Patient-Specific Quality Assurance of Deformable Image Registrations Using Atlas for Adaptive Radiotherapy of Lung Cancer

Abstract

<h3>Purpose/Objective(s)</h3> Evaluating the uncertainty of deformable image registration (DIR) is challenging because the ground truth is often unavailable. We developed an automated method to create realistic deformations via an atlas and assess DIR algorithms on a patient-specific level for adaptive radiotherapy. <h3>Materials/Methods</h3> A library of deformations was created by extracting the longitudinal anatomical changes observed from an atlas of 60 locally advanced non-small cell lung cancer patients treated with IMRT (60Gy, 30 fx). The deformation vector field (DVF) between the planning CT (pCT) and last weekly CBCT of an atlas patient was derived via a free-form DIR algorithm and served as a ground truth pattern. An inquiry patient was first matched to an atlas patient on gross tumor volume (GTV) location and volume. The pCT of an inquiry patient was then deformably registered to that of an atlas patient to establish a voxel-based correspondence. Then, the deformation pattern of the atlas patient was transferred to the inquiry patient pCT by applying the known DVF to the corresponding voxels, resulting in a digital CT phantom. Subsequently, this CT phantom was appended with CBCT artifacts using a physics-based augmentation (Alam, PMB 2021) to imitate a weekly CBCT. To evaluate DIRs, a large deformation diffeomorphic metric mapping (LDDMM) algorithm, and two commercial systems (C1 C2) were used between the phantom (both simulated CT and CBCT) and actual pCT of the inquiry patient. Derived DVFs were compared to the ground truth. The voxel-level geometric and dosimetric uncertainties of DIRs were calculated using the DVF95% errors, fraction of volume with ≤1.5mm errors (V1.5mm) and mean dose errors (D_mean) for GTV and esophagus. <h3>Results</h3> In a retrospective evaluation of 10 paired inquiry and atlas patients, the actual deformations observed between the pCT and weekly CBCTs of inquiry patients were well-contained by the combined patterns from the matched atlas, validating the feasibility of an atlas. LDDMM performed consistently well, especially for the CT-CBCT registration (p<0.01, two-tailed t-test). GTV mean dose is less sensitive to DIR errors than esophagus mean dose. <h3>Conclusion</h3> It is feasible to augment anatomical changes for a particular inquiry patient using deformation patterns from matched patients in an atlas and evaluate the uncertainty of DIR algorithms in realistic simulations. Integration of such an automated program facilitates the clinical implementation of adaptive radiotherapy that involves longitudinal imaging studies.