Achievable Accuracy of DIR for Tumor/Organ With Large Progressive Shrinkage During the Radiation Treatment: A Bio-Tissue Phantom Study

Abstract

<h3>Purpose/Objective(s)</h3> The purpose of the study was to investigate voxel-wise accuracy of different deformable registration (DIR) schemes on tumor/organ with large shrinkage during radiotherapy, and to find the most suitable schemes for specific clinical scenarios. <h3>Materials/Methods</h3> An intact porcine liver lobe was utilized and 60 gold markers of 0.5 × 0.5 × 1mm were implanted both on the surface and inside to construct the phantom. To simulate the progressive radiation-induced shrinkage, it was heated by a microwave incrementally from 60 to 200s in 7 phases. For each phase, the phantom was heated and then CT scanned. Two extra image sets were generated from the original CTs by: 1) overriding the gold markers (feature image), and 2) overriding the entire phantom to the mean intensity (featureless image). Ten DIR schemes were evaluated on the secondary images to simulate different combinations of the boundary conditions, availability of image feature or intermediate imaging, and different algorithms as listed in Table 1 (Direct: D, Sequential: S; Surface: SU, Surface-marker: SM; image feature: IF, featureless: NF; IM-DIR: IM, BM-DIR: BM). For example, D/SM/NF/BM means a direct BM-DIR with surface-marker constrain on featureless image. The surface constrain forced the surface matching, while surface-marker constrain further enforced the matching of all surface markers. Direct DIR was between initial and target phase, while sequential DIR utilized all intermediate phases. IM-DIR is intensity based while BM-DIR is a hybrid DIR based on bio-mechanical model. Target registration errors (TRE) using the internal marks as ground truth and Jacobian map (JM) were evaluated for all schemes. <h3>Results</h3> The volume reduction was about 20%, 30% and 40% after 90s, 130s and 200s of heating, respectively. The TREs of all 10 DIR schemes are listed in Table 1. Without image features and with surface constrain, the schemes 1-4 achieved average TRE of 3.8 to 5.0 mm, while schemes 5 to 7 with the surface-marker constrain improved to 2.4 to 3.2mm. The best TRE of 1.9mm was from scheme 10, which combined the image feature, surface-marker constrain and intermediate imaging, but appeared large JM irregularity. The scheme 6 and 7, which utilized BM-DIR with better boundary condition, achieved the second-best TRE of 2.5 and 2.4mm and 90% points with TRE less than 4mm, as well as smooth JM. <h3>Conclusion</h3> Achievable DIR accuracy for tumors with large progressive shrinkage depends on multiple factors such as image feature within and around tumor, volume reduction level, and the underlying regularization of DIR. For tumors with salient feature surrounding its boundary, but in the absence of internal contrast, the hybrid DIR method (scheme 6) could achieve registration accuracy within 4mm.