FT-Reg: Unsupervised Multimodal Medical Image Registration Using Dynamic Feature Translation

Multimodal registration plays a crucial role in medical analysis, necessitating the precise alignment of images from various imaging modalities. However, inherent differences in appearance and imaging characteristics present considerable challenges. To address these challenges, we introduce a novel unsupervised approach termed FT-Reg for multimodal deformable registration via dynamic feature-to-feature (F2F) translation. This method departs from traditional reliance on image-to-image (I2I) translation for multimodal registration by employing a feature translation scheme that aligns the more distinctive extracted features. We demonstrate that this framework significantly bolsters the network′s capacity to predict more precise deformation fields. Moreover, both the translation and deformation networks adopt an iterative training strategy, which further enhances the registration efficacy. Additionally, we incorporate a samplewise dynamic neural network that adaptively applies the translator to specific feature layers, accounting for sample diversity. Extensive experiments on the Learn2Reg and iSeg benchmark datasets demonstrate the effectiveness of our proposed method. Comparative analyses with other state-of-the-art (SOTA) approaches further confirm the superior registration quality achieved by our method. In addition, we demonstrate the practical application of our method on the real-scanned clinical low-field and high-field magnetic resonance imaging (MRI) datasets. The code is publicly available at https://github.com/HA-xf/FT-Reg.