Utilizing Deformable Image Registration and Quality Assurance to Automate Analysis of Anatomic Changes After Head and Neck Radiation Therapy

Abstract

Deformable image registration (DIR) algorithms are clinically utilized in radiation therapy without quality assurance (QA) of the deformation vector fields (DVFs) produced. This study investigated the feasibility of simultaneous automated detection of anatomic changes and DIR-QA for patients who receive head and neck (H&N) radiation therapy. For 10 patients with H&N cancer, three DIR algorithms were used to map initial planning computed tomography (CT) images to posttreatment images and analyze anatomic changes in organs-at-risk (OARs) and gross tumor volumes (GTVs). Automated analysis of anatomic changes was carried out by computing the Jacobian determinant, |J|, of the deformation vector field (DVF) resulting from each of 3 different, commercially available DIRs, which map image intensity between images. |J| is the relative deformation of each mapped voxel, and we focused on the 50% deformation level, i.e., where voxels expand or contract by at least 50% (J50). The Jacobian volume histogram (JVH) for OARs and GTVs described the volume of each structure, which deformed at a specified level. In order to differentiate DVF error from true deformation in QA of the DVFs, we utilized the magnitude of vector curl, |C|, and knowledge of limitations in dense tissue deformation (e.g., in bone we expect J50 = 0). Deformation vector field regions with relatively large |C| and/or large deformation in dense tissue were considered erroneous. Analysis of |J| revealed >50% regression (J50) in up to 48% of GTV volumes (J50 >10% in 3/10 cases) and up to 70% volume of parotid glands (J50>10% in 6/10 cases); however, these relatively large volume changes were not consistent across the three DIR algorithms with differences in J50 up to 7 cm3 of GTV and up to 16 cm3 of parotids. Inconsistent deformations at J50 in up to 5 cm3 of tissue were also detected in the oral cavity (6/10) and larynx (3/10 cases), but not in the esophagus (maximum difference of 0.9 cm3) or spinal cord (maximum difference of 0.9 cm3). For mandible, volumes of large deformation up to 23 cm3 were detected implying DVF errors. Thresholding |C| at 0.8 (80% relative rotation) allowed rapid identification of errors in DVFs. For GTVs, at least 1 DVF per patient could be identified and disregarded due to non-physical, large rotations in each of the 3/10 cases with large differences in J50(>5 cm3 with |C|>0.8). No single algorithm was consistently superior to the others in terms of minimizing large, unrealistic deformations across the ten cases considered. Fully automated detection of anatomic response to radiation therapy is not currently feasible based on the presence of implausible deformations and rotations; however, by incorporating DVF-QA utilizing Jacobian and curl, it may be possible to rule-out DVFs and find realistic, patient-specific anatomic variations.