Abstract
BACKGROUND AND PURPOSE: Deep learning-based tumor tracking is promising for real-time magnetic-resonance-imaging (MRI)-guided radiotherapy. We investigate the applicability of a tumor tracking model developed for 0.35 T MRI-linac sagittal cine-MRI for 1.5 T interleaved orthogonal cine-MRI and implement transfer learning to further improve its performance. MATERIALS AND METHODS: We collected 3600 cine-MRI frames in sagittal, coronal and axial planes from 24 patients (validation 10, testing 14) treated on a 1.5 T MRI-linac, where two expert clinicians manually segmented target labels. A transformer-based deformation model trained on 0.35T MRI-linac images (baseline model, BL) was evaluated and used as a starting point to train patient-specific (PS) models. The Dice similarity coefficient (DSC) and the surface distance (50th and 95th percentiles, SD50%, SD95%) were used to compare the obtained target segmentations with the ground truth labels. The percentage of negative Jacobian determinant values (NegJ), accounting for the folding pixel ratio, was determined. RESULTS: Outperformed by all the PS models, the BL model averaged in a DSC of 0.85, SD50% of 1.9 mm, SD95% of 5.9 mm and NegJ of 0.45 % in testing. The best PS model averaged in a DSC of 0.90, SD50% of 1.3 mm, SD95% of 3.9 mm and NegJ of 0.02 % in testing. CONCLUSION: We have found the 0.35 T model trained on sagittal cine-MRIs cannot be directly applied to a 1.5 T interleaved orthogonal cine-MRI system. However, PS transfer learning could improve the target tracking performance and reach an accuracy comparable to the inter-observer variability.