Title: L-TGVN: Utilizing Longitudinal Priors for Accelerated Personalized MRI

Magnetic Resonance Imaging (MRI) offers superior soft-tissue differentiation without exposing patients to ionizing radiation. However, the lengthy data acquisition periods often lead to patient discomfort, increased examination costs, and reduced scanner efficiency. A prevalent strategy to shorten scan durations involves acquiring a reduced set of measurements. This approach transforms image reconstruction into an ill-posed linear inverse problem, necessitating the integration of prior knowledge to recover images suitable for diagnostic purposes.

In the context of follow-up examinations, a patient’s most recent prior scan serves as a potent, subject-specific source of contextual information. Nevertheless, its practical application is hindered by several challenges, such as temporal alterations like disease progression, misalignment issues between different scans, and variations in acquisition protocols. To address these complexities, this study presents L-TGVN (Longitudinal Trust-Guided Variational Network). This method employs prior scans as auxiliary data to reconstruct current images from heavily undersampled inputs. A key feature of L-TGVN is its mechanism to limit the influence of prior scans, ensuring consistency with the newly acquired measurements.

Distinct from numerous existing longitudinal reconstruction techniques, L-TGVN eliminates the need for explicit pre-registration between the prior and current scans. Additionally, it is designed to handle discrepancies in acquisition protocols across different visits, such as shifts in sequence parameters. The performance of L-TGVN was assessed against baselines with comparable capacity, including both prior-guided approaches and methods that exclude longitudinal priors. The results demonstrated consistent enhancements in standard quantitative metrics and superior retention of fine structural details, particularly at high acceleration rates. The source code for L-TGVN is accessible at github.com/sodicksonlab/L-TGVN.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC