Mamba-based Framework for segmentation joint registration

The main purpose of deformable image registration is to improve the registration accuracy of the model by minimizing the difference between a pair of fixed and moving images. Since many anatomical features are closely related to each other, using the supervision of auxiliary tasks (such as supervised anatomical segmentation) has the potential to enhance the realism of the distorted image after registration. In this work, we adopt a multi-task learning framework to formulate registration and segmentation as a joint problem, in which we use a shared encoder to extract features and then use respective decoders for prediction. This structure can effectively reduce the risk of overfitting and reduce the parameters of the network through parameter sharing. Firstly, we propose a channel attention module in the encoding layer to selectively enhance the channel features that are more crucial for the current task according to the characteristics of the input data. On the other hand, Mamba has demonstrated its excellence in long-range modeling with linear complexity, especially in the field of natural language processing. Inspired by TransMorph, we integrate the mamba block into the encoder to capture long-range spatial relationships and optimize memory usage. In this paper, we design a mamba network specifically for the 3D image registration task. Using the U-shaped network architecture, we calculate the deformation field based on the target and source volumes and propose a fine-grained feature extraction module for high-dimensional feature learning. At the same time, the decoder part designs a Selective Feature Fusion (SFF) module. By estimating the attention maps of the two features of segmentation and registration, it helps the model selectively focus on significant regions with a very low computational burden, allowing the network to learn the information needed in each encoder. At the same time, different encoders can use the features of another encoder to enhance their own features. Qualitative and quantitative results show that in our specific experimental setup, the proposed method is superior to other methods in the segmentation and registration tasks of 3D ultrasound images.