Room: AAPM ePoster Library
Magnetic resonance-guided radiation therapy (MRgRT) is capable of increasing the precision of beam delivery by tracking a tumor target on continuous 2D cine MRI and gating the radiation beam. Tracking results are often not accurate or robust due to the low signal-to-noise ratio (SNR) on the 0.35 T Viewray MRI-Linac, leading to inconsistent beam gating. This research proposes the use of a deep learning-based deformable image registration (DIR) method, LiteFlowNet, and a multiple registration reference strategy (MRRS) to improve the accuracy and robustness of target tracking at low field.
2D cine MRIs (4 frames/s) of 5 lung and 6 pancreas cancer patients, 2 fractions per patient and 40 frames per fraction, were retrospectively obtained and analyzed. Targets in all frames were manually contoured by two observers. LiteFlowNet + MRRS was used to register 2D MRIs frame-by-frame to track the target. Results were compared with traditional Demons DIR algorithm + single reference registration strategy (SRRS) and Demons + MRRS. SRRS only utilizes one reference frame as the fixed image in the DIR computation. MRRS conducts 10 independent DIRs using 10 reference frames that cover the typical respiration amplitude and then derives the tracking contour by merging all 10 DIR results.
Four evaluation metrics, dice similarity score (DSC), centroid, mean, and Hausdorff distance were computed. The results suggested that LiteFlowNet + MRRS consistently archived the highest accuracy compared to ground-truth manual contours with DSC = 0.92±0.04, where DSC = 0.81±0.09 between observers. Results also showed that Demons with SRRS or MRRS could not provide robust and accurate results with DSC scores = 0.63±0.29 and 0.59±0.31, respectively.
LiteFlowNet with the multiple registration reference strategy could robustly track tumor targets in 2D Cine MRIs.
Image-guided Therapy, Registration, MRI