A edge prior constraint Mamba network for medical image super-resolution generation

Abstract

Existing deep learning-based algorithms for super-resolution generation of medical images are usually based on convolutional architecture or transformer module. Algorithms based on convolutional architecture are limited by the inherent inductive bias to efficiently acquire global contextual information, while algorithms based on transformer module cannot be practically deployed due to the high computational cost. To overcome these problems, a cyclic scanning Mamba network is guided by edge prior knowledge constraints for achieving super-resolution generation of medical images. In this paper, we propose to integrate the Mamba module into the Unet network to acquire long range dependencies between regions in medical images at a small computational cost. In addition, a sequence cyclic scanning component is designed to process the input image sequences from both forward and backward directions, which enhances the sensitivity of our model to changes in orientation information. Meanwhile, an edge prior control module is developed to add additional steering features and constraints to the generation process. The results of our comparison and ablation experiments show that the super-resolution performance and computational resource occupancy outperform existing methods on IXI and fastMRI medical image datasets. The downstream visual task of brain tumour segmentation using a medical image segmentation network also shows the effectiveness of our method with a mean Dice Score of 57.73 % on the BraTS2021 dataset.