Transformer–Mamba based dynamic threshold generation for dual frame network and its application in compressed sensing MRI

Compressed sensing magnetic resonance imaging (CSMRI), a typically accelerated data acquisition technology in MRI system, aims to recovery high-quality MR images from fewer k-space data. Recent deep unfolded methods have achieved remarkable performance in terms of reconstruction quality, but there is still room for reconstruction performance improvement due to the limitation of representation ability of the prior network. To improve the representation ability of the prior network as well as the reconstruction performance, we propose a novel Transformer–Mamba-based dual frame network named DualTMNet and serving it as a prior network for constructing an unfolded CSMRI algorithm. Besides the synthesis and analysis frames, the adaptive thresholds in dual frame network are also important for high representation ability of the overall network, we elaborate a dynamic threshold generation module to learn the threshold vector from input instances adaptively. Therein, a noisy intensity generation module (NGM) and Transformer–Mamba cross-fusion module (TMCF) are meticulously elaborated in the dynamic threshold generation module. TMCF introduces a Mamba module and a Transformer module to exploit the global contexts and complementary non-local information, respectively. By doing so, the proposed network can effectively extract important features as well as generate the faithful thresholds for image filtering. Via an end-to-end supervised learning strategy, the dynamic threshold generation module and the dual frame network are jointly trained from the labels and their counterparts. Serving the proposed DualTMNet as a prior network, we illustrate its performance on the task of CSMRI. Extensive experiments demonstrate the proposed method can achieve higher-quality reconstructions over the benchmark competitors in terms of both qualitative and quantitative quality.