Clinical evaluation of free-breathing cardiac multi-parametric mapping using dictionary-based motion correction

Free-breathing cardiac multi-parametric mapping is clinically important but requires accurate motion correction (MoCo). The clinical adoption of the dictionary-matching and low-rank (DM + LR) method remains limited due to computational bottlenecks and a lack of clinical validation. This study aimed to develop and validate a modified dictionary-matching and low-rank (mDM + LR) MoCo approach with improved computational efficiency and diagnostic performance in free-breathing cardiac T1/T2 mapping.

This prospective study enrolled 130 patients with cardiac diseases and 23 healthy controls (HCs). All participants underwent cardiac magnetic resonance imaging (MRI) on a 3T scanner (uMR 790) using electrocardiogram-gated balanced Steady-State Free Precession (bSSFP)-based multimapping for joint T1/T2 mapping under free-breathing conditions. Breath-hold multimapping, Modified Look-Locker Inversion recovery (MOLLI), and T2 mapping served as reference standards in a subset of 19 patients. The mDM + LR MoCo method integrated a pre-trained multi-layer perceptron (MLP), trained on 12.5 million extended phase graph (EPG)-simulated samples, to map T1, T2, and RR-interval history to signals, reducing the runtime to ~25 seconds per sample. MoCo accuracy was evaluated against non-MoCo and parametric image registration with total variation-regularization (pTVreg) using quantitative metrics [Dice similarity coefficient (DSC) scores, mean contour distance (MCD) values, and relative dictionary-matching errors (RDMEs)], qualitative map scores assessed by two blinded readers, and T1/T2 quantification accuracy via correlation and Bland-Altman analyses against breath-hold references. Diagnostic performance (i.e., sensitivity, specificity, and accuracy) was assessed using thresholds derived from HC breath-hold data. Statistical analyses included the Shapiro-Wilk test, t-test or Mann-Whitney U test, Wilcoxon signed-rank test, intraclass correlation coefficients (ICCs), and Bonferroni correction (significance: P<0.05).

In the patients, mDM + LR outperformed non-MoCo and pTVreg in quantitative metrics such as DSC scores (78.0%±7.6% vs. 61.4%±13.3% and 74.5%±11.2%), MCD values (1.20±0.40 vs. 2.41±1.12 and 1.48±0.72 voxels), and RDMEs (8.4%±2.3% vs. 14.6%±3.9% and 9.9%±3.0%), as well as qualitative scores such as map quality scores (T1/T2: 4.65±0.58/4.69±0.49 vs. 3.72±0.81/3.56±0.75 and 3.76±0.78/3.87±0.75, all P<0.01). The mDM + LR method also resulted in higher correlations between global T1/T2 values and breath-holding reference values (r=0.81/0.80 vs. 0.53/0.46 and 0.70/0.64), improved diagnostic specificity (93%/100% vs. 21%/69% and 64%/81%), and improved diagnostic accuracy (89%/100% vs. 42%/74% and 68%/84%). No statistically significant difference was observed between the DM + LR and mDM + LR results. The processing time for mDM + LR was approximately 25 seconds per sample.

mDM + LR significantly improves MoCo, quantification accuracy, and diagnostic performance for free-breathing multi-parametric mapping and thus could be applied in clinical settings.