Title:A Trust-Guided Approach to MR Image Reconstruction with Side Information

Abstract:Reducing MRI scan times can improve patient care and lower healthcare costs. Many acceleration methods are designed to reconstruct diagnostic-quality images from limited sets of acquired $\textit{k}$-space data. This task can be framed as a linear inverse problem (LIP), where, as a result of undersampling, the forward operator may become rank-deficient or exhibit small singular values. This results in ambiguities in reconstruction, in which multiple generally incorrect or non-diagnostic images can map to the same acquired data. To address such ambiguities, it is crucial to incorporate prior knowledge, for example in the form of regularization. Another form of prior knowledge less commonly used in medical imaging is contextual side information garnered from other sources than the current acquisition. Here, we propose the $\textbf{T}$rust-$\textbf{G}$uided $\textbf{V}$ariational $\textbf{N}$etwork $\textbf{(TGVN)}$, a novel end-to-end deep learning framework that effectively integrates side information into LIPs. TGVN eliminates undesirable solutions from the ambiguous space of the forward operator while remaining faithful to the acquired data. We demonstrate its effectiveness in multi-coil, multi-contrast MR image reconstruction, where incomplete or low-quality measurements from one contrast are used as side information to reconstruct high-quality images of another contrast from heavily under-sampled data. Our method is robust across different contrasts, anatomies, and field strengths. Compared to baselines that also utilize side information, TGVN achieves superior image quality at challenging under-sampling levels, drastically speeding up acquisition while minimizing hallucinations. Our approach is also versatile enough to incorporate many different types of side information (including previous scans or even text) into any LIP.