IIRP-Net: Iterative Inference Residual Pyramid Network for Enhanced Image Registration

Deep learning-based image registration (DLIR) meth-ods have achieved remarkable success in deformable im-age registration. We observe that iterative inference can exploit the well-trained registration network to the fullest extent. In this work, we propose a novel Iterative Inference Residual Pyramid Network (IIRP-Net) to enhance registration performance without any additional training costs. In IIRP-Net, we construct a streamlined pyramid registration network consisting of a feature extractor and residual flow estimators (RP-Net) to achieve generalized capabilities in feature extraction and registration. Then, in the inference phase, IIRP-Net employs an iterative inference strategy to enhance RP-Net by iteratively reutilizing residual flow es-timators from coarse to fine. The number of iterations is adaptively determined by the proposed IterStop mecha-nism. We conduct extensive experiments on the FLARE and Mindboggle datasets and the results verify the effectiveness of the proposed method, outperforming state-of-the-art de-formable image registration methods. Our code is available at https://github.com/Torbjorn1997/IIRP-Net.

1. Introduction

Deformable image registration is an important task in Computer vision, focusing on establishing non-linear dense cor-respondences between two n-D images. It has widespread applications, particularly in medical image analysis [6], [29], [32] and remote sensing [9]. Traditional deformable regis-tration methods usually formulate image registration as an optimization task and attempt to minimize the energy function in an iterative manner. Common intensity-based op-timization methods [5], [33], [37] utilize the intensity differ-ences between images as the energy function. However, in regions with weak texture, the gradient tends to be small, leading to optimization falling into local minima [20]. To address this issue, feature-based optimization methods are proposed, which utilize various feature representation functions such as graph spectral representation [21], structure tensor [35], and Gabor features [28], [36]. These methods ex-tend the image intensity information into more comprehen-sive feature information, enhancing the capability to handle areas where intensity-based methods might struggle. Traditional optimization registration methods treat the registration task as an independent iterative optimization problem. When the target image pairs exhibit significant anatomical appearance variations, the registration time increases dra-matically [25].