Adaptive Conditional Contrast-Agnostic Deformable Image Registration with Uncertainty Estimation

TL;DR

AC-CAR introduces an adaptive, contrast-agnostic deformable image registration framework that generalizes to unseen imaging contrasts by combining random convolution-based contrast augmentation with an adaptive conditional feature modulator and a contrast-invariant latent regularization. A separate variance network provides contrast-agnostic uncertainty estimates, enhancing trustworthiness for multi-contrast registration. The method demonstrates superior accuracy and generalization across 3D brain and 2D cardiac datasets, with meaningful uncertainty maps and robust feature invariance compared with state-of-the-art baselines. This approach enables reliable, efficient registration in scenarios with diverse or unknown contrasts, with practical implications for multi-contrast MRI analysis and downstream clinical tasks.

Abstract

Deformable multi-contrast image registration is a challenging yet crucial task due to the complex, non-linear intensity relationships across different imaging contrasts. Conventional registration methods typically rely on iterative optimization of the deformation field, which is time-consuming. Although recent learning-based approaches enable fast and accurate registration during inference, their generalizability remains limited to the specific contrasts observed during training. In this work, we propose an adaptive conditional contrast-agnostic deformable image registration framework (AC-CAR) based on a random convolution-based contrast augmentation scheme. AC-CAR can generalize to arbitrary imaging contrasts without observing them during training. To encourage contrast-invariant feature learning, we propose an adaptive conditional feature modulator (ACFM) that adaptively modulates the features and the contrast-invariant latent regularization to enforce the consistency of the learned feature across different imaging contrasts. Additionally, we enable our framework to provide contrast-agnostic registration uncertainty by integrating a variance network that leverages the contrast-agnostic registration encoder to improve the trustworthiness and reliability of AC-CAR. Experimental results demonstrate that AC-CAR outperforms baseline methods in registration accuracy and exhibits superior generalization to unseen imaging contrasts. Code is available at https://github.com/Yinsong0510/AC-CAR.

Figures (10)

• Figure 1: Overview of the proposed registration framework. We first simulate two augmented images for each moving and fixed image with the contrast augmentation module. The augmented images are then used as the network inputs. At the encoder part, we use our proposed adaptive conditional feature modulator (ACFM) and contrast-invariant latent regularization (CLR) to extract contrast-invariant features. The learned features are then used to estimate the deformation field. The deformation field is then used to warp the pre-augmented moving image. The similarity loss is calculated on the warped image and the pre-augmented fixed image.
• Figure 2: Overview of the proposed Adaptive Conditional Feature Modulation Module (ACFM).
• Figure 3: Overview of the proposed contrast-agnostic uncertainty estimation framework.
• Figure 4: Qualitative results on CamCAN dataset. We present the registration results of the middle slices along the z-axis of the same volume for illustration. The first row shows the error map of our proposed method against the baseline. The Dice score and HD95 of the whole volume are shown at the bottom left of the error map. The second row shows the warped images overlaid with the deformation fields.
• Figure 5: Qualitative results on CMRxRecon dataset using T1 Mapping data for training. We show the results of registering misaligned images of $\mathrm{TI_{i=2,4,6,8}}$ to $\mathrm{TI_{1}}$. Columns 2-5 present the warped images overlaid with the deformation fields, and columns 6-10 present the error map of our proposed method against the baseline. The Dice score and HD95 are shown at the top left of the error map.