. This study evaluates the feasibility of fully automated semantic segmentation of three main hallmarks of CF: bronchiectasis, bronchial wall thickening, and bronchial mucus. To address the challenges of low proton MRI signal and resolution, we propose a novel Reinforcement learning for deep Supervision adapted to nnU-Net (RiSeNet). This approach enhances the standard nnU-Net by dynamically adjusting deep supervision weights during training through reinforcement learning. We compare RiSeNet against both the baseline nnU-Net and selected state-of-the-art architectures: SAMed and nnSAM for global context modeling, MedNeXt for large-scale feature capture, and U-Mamba for efficient volumetric processing. All models were evaluated using registered same-day CT-derived ground truth labels. Results demonstrate RiSeNet's superior performance in both accuracy and efficiency when handling the unique challenges of UTE MRI segmentation.