L-TGVN: Leveraging Longitudinal Priors for Personalized Rapid MRI

Abstract: MRI provides excellent soft-tissue contrast without ionizing radiation, but long acquisition times increase patient discomfort while also raising exam costs and limiting scanner throughput. A common approach to reduce scan time is to acquire fewer measurements, which yields an ill-posed linear inverse problem; recovering diagnostic-quality images therefore requires incorporating prior knowledge beyond the measured data. In follow-up exams, the most recent prior scan of a patient can provide a highly informative subject-specific context, but practical use is complicated by temporal changes (including pathology progression), misalignment between scans, and protocol drift across acquisitions. In this work, we introduce L-TGVN, a Longitudinal Trust-Guided Variational Network that leverages prior scans as side information to reconstruct the current scan from heavily undersampled measurements. Crucially, L-TGVN constrains the influence of prior scans to be consistent with the acquired measurements. Unlike many existing longitudinal reconstruction methods, it does not require explicit pre-registration between prior and current scans. It further accommodates differences in acquisition protocols across visits (e.g., changes in sequence parameters). We evaluate L-TGVN against matched-capacity baselines, including prior-guided methods and methods that do not use longitudinal priors, and observe consistent improvements in standard quantitative metrics together with better preservation of fine structures at challenging accelerations. Source code is available at github.com/sodicksonlab/L-TGVN.

• The paper introduces L-TGVN, which integrates subject-specific longitudinal priors through trust-guided regularization to tackle the ill-posed nature of undersampled MRI data.
• It employs a dual-encoder U-Net and iterative unrolling to implicitly align and fuse prior and current scans, yielding superior metrics like SSIM and PSNR at aggressive acceleration rates.
• Results demonstrate that L-TGVN preserves fine neuroanatomical details and outperforms baseline methods, validating subject-specific prior exploitation without explicit registration.

L-TGVN: Leveraging Longitudinal Priors for Subject-Specific Accelerated MRI Reconstruction

Problem Context and Motivation

The acceleration of Magnetic Resonance Imaging (MRI) through undersampling in k-space is limited by the resulting ill-posedness of the inverse problem, which leads to degraded image quality, especially at high acceleration factors. While classical techniques—such as parallel imaging and compressed sensing—impose population-based priors, these constraints are insufficient to fully disambiguate heavily undersampled data. Longitudinal studies, where prior scans for a given subject are available, present an opportunity to inject high-fidelity, subject-specific context into the reconstruction process. However, robustly leveraging this information is challenged by temporal anatomical changes, misalignment, and protocol drift.

Methodological Framework: The L-TGVN Architecture

L-TGVN (Longitudinal Trust-Guided Variational Network) addresses the aforementioned challenges by tightly integrating prior scans as side information, with a structural safeguard to constrain their influence to components of the reconstruction that are ambiguously determined by the current measurement operator.

Trust-Guided Regularization:

The core innovation lies in partitioning the solution space according to the singular value spectrum of the forward operator, such that guidance from longitudinal priors is only exerted on directions insufficiently determined by acquired data. The differentiable singular value threshold utilized in the projection operator $P_\delta$ enables end-to-end learning of the data-prior fusion boundary.

Conditional Dual-Encoder Network:

A dual-encoder, shared-decoder U-Net conditions both on the prior scan stack and the current intermediate reconstruction, with feature-level fusion at each resolution scale. This design, avoiding explicit pre-registration, allows implicit alignment and protocol adaptation within the network, which is critical in the face of inter-scan misalignments and protocol heterogeneity.

Iterative Unrolling with Data Consistency Enforcement:

L-TGVN employs multiple (T) Landweber iterations with embedded trust-guided updates, enforcing consistency with the acquired k-space while regularizing only the underdetermined subspace via side information.

Empirical Evaluation

Experiments were conducted on an extensive multi-coil brain MRI dataset, encompassing heterogeneous acquisition parameters, scanner types, and significant temporal intervals between prior and follow-up scans (median 326 days). Three levels of k-space acceleration (20x, 15x, 10x) were retrospectively simulated.

Baselines:

Comparisons included E2E-VN (population-prior, no longitudinal input), and matched-capacity longitudinal prior methods—MTrans and DMSI—adapted from multi-contrast settings.

Results:

L-TGVN produced consistently superior quantitative metrics across all acceleration factors:

• In the most challenging scenario (20x acceleration), L-TGVN achieved a mean SSIM of 90.89 (vs. 87.65 for E2E-VN and 77.05/53.78 for MTrans/DMSI), PSNR of 31.27 dB (vs. 29.13 dB for E2E-VN), and lowest NRMSE.
• Qualitative inspection revealed that L-TGVN uniquely preserved subtle neuroanatomical features and pathologies, while alternative longitudinal methods produced oversmoothed results, often failing to recover fine structure.
• Statistical significance was confirmed via a Wilcoxon signed-rank test ($p < 0.01$).

Notably, the claim that explicitly incorporating subject-specific priors—constrained by a trust-guided mechanism—yields marked improvements even over increased regularizer capacity is strongly substantiated.

Theoretical and Practical Implications

The L-TGVN paradigm demonstrates that it is feasible, and practically advantageous, to integrate longitudinal side information without explicit image-domain registration, and in a manner robust to temporal evolution and protocol variance. By anchoring the influence of prior scans within an operator-informed trust region, the approach mitigates the twin risks of hallucination (over-reliance on priors) and underutilization (insufficient exploitation of subject-specific context).

From a systems perspective, the model can be integrated into clinical MRI workflows where prior exams exist, providing substantial acceleration benefits and improved diagnostic reliability, especially for follow-up studies.

Limitations and Prospective Advancements

While the study leverages only the most recent prior and a limited spatial stack, extensions to richer longitudinal context (multiple prior studies, multi-series inputs, integration of clinical metadata, or reader studies) are natural next steps. The architecture’s modularity also enables prospective augmentation with uncertainty quantification, allowing real-time flagging of regions where prior and current data are in conflict—a key prerequisite for clinical deployment.

Additionally, the framework invites exploration into broader domains of medical imaging where longitudinal or cross-modal priors are available, and it portends applications in emerging adaptive acquisition paradigms.

Conclusion

L-TGVN introduces a principled, trust-guided fusion of longitudinal prior scans for rapid MRI reconstruction, yielding consistent gains in fidelity and robustness at aggressive acceleration rates. By aligning the influence of side information with the data ambiguity structure, the method resolves the inherent tradeoff between prior exploitation and data consistency. These results advance the state of personalized, data-driven MRI reconstruction and motivate further research toward adaptive, uncertainty-aware, and clinically integrated longitudinal imaging pipelines.