TU‐CD‐BRA‐09: Accuracy of Deformable Image Registration in Homogeneous Regions: The Impact of Regularisation

Abstract

PurposeDeformable Image Registration (DIR) consists of computing a deformation field to provide an anatomical mapping between the images to be aligned. The deformation field is regularized to prevent over‐fitting, ensure numerical stability, and support the assumption that natural deformations are smooth. The aim of this work is to evaluate the impact of the regularization on the accuracy of the DIR in homogeneous regions where strong image features are not present to guide the mapping.MethodsThe “DIR Validation dataset” (http://www.creatis.insa‐lyon.fr/rio/dir_validation_data) providesbenchmark data for evaluating 3D CT registration algorithms. It consists of six 4DCT lung datasets, with landmarks identified on the end‐exhalation and end‐inhalation phases. Images were registered using image‐based commercial DIR software (Mirada RTx, Mirada Medical, UK). Lungs were segmented and the contours used for structure‐based registration method, thus simulating a fully homogenous region by blinding registration to the underlying image. Target and Surface registration errors (TRE, SRE) were measured at different values of the regularization parameters.ResultsStructure‐based registration showed similar boundary alignment at low regularisation (mean SFE: 0.41 ± 0.07mm) compared to image‐based DIR (0.38±0.17mm), but internal TRE was much worse (5.1±2.1mm cf. 1.4±0.6mm). As regularisation increased internal TRE decreased slightly (minimum 4.2±1.9mm) however this corresponded to an increase in SFE (maximum: 0.9±0.18mm).ConclusionWhile image‐based DIR can align the internal structures in this case, sometimes few internal structures may be visible in clinical practice (e.g. some tumours, liver, brain). The use of structure‐based registration simulated registration within homogenous regions, while allowing comparison with internal TRE. This study demonstrates that while regularization creates a smooth deformation within such regions but it is not necessarily an accurate one, even where the structure boundary is well aligned. Assessment of the deformation grid is helpful for evaluating registrations in such areas.A. Larrue, A.Cifor, and M.Gooding are employees of Mirada Medical