Robust Water-Fat Separation in Nasal MRI Using Hybrid Model and Data Driven Network

Motivation: Robust water-fat separation is crucial for nasal lesion detection and tracking. Existing methods often suffer from limited contrast and issues with water-fat swaps. Goal(s): We aim to accurately extract water/fat-only images from multi-contrast nasal MRI using two-echo FSE sequences. Approach: We propose a hybrid model and data driven network to provide reliable water-fat separation from two-echo signals based on end-to-end learning. A 3D physical model is also used to guide the network to learn layer-wise features. Results: The proposed framework outperforms the representative methods, with PSNR of 50.57 and SSIM of 0.9983. Visually, our method also provides more precise separation of nasal structures. Impact: Our proposed hybrid separation framework combines multi-task networks with a 3D physical model to enhance the robustness of water/fat separation in multi-contrast nasopharyngeal MRI images, significantly expanding its clinical applicability.