Purpose: We evaluated the accuracy of deformable image registration (DIR) for the pelvic region between magnetic resonance (MR) and computed tomography (CT) images using two different software programs.
Methods: This study used CT, T1-weighted (T1w) MR, and T2-weighted (T2w) MR images of the pelvic region of five male patients. DIRs were performed using Elastix, an open-source software, and Velocity AI, a commercial software. With Elastix, we used six patterns of cost functions with regularization weights λ = 0.01, 0.1, 1, 10, 100, and 200. For Velocity AI, we used a MR Corrected Deformable algorithm. The DIR accuracy was evaluated based on the Dice similarity coefficient (DSC) for prostate, bladder, rectum, and left and right femoral heads.
Results: The mean DSCs from all structures in T1w images were 0.84 ± 0.08 (rigid, Elastix), 0.88 ± 0.05 (DIR, λ = 0.1), 0.86 ± 0.09 (DIR, λ = 10), 0.85 ± 0.08 (DIR, λ = 200), and 0.86 ± 0.07 (DIR, Velocity AI), respectively. For the T2w images, the mean DSCs were 0.85 ± 0.08 (rigid, Elastix), 0.85 ± 0.09 (DIR, λ = 0.1), 0.86 ± 0.09 (DIR, λ = 10), 0.85 ± 0.08 (DIR, λ = 200), and 0.85 ± 0.08 (DIR, Velocity AI), respectively. DIR accuracy in elastix was as good as Velocity AI when λ = 0.1 (10) between CT-T1w (CT-T2w) images. The λ at which the mean DSC was highest differed for each patient.
Conclusion: The results suggested that DIR accuracy comparable to that of commercial DIR systems could be achieved by verifying the penalty weight for DIR between CT and MR images for each sequence and each facility.
Funding Support, Disclosures, and Conflict of Interest: This study was partially supported by the grant from the Japan Agency for Medical Research and Development.