Deformable Image Registration with Multi-scale Feature Fusion from Shared Encoder, Auxiliary and Pyramid Decoders

TL;DR

This work introduces a deformable image registration framework based on a multi-scale pyramid network with a shared encoder and a novel shared auxiliary decoder for both image pairs. A multi-scale feature fusion block (MSFB) combines low-level, high-level, and prior-scale deformation information to produce accurate, scale-consistent deformation fields, trained with a loss that blends normalized cross-correlation terms and deformation regularization. On the Learn2Reg LUMIR dataset, the approach delivers higher registration accuracy and smoother, more plausible deformations than several baselines. The method advances unsupervised DIR by efficiently leveraging cross-scale, cross-image features to handle complex deformations in medical imaging tasks.

Abstract

In this work, we propose a novel deformable convolutional pyramid network for unsupervised image registration. Specifically, the proposed network enhances the traditional pyramid network by adding an additional shared auxiliary decoder for image pairs. This decoder provides multi-scale high-level feature information from unblended image pairs for the registration task. During the registration process, we also design a multi-scale feature fusion block to extract the most beneficial features for the registration task from both global and local contexts. Validation results indicate that this method can capture complex deformations while achieving higher registration accuracy and maintaining smooth and plausible deformations.

Figures (1)

• Figure 1: An example of our method results in registration. From left to right are moving image, deformation field, warped image and fixed image. The red box highlights the significant areas.