SU‐F‐J‐84: Comparison of Quantitative Deformable Image Registration Evaluation Tools: Application to Prostate IGART

Abstract

Purpose:Errors in displacement vector fields (DVFs) generated by Deformable Image Registration (DIR) algorithms can give rise to significant uncertainties in contour propagation and dose accumulation in Image‐Guided Adaptive Radiotherapy (IGART). The purpose of this work is to assess the accuracy of two DIR algorithms using a variety of quality metrics for prostate IGART.Methods:Pelvic CT images were selected from an anonymized database of nineteen prostate patients who underwent 8–12 serial scans during radiotherapy. Prostate, bladder, and rectum were contoured on 34 image‐sets for three patients by the same physician. The planning CT was deformably‐registered to daily CT using three variants of the Small deformation Inverse Consistent Linear Elastic (SICLE) algorithm: Grayscale‐driven (G), Contour‐driven (C, which utilizes segmented structures to drive DIR), combined (G+C); and also grayscale ITK demons (Gd). The accuracy of G, C, G+C SICLE and Gd registrations were evaluated using a new metric Edge Gradient Distance to Agreement (EGDTA) and other commonly‐used metrics such as Pearson Correlation Coefficient (PCC), Dice Similarity Index (DSI) and Hausdorff Distance (HD).Results:C and G+C demonstrated much better performance at organ boundaries, revealing the lowest HD and highest DSI, in prostate, bladder and rectum. G+C demonstrated the lowest mean EGDTA (1.14 mm), which corresponds to highest registration quality, compared to G and C DVFs (1.16 and 2.34 mm). However, demons DIR showed the best overall performance, revealing lowest EGDTA (0.73 mm) and highest PCC (0.85).Conclusion:As expected, both C‐ and C+G SICLE more accurately reproduce manually‐contoured target datasets than G‐SICLE or Gd using HD and DSI metrics. In general, the Gd appears to have difficulty reproducing large daily position and shape changes in the rectum and bladder. However, Gd outperforms SICLE in terms of EGDTA and PCC metrics, possibly at the expense of topological quality of the estimated DVFs.