SU‐E‐J‐72: Shift Invariant Feature Transform (SIFT) Based Image Stitching for Panoramic Cone Beam CT (CBCT)

Purpose: Panoramic CBCT is a novel imaging technique that stitches together projection images from multiple views to increase the imaging volume to as large as practically needed using half (200 degree gantry rotation) scans only. In this study, we investigated a SIFT‐based method to improve the accuracy of image stitching for panoramic CBCT. Methods: Panoramic CBCT data were acquired for a torso phantom using two half scans: (A) the detector at the center position and (B) the detector shifted 148mm toward the right side of the phantom. For each projection image of scan A, a corresponding projection image of the same gantry angle was interpolated for scan B using images of two neighboring gantry angles. The SIFT method was applied to extract ‘matching’ feature points in the overlapping region of the projection images of the same gantry angle from each scan. The rigid (translation and rotation) transform matrix derived from this analysis was then used to stitch together the two projection images for CBCT reconstruction using the simultaneous algebraic reconstruction technique (SART) method. Three overlapping (5‐cm, 3‐cm and 2‐cm) sizes were used for the SIFT analysis to determine the optimal imaging parameters. Results: The CBCT reconstructions using the SIFT stitching method showed significant artifact reduction compared with the direct stitching method. The success rate of matching (less than 0.5 degrees rotation and 2 pixels horizontal/vertical shifts) was close to 100% for 5‐cm overlapping and gradually reduced as the overlapping size decreased. The minimum overlapping size we could reliably apply our method was 2 cm (93.8% success rate) below which significant operator involvement was required to determine the applicable rigid transform parameters. Conclusion: We have demonstrated the effectiveness of the SIFT stitching method for panoramic CBCT reconstruction. We are currently improving the registration technique to further reduce the required overlapping size. This work was partially supported by a DOD grant DOD W81XWH1010862