OBJECTIVE: MRI-guided radiotherapy (MRIgRT) is an emerging treatment technique where anatomical and pathological structures are imaged through integrated MR-radiotherapy units. This work aims 1) at assessing the accuracy of optical-flow-based motion tracking in liver cine-MRI sequences; and 2) at testing a MRIgRT workflow combining similarity-based image matching with image registration. METHODS: After an initialization stage, a set of template images is collected and registered to the first frame of the cine-MRI sequence. Subsequent incoming frames are either matched to the most similar template image or registered to the first frame when the similarity index is lower than a given threshold. The tracking accuracy was evaluated by considering ground-truth liver landmarks trajectories, as obtained through the scale-invariant features transform (SIFT). RESULTS: Results on a population of 30 liver subjects show that the median difference between SIFT- and optical flow-based landmarks trajectories is 1.0 mm, i.e., lower than the cine-MRI pixel size (1.28 mm). The computational time of the motion tracking workflow (<50 ms) is suitable for real-time motion compensation in MRIgRT. Such time could be further reduced to ≍30 ms with limited loss of accuracy by the combined image matching/registration approach. CONCLUSION: The reported workflow allows us to track liver motion with accuracy comparable to robust feature matching. Its computational time is suitable for online motion monitoring. SIGNIFICANCE: Real-time feedback on the patient anatomy is a crucial requirement for the treatment of mobile tumors using advanced motion mitigation strategies.
OBJECTIVE: MRI-guided radiotherapy (MRIgRT) is an emerging treatment technique where anatomical and pathological structures are imaged through integrated MR-radiotherapy units. This work aims 1) at assessing the accuracy of optical-flow-based motion tracking in liver cine-MRI sequences; and 2) at testing a MRIgRT workflow combining similarity-based image matching with image registration. METHODS: After an initialization stage, a set of template images is collected and registered to the first frame of the cine-MRI sequence. Subsequent incoming frames are either matched to the most similar template image or registered to the first frame when the similarity index is lower than a given threshold. The tracking accuracy was evaluated by considering ground-truth liver landmarks trajectories, as obtained through the scale-invariant features transform (SIFT). RESULTS: Results on a population of 30 liver subjects show that the median difference between SIFT- and optical flow-based landmarks trajectories is 1.0 mm, i.e., lower than the cine-MRI pixel size (1.28 mm). The computational time of the motion tracking workflow (<50 ms) is suitable for real-time motion compensation in MRIgRT. Such time could be further reduced to ≍30 ms with limited loss of accuracy by the combined image matching/registration approach. CONCLUSION: The reported workflow allows us to track liver motion with accuracy comparable to robust feature matching. Its computational time is suitable for online motion monitoring. SIGNIFICANCE: Real-time feedback on the patient anatomy is a crucial requirement for the treatment of mobile tumors using advanced motion mitigation strategies.
Authors: Naoyuki Shono; Brian Ninni; Franklin King; Takahisa Kato; Junichi Tokuda; Takahiro Fujimoto; Kemal Tuncali; Nobuhiko Hata Journal: Med Phys Date: 2020-03-28 Impact factor: 4.071