TSMR-Net: a two-stage multimodal medical image registration method via pseudo-image generation and deformable registration

Abstract

Multimodal medical image registration is critical for accurate diagnosis, treatment planning, and surgical guidance. However, differences in imaging mechanisms cause substantial appearance discrepancies between modalities, hindering effective feature extraction and similarity measurement. We propose TSMR-Net, a two-stage multimodal registration framework. In the first stage, an Intensity Distribution Regression module nonlinearly transforms the fixed image into a modality-consistent generated fixed-like image, reducing inter-modality appearance gaps. In the second stage, a deformable registration network aligns the generated fixed-like and moving images using a unimodal similarity metric. The architecture incorporates a parallel downsampling module for multi-scale spatial feature capture and residual skip connections with a 3D channel interaction module to enhance feature propagation. Experiments on IXI and BraTS2023 datasets show that TSMR-Net outperforms state-of-the-art methods in alignment precision, structural consistency, and deformation stability. These findings validate the two-stage strategy’s effectiveness in bridging modality gaps and improving registration accuracy. TSMR-Net provides a scalable, robust solution for diverse multimodal registration tasks with strong potential for clinical application.