Learning Deformable Intra-Patient Liver Registration with Graph Cross-Attention

Liver registration is a critical component of computer-assisted hepatic interventions. Accurate alignment of the liver surface and internal vascular network is essential for planning and evaluating interventions, and follow-up monitoring. While conventional techniques achieve high precision in spatial alignment, they often incur high computation times. Therefore, deep learning methods have been proposed for fast registration, albeit with limited generalization to real-world scenarios. In this work, we introduce a novel learnable graph cross-attention mechanism specifically designed for the alignment of the liver vascular system. We build on related work, and extract feature pyramids from the inputs in a dual-stream fashion. Then, we represent the features at the coarsest-scale as graphs, learn critical structural information, and model the relationship between the two graphs through cross-attention. Our proposed module is a novel solution that can enhance the performance of various encoder-decoder architectures. To the best of our knowledge, this is the first application of a graph cross-attention mechanism for liver registration. We carried out the experimental validation on 20 procedures from the Innsbruck University Hospital, demonstrating that our graph module yields significant improvements, with achieved enhancements ranging from 5 % to 7 %.