Evaluation of volumetric breathing motion prediction of the stomach using dynamic golden-angle radial MRI

Abstract Objective : The motion of abdominal organs complicates accurate radiotherapy planning and delivery. Accurate short-term prediction of respiratory motion is essential for image-guided adaptive radiotherapy in the abdomen. This study evaluates the performance of a kernel ridge regression–based breathing motion prediction framework (KRR-BMP) for abdominal organs, with particular emphasis on the stomach and its centerline. Approach : Dynamic MRI datasets from 27 scans acquired in 16 patients were processed to generate low-rank breathing motion models based on principal component analysis (PCA) of deformation vector fields (DVFs). KRR-BMP was used to predict future PCA coefficients of breathing motion, enabling reconstruction of volumetric breathing motion deformation fields at specified prediction horizons. The accuracy of predicted breathing DVFs was evaluated across multiple organs and anatomical structures, training set sizes, PCA ranks, and prediction horizons. Main results : At a 340 ms prediction horizon, KRR-BMP achieved sub-voxel accuracy across all evaluated organs, with population-mean endpoint errors below 1.3 mm and 95th percentile errors consistently below 3 mm. Error distributions were relatively narrow across subjects and training sets, demonstrating robustness and generalizability. Reductions in training data led to predictable increases in error, particularly for gastric structures, although population-level upper-bound errors remained within clinically relevant ranges. A PCA rank of two provided the best balance between accuracy and robustness, while higher ranks increased inter-subject variability. Significance : KRR-BMP provides accurate and robust short-term respiratory motion prediction for abdominal organs in MR-guided radiotherapy. The demonstrated performance for stomach subregions establishes a foundation for future integration of respiratory motion prediction with real-time gastric motility modeling and motion-aware adaptive treatment workflows.