Magnetic Resonance Imaging (MRI) has limited applications in clinical scenarios, particularly in dynamic imaging and interventional surgery, due to its long data collection time. Compressed sensing (CS) theory can shorten MRI scan time by undersampling k-space data, but the reconstruction quality is highly dependent on effective prior constraints. In recent years, it shows excellent performance in MR image reconstruction by adopting deep learning methods, which performance usually depends on large-scale, high-quality clinical training datasets, while obtaining clinical data is often very difficult in practice. To address this challenge,we propose a multi-prior constrained reconstruction method for undersampled MRI using reference-driven Deep Image Prior (DIP) and Regularization by Denoising (RED). A key advantage of our method is its independence from large pre-trained datasets; instead, the entire reconstruction is guided by only a single reference image, offering a low-data-dependency pathway for undersampled MRI reconstruction.
The proposed method capitalizes on the structural similarity between the reference and target images, constructing a two-stage process core-driven by a single reference image. In the first stage, the denoising engine within the RED is trained using the reference image, thereby embedding structural priors into the RED regularization. In the second stage, the same reference image guides the DIP-based reconstruction. A multiprior collaborative optimization model is designed, by combining k-space data consistency constraints, wavelet domain structural similarity constraints, and RED regularization. By optimizing the network parameters in this model, high-quality reconstruction from undersampled data is achieved. This design allows the same reference image to drive both the RED denoising engine training and the DIP-based learning, ensuring consistent and efficient utilization of structural priors throughout the workflow. The integration of multi-domain priors, which span k-space, wavelet, and image domains, enables collaborative optimization from multiple dimensions and thus further enhances reconstruction performance.
Experimental results show that the proposed method surpasses existing comparable techniques in both visual assessment and quantitative metrics, more accurately recovering tissue details and edge structures. Importantly, it substantially reduces the reliance on large-scale pre-training data. Through the reference-driven strategy and multi-prior fusion mechanism, this method jointly improves the accuracy and effectiveness of prior incorporation, ensuring reconstruction fidelity and robustness. In summary, this work presents a promising technical pathway to overcome the challenges of clinical MR data acquisition and achieve high-quality reconstruction.