Generalization of muscle segmentation models for 3D fat fraction quantification with different Dixon MRI protocols

One of the main obstacles to deploying deep learning (DL) based segmentation models for analysis of MR images in clinical routine practice is their limited ability to generalize across different scanners and imaging protocols. Variability in MRI hardware and acquisition settings, such as field strength, gradient performance, and sequence parameters, can significantly affect image characteristics, making DL models trained under controlled conditions less robust in real-world clinical environments. In the diagnosis of neuromuscular diseases (NMD), Dixon MRI is a widely used imaging protocol for the assessment of muscular fattening due to its ability to separate fat and water signals by exploiting the slight frequency difference between fat and water protons. Dixon MRI yields several images with different contrast: in-phase (IP), out-of-phase (OP), fat-only (F), and water-only (W), from which the proton density fat fraction (PDFF) image is calculated as F/F+W . Per muscle 3D fat fraction (%FF) can then be quantified from the PDFF image provided that a 3D segmentation of each muscle in the Dixon MRI images is available. In our center, a variety of Dixon MRI protocols and scanners are used depending on the clinical context. Diagnostic scans, where speed is essential, typically use 2-point or 3-point Dixon sequences. In contrast, clinical studies prioritize accurate fat quantification and therefore employ 6-point Dixon imaging. The use of both 1.5T and 3T scanners further increases protocol diversity, underscoring the need for segmentation models that can generalize across these variations. In existing literature, a range of combinations of Dixon image types have been used as input to convolutional neural networks (CNNs) for 3D muscle segmentation. Some studies use the F and W images, while others include the IP and OP images, or rely solely on the IP or F images. This diversity in input strategies reflects the lack of consensus on the optimal image type combination for robust segmentation performance. This study investigates the generalizability of muscle segmentation models trained on Dixon MRI images of the proximal leg and examines the impact of the choice of different input image types on the robustness of model performance across different datasets acquired with different acquisition setups.