Customizing Native T1 Mapping: the effects of compressed sensing, deep learning-based denoising and high-resolution on measurement of native myocardial T1

Figures

• 

  Fig. 1. Basal slices of the 10 T1 mapping methods: the first row shows images reconstructed using conventional CS reconstruction with acceleration factors 3–5 and varying spatial resolutions (1.6×1.6×10 mm3, 1.8×1.8×10 mm3, and 2.0×2.0×10 mm3 for CS3 and 2.0×2.0×10 mm3 for CS4 and CS5); the second row displays the same acquisitions reconstructed with DL-denoising. CS compressed sensing, DL deep learning

• 

  Fig. 2. Qualitative assessment of blurring, aliasing, susceptibility artifacts, and overall IQ for each T1 mapping method. Image quality was rated using a 4-point Likert scale (1 = excellent, 2 = good, 3 = fair, 4 = poor), with stacked bar plots showing the percentage distribution of scores across three independently blinded readers. Susceptibility artifact scores remained stable across all ten mapping methods. DL-denoising consistently improved image quality and reduced blurring across all configurations, with only a marginal increase in aliasing severity. Higher CS acceleration levels were associated with a gradual degradation in image quality and increased blurring and aliasing artifacts. Statistically significant post hoc pairwise differences are indicated for comparisons with the reference standard (asterisk) and for comparisons between the two reconstruction approaches (conventional CS vs DL-denoising) within the same acquisition (light blue bar); the reference standard method is highlighted with a red box. IQ image quality, CS compressed sensing, DL deep learning

• 

  Fig. 3. Bull’s eye maps showing mean T1 values per AHA segment for each evaluated T1 mapping method. Each map is color-coded using the Philips 3T color scale and illustrates the spatial distribution of average T1 values across myocardial segments, highlighting very close mean values across methods in both acquisition and reconstruction strategies, based on per-segment comparison. The top row displays the mapping methods reconstructed using the conventional CS reconstruction at different spatial resolutions and CS acceleration levels, while the bottom row shows the corresponding methods with DL-denoising reconstruction from the same acquisitions. CS compressed sensing, DL deep learning

• 

  Fig. 4. Bull’s eye maps showing SD of T1 values per AHA segment for each evaluated T1 mapping method. Each map is color-coded using the Viridis scale and illustrates the spatial distribution of variability across myocardial segments, revealing localized differences. Higher SD values were observed in mid-ventricular and basal lateral segments as well as in apical regions, likely reflecting their higher susceptibility to technical and physiological artifacts. The top row displays the mapping methods reconstructed using the conventional CS reconstruction at different spatial resolutions and CS acceleration levels, while the bottom row shows the corresponding methods with DL-denoising reconstruction from the same acquisitions. CS compressed sensing, DL deep learning, SD standard deviation

• 

  Fig. 5. Segmental agreement of native T1 mean values across the ten mapping methods, assessed using the two-way random-effects ICC2. The bull’s eye representation follows the standard AHA 16-segment model. Overall, a moderate level of inter-method agreement was observed across most myocardial segments. Notably, the lowest concordance was found in the lateral basal and mid-ventricular segments (segments 5 and 6), which are typically more susceptible to motion and susceptibility-related artifacts, potentially explaining the reduced reliability observed in these areas. Dashed lines in the color scale indicate conventional ICC thresholds for poor (<0.50), moderate (0.50–0.75), good (0.75–0.90), and excellent (>0.90) agreement. ICC intraclass correlation coefficient

• 

  Fig. 6. Pairwise ICC2 maps showing segmental agreement between each T1 mapping method and the reference (CS3 2×2×10). DL CS3 2×2×10 showed the highest concordance across all segments, while agreement decreased with higher acceleration and resolution, particularly in lateral basal segments. Dashed lines in the color scale indicate conventional ICC thresholds for poor (<0.50), moderate (0.50–0.75), good (0.75–0.90), and excellent (>0.90) agreement. ICC intraclass correlation coefficient, CS compressed sensing, DL deep learning, SD standard deviation

• 

  Fig. 7. Bull’s eye maps showing the 95% LoA widths derived from Bland–Altman analysis comparing each T1 mapping method to the reference standard (CS3 2×2×10), at the level of individual AHA segments, expressed in ms. LoA widths reflect the spread of agreement and remained below 200 ms for most segment–method combinations, indicating moderate variability in line with the intrinsic nature of T1 mapping. Broader LoA widths were predominantly observed in segments 4–6 (basal lateral wall) and apical segments (13–16), regions known to be susceptible to motion and susceptibility-related artifacts. Among all methods, DL CS3 2×2×10 demonstrated the narrowest LoA widths, including in the most challenging segments, underscoring its superior agreement and robustness. LoA limits of agreement, CS compressed sensing, DL deep learning

• 

  Fig. 8. Bull’s eye maps showing segmental bias values between each T1 mapping method and the reference standard (CS3 2×2×10), expressed in ms. The overall bias ranged from −21 ms to +22 ms, indicating a generally low degree of systematic deviation. High-acceleration methods (e.g., CS4 2×2×10 and CS5 2×2×10) tended to overestimate T1 values relative to the reference, with several segments exceeding +10 ms. In contrast, DL-denoising reconstructions, particularly when applied to high-resolution datasets (e.g., DL CS3 1.6×1.6×10 and DL CS3 1.8×1.8×10), showed consistent underestimation, with negative biases. The smallest biases, often near zero, were observed in the high-resolution CS3 acquisitions reconstructed with conventional CS reconstruction (CS3 1.6×1.6×10 and CS3 1.8×1.8×10), indicating excellent agreement with the reference across myocardial segments. CS compressed sensing, DL deep learning

• 

  Fig. 9. Bull’s eye maps illustrating the results of the TOST for statistical equivalence between each T1 mapping method and the reference standard (CS3 2×2×10) at the level of individual AHA segments. Equivalence was tested using a predefined margin of ±30 ms. Green segments indicate statistical equivalence, while red segments represent regions where equivalence was not achieved; the bias value for each segment is reported within the corresponding region. The vast majority of segment-wise comparisons demonstrated equivalence; exceptions were limited to segment 5 for the high-resolution method CS3 1.6×1.6×10, and segments 4 and 5 for the accelerated method CS4 2×2×10, regions located in the inferior-lateral wall, which are known to be particularly prone to motion and susceptibility artifacts, contributing to reduced measurement reliability in these areas. TOST two one-sided Tests, CS compressed sensing, DL deep learning

• 

  Fig. 10. In the general MOLLI 5(3)3 acquisition scheme (a), after a 180° inversion pulse (blue bar), five images are sequentially acquired during diastole, each corresponding to a different inversion time (TI) over five consecutive heartbeats. This is followed by a three-heartbeat recovery period and a second inversion pulse, allowing the acquisition of three additional images with slightly shifted TIs. Green boxes indicate the acquisition windows. (b) Increasing CS acceleration factors (CS3, CS4, CS5) shorten the acquisition window duration (135.4 ms, 102.1 ms, and 82.2 ms, respectively), thereby improving temporal resolution of the corresponding images. (c) Native T1 maps obtained from the corresponding acquisitions. CS compressed sensing