WTDL-Net: Medical Image Registration Based on Wavelet Transform and Multi-Scale Deep Learning

Abstract Three-dimensional (3D) medical image registration has drawn substantial research attention. In comparison to traditional approaches, deep learning techniques present significant advantages in terms of speed and accuracy. However, large deformations and complex transformations pose challenges for single-modality image registration. In this study, we propose WTDL - Net. It is a multi - scale registration network that incorporates wavelet transform. First, low-frequency sub-images generated by WT at various resolutions were used as inputs to the multiscale registration network. Coarse-to-fine registration was achieved by analyzing image information at different resolutions. Second, the high-frequency components derived from the WT were combined to create a High-Frequency Infographic. This Infographic is applied to constrain multilevel registration, thereby enhancing the optimization of registration details. The proposed approach is demonstrated to be superior to current deep - learning - based registration techniques through comprehensive quantitative and qualitative evaluations conducted over four MR brain scan datasets.