|
Image registration algorithms often depend on model parameters that can substantially impact registration accuracy. Current strategies for optimizing registration performance depend on retrospective assessment of accuracy measures such as target registration error (TRE) to identify the most appropriate model parameterization. However, this process of hyperparameter tuning may not produce results that adequately generalize to inter- and intra-dataset variabilities. In this work, we present an analysis framework based on the Akaike Information Criterion (AIC) that permits dynamic runtime adaptation of model parameters by maximizing the informativeness of the registration model with respect to the specific instance of available data constraints. We implement this parameter adaptation framework within a frequency band-limited boundary condition reconstruction approach to efficiently resolve modal harmonics of soft tissue deformation during image registration. We find that minimization of the AIC measure can be used as a surrogate for optimizing TRE when determining optimal model parameterization. Our registration approach automatically tunes model complexity to match informational constraints via an AIC-weighted ensemble model consisting of a collection of registration candidates computed in parallel. Within the context of image-to-physical registration on a registration challenge dataset, we show that our method achieves TRE comparable to other state-of-the-art methods of 4.86±1.07 mm, without the need for any hyperparameter tuning. When an exhaustive fine-tuning approach is applied, the band-limited reconstruction approach exhibits average TRE of 2.99±0.66 mm, which outperforms all other state-of-the-art registration methods contributed to the registration challenge. This technique is expected to improve registration accuracy and robustness by providing an information theoretically optimal strategy to adjust model parameters in a fully prospective manner when generalizing a registration algorithm to new data.
|
