Affine registration for medical images via coarse-to-fine Swin Transformer

Affine registration is a pivotal task in medical image processing, especially when automated methods are needed to handle large-scale datasets. Despite its importance, there remains a notable gap in assessing the standalone performance of affine registration networks. Traditional optimization-based methods often suffer from inefficiency and limited generalizability. CNN-based approaches, while useful, struggle with receptive field limitations in managing long-range dependencies. Similarly, pure Vision Transformer (ViT)-based methods grapple with high computational demands due to their quadratic scaling of time complexity with the number of image patches, resulting in significant training and inference time. This paper introduces the Coarse-to-Fine Swin Vision Transformer (C2FSwinT), a model designed for fast and robust image affine registration. Central to our model is the Swin Transformer Block, which confines self-attention to local windows, markedly reducing computational complexity. Our architecture further incorporates a multi-stage progressive approach to enhance feature extraction. Additionally, the model integrates techniques such as pre-processing for centroid registration, non-overlapped patch embedding, and decoupled atomic transformation to refine the learning of affine transformation matrices. Experimental evaluations demonstrate that C2FSwinT surpasses other state-of-the-art methods, including conventional optimization-based and ViT-based approaches, in terms of registration accuracy, generalizability, and efficiency in training and inference times.