HyperPredict: Estimating Hyperparameter Effects for Instance-Specific Regularization in Deformable Image Registration

TL;DR

Hyperparameter selection in deformable image registration is crucial but traditionally data- and label-intensive. HyperPredict learns a conditional mapping from an encoded image pair and a hyperparameter value to predicted registration metrics, enabling test-time, instance-specific hyperparameter selection without labeled data. Using two backends (cLapIRN and NiftyReg) and brain MRI data, it demonstrates accurate metric prediction, competitive Dice scores, and reduced deformation folding when optimizing per pair. The approach offers substantial test-time efficiency and flexibility in choosing region- or task-specific regularization, with potential applicability beyond brain imaging to other registration scenarios.

Abstract

Methods for medical image registration infer geometric transformations that align pairs/groups of images by maximising an image similarity metric. This problem is ill-posed as several solutions may have equivalent likelihoods, also optimising purely for image similarity can yield implausible transformations. For these reasons regularization terms are essential to obtain meaningful registration results. However, this requires the introduction of at least one hyperparameter often termed $λ$, that serves as a tradeoff between loss terms. In some situations, the quality of the estimated transformation greatly depends on hyperparameter choice, and different choices may be required depending on the characteristics of the data. Analyzing the effect of these hyperparameters requires labelled data, which is not commonly available at test-time. In this paper, we propose a method for evaluating the influence of hyperparameters and subsequently selecting an optimal value for given image pairs. Our approach which we call HyperPredict, implements a Multi-Layer Perceptron that learns the effect of selecting particular hyperparameters for registering an image pair by predicting the resulting segmentation overlap and measure of deformation smoothness. This approach enables us to select optimal hyperparameters at test time without requiring labelled data, removing the need for a one-size-fits-all cross-validation approach. Furthermore, the criteria used to define optimal hyperparameter is flexible post-training, allowing us to efficiently choose specific properties. We evaluate our proposed method on the OASIS brain MR dataset using a recent deep learning approach(cLapIRN) and an algorithmic method(Niftyreg). Our results demonstrate good performance in predicting the effects of regularization hyperparameters and highlight the benefits of our image-pair specific approach to hyperparameter selection.

Figures (16)

• Figure 1: Overview of the method: (a) We learn the parameters $g_{\theta}$ (yellow box) of an MLP that maps the encoded representation ($e_{o}$) of an image pair (m, f) and a set of hyperparameters ($\lambda$) to evaluation metrics (Dice and folded voxels). At train time, we optimize $g_{\theta}$ by computing the target metrics of the input (blue boxes and dashed blue lines) using a registration algorithm. (b) At test time, given a set of input, we predict the effect for each hyperparameter value. We select the optimal hyperparameter, $\lambda^*$, and use that in the registration algorithm to obtain a desired warped image (red box). The green shapes (encoder and registration) are pretrained.
• Figure 2: Error Distribution Plots on Both HyperPredict Models. Left: Difference in predicted and target dice coefficients for selected hyperparameter values. Right: Difference in predicted and target %nfv for selected hyperparameter values. For visualization purpose, we display results for selected values.
• Figure 3: Validating HyperPredict. Left: Bland-Altman plots showing the agreement between the predicted vs target Dice scores across the entire population for both models (each point represents the average Dice for a particular image pair). Right: Bland-Altman plots showing the agreement between the predicted vs target %nfv for the entire population.
• Figure 4: HyperPredict vs Cross-Validation. Left: Comparison of the Dice Coefficient on both methods for selected labels across the population. Right: Comparison of the %nfv on both methods across the population.
• Figure 5: HyperPredict vs Cross-Validation. Top: Worse case analysis on HyperPredict$_\text{clap}$Bottom: Worse case analysis on HyperPredict$_\text{nr}$.