SSCA-DIPNet: structured sparse and channel-attentive deep image prior network for highly-accelerated MRI

magnetic resonance imaging (MRI) aids diagnosis yet suffers slow scanning. k-space undersampling causes ill-posed reconstruction, with existing methods flawed. This study builds a data-efficient network to tackle high-acceleration, limited-data MRI reconstruction. We proposed the structured sparse and channel-attentive deep image prior network (SSCA-DIPNet), which integrated three innovative modules based on ISTA-Net+: (1) a channel-adaptive structured sparsity module (SSM) for differential feature preservation, (2) a channel-attention deep image prior (DIP) module for artifact repair, (3) a globally learnable fusion mechanism combined with symmetric loss to enhance model generalization. Two branch models were derived from ISTA-Net+ baseline: SS-Net with only the SSM, and CA-DIPNet with only the channel-attention DIP module. Experiments on two clinical datasets (179 training samples each) and supplementary fast MRI raw k-space validation are performed under 5×/10×/20× acceleration and evaluated via peak signal-to-noise ratio and structural similarity index. The proposed model consistently outperforms ISTA-Net+, SS-Net and CA-DIPNet across all scenarios. Using scarce data, SSCA-DIPNet yields robust reconstruction with refined anatomical details and reduced artifacts. It balances reconstruction performance and data efficiency, offering a fresh paradigm for iterative-unfolding MRI reconstruction to support clinical fast scanning under limited data.